├── README.md ├── data └── scripts │ ├── eval_subroot.py │ ├── make_dp_dataset.py │ ├── make_subroot_dataset.py │ ├── make_wsd_pred.py │ ├── mergy_subroot_feature.py │ ├── recover_dataset.py │ ├── subroot_decomposition.py │ └── subroot_replace_unk.py ├── node2vec └── train.py ├── onmt ├── Highway.py ├── Loss.py ├── ModelConstructor.py ├── Models.py ├── Optim.py ├── SubwordElmo.py ├── Trainer.py ├── Utils.py ├── __init__.py ├── __init__.pyc ├── io │ ├── AudioDataset.py │ ├── DatasetBase.py │ ├── IO.py │ ├── IO.pyc │ ├── ImageDataset.py │ ├── TextDataset.py │ ├── __init__.py │ ├── __init__.pyc │ └── __pycache__ │ │ ├── AudioDataset.cpython-36.pyc │ │ ├── DatasetBase.cpython-36.pyc │ │ ├── IO.cpython-36.pyc │ │ ├── ImageDataset.cpython-36.pyc │ │ ├── TextDataset.cpython-36.pyc │ │ └── __init__.cpython-36.pyc ├── modules │ ├── AudioEncoder.py │ ├── Conv2Conv.py │ ├── ConvMultiStepAttention.py │ ├── CopyGenerator.py │ ├── Embeddings.py │ ├── Gate.py │ ├── GlobalAttention.py │ ├── ImageEncoder.py │ ├── MultiHeadedAttn.py │ ├── SRU.py │ ├── StackedRNN.py │ ├── StructuredAttention.py │ ├── Transformer.py │ ├── UtilClass.py │ ├── WeightNorm.py │ ├── __init__.py │ └── __pycache__ │ │ ├── AudioEncoder.cpython-36.pyc │ │ ├── Conv2Conv.cpython-36.pyc │ │ ├── ConvMultiStepAttention.cpython-36.pyc │ │ ├── CopyGenerator.cpython-36.pyc │ │ ├── Embeddings.cpython-36.pyc │ │ ├── Gate.cpython-36.pyc │ │ ├── GlobalAttention.cpython-36.pyc │ │ ├── ImageEncoder.cpython-36.pyc │ │ ├── MultiHeadedAttn.cpython-36.pyc │ │ ├── SRU.cpython-36.pyc │ │ ├── StackedRNN.cpython-36.pyc │ │ ├── StructuredAttention.cpython-36.pyc │ │ ├── Transformer.cpython-36.pyc │ │ ├── UtilClass.cpython-36.pyc │ │ ├── WeightNorm.cpython-36.pyc │ │ └── __init__.cpython-36.pyc ├── opts.py └── translate │ ├── Beam.py │ ├── Penalties.py │ ├── Translation.py │ ├── TranslationServer.py │ ├── Translator.py │ ├── __init__.py │ └── __pycache__ │ ├── Beam.cpython-36.pyc │ ├── Penalties.cpython-36.pyc │ ├── Translation.cpython-36.pyc │ ├── TranslationServer.cpython-36.pyc │ ├── Translator.cpython-36.pyc │ └── __init__.cpython-36.pyc ├── preprocess.py ├── requirements.opt.txt ├── requirements.txt ├── resources └── seq2seq4dp.pdf ├── screenshots └── seq2seq_model.png ├── server.py ├── setup.py ├── subroot ├── README.md ├── RUN.md ├── dnn_pytorch │ ├── dnn_utils.py │ ├── generate_features.py │ ├── loader.py │ ├── nn.py │ ├── tag.py │ ├── train.py │ └── utils.py └── subroot │ ├── preprocess.py │ ├── stat.py │ ├── test.py │ └── train.py ├── tools ├── README.md ├── apply_bpe.py ├── average_models.py ├── bpe_pipeline.sh ├── detokenize.perl ├── embeddings_to_torch.py ├── extract_embeddings.py ├── learn_bpe.py ├── multi-bleu-detok.perl ├── multi-bleu.perl ├── nonbreaking_prefixes │ ├── README.txt │ ├── nonbreaking_prefix.ca │ ├── nonbreaking_prefix.cs │ ├── nonbreaking_prefix.de │ ├── nonbreaking_prefix.el │ ├── nonbreaking_prefix.en │ ├── nonbreaking_prefix.es │ ├── nonbreaking_prefix.fi │ ├── nonbreaking_prefix.fr │ ├── nonbreaking_prefix.ga │ ├── nonbreaking_prefix.hu │ ├── nonbreaking_prefix.is │ ├── nonbreaking_prefix.it │ ├── nonbreaking_prefix.lt │ ├── nonbreaking_prefix.lv │ ├── nonbreaking_prefix.nl │ ├── nonbreaking_prefix.pl │ ├── nonbreaking_prefix.ro │ ├── nonbreaking_prefix.ru │ ├── nonbreaking_prefix.sk │ ├── nonbreaking_prefix.sl │ ├── nonbreaking_prefix.sv │ ├── nonbreaking_prefix.ta │ ├── nonbreaking_prefix.yue │ └── nonbreaking_prefix.zh ├── release_model.py ├── test_rouge.py └── tokenizer.perl ├── train.py └── translate.py /README.md: -------------------------------------------------------------------------------- 1 | # Sequence to sequence model for dependency parsing based on OpenNMT-py 2 | 3 | This is a Seq2seq model implemented based on [OpenNMT-py](http://opennmt.net/OpenNMT-py/). It is designed to be presents a seq2seq dependency parser by directly predicting the relative position of head for each given presents a seq2seq dependency parser by directly predicting the relative position of head for each given word, which therefore results in a truly end-to-end seq2seq dependency parser for the first time.word, which therefore results in a truly end-to-end seq2seq dependency parser. 4 | 5 | Enjoying the advantage of seq2seq modeling, we enrich a series of embedding enhancement, including firstly introduced subword and node2vec augmentation. Meanwhile, we propose a beam search decoder with tree constraint and subroot decomposition over the sequence to furthermore enhance our seq2seq parser. 6 | 7 | The framework of the proposed seq2seq model: 8 |
9 | 10 | ## Requirements 11 | 12 | ```bash 13 | pip install -r requirements.txt 14 | ``` 15 | this project is tested on pytorch 0.3.1, the other version may need some modification. 16 | 17 | ## Quickstart 18 | 19 | ### Step 1: Convert the dependency parsing dataset 20 | 21 | ```bash 22 | python data/scripts/make_dp_dataset.py 23 | ``` 24 | 25 | 26 | ### Step 2: Preprocessing the data 27 | 28 | ```bash 29 | python preprocess.py -train_src data/input/dp/src_ptb_sd_train.input -train_tgt data/input/dp/tgt_ptb_sd_train.input -valid_src data/input/dp/src_ptb_sd_dev.input -valid_tgt data/input/dp/tgt_ptb_sd_dev.input -save_data data/temp/dp/dp 30 | ``` 31 | We will be working with some example data in `data/` folder. 32 | 33 | The data consists of parallel source (`src`) and target (`tgt`) data containing one sentence per line with tokens separated by a space: 34 | 35 | * `src-train.txt` 36 | * `tgt-train.txt` 37 | * `src-val.txt` 38 | * `tgt-val.txt` 39 | 40 | Validation files are required and used to evaluate the convergence of the training. It usually contains no more than 5000 sentences. 41 | 42 | 43 | After running the preprocessing, the following files are generated: 44 | 45 | * `dp.train.pt`: serialized PyTorch file containing training data 46 | * `dp.valid.pt`: serialized PyTorch file containing validation data 47 | * `dp.vocab.pt`: serialized PyTorch file containing vocabulary data 48 | 49 | 50 | Internally the system never touches the words themselves, but uses these indices. 51 | 52 | ### Step 2: Make the pretrain embedding 53 | 54 | ```bash 55 | python tools/embeddings_to_torch.py -emb_file_enc data/pretrain/glove.6B.100d.txt -dict_file data/temp/dp/dp.vocab.pt -output_file data/temp/dp/en_embeddings -type GloVe 56 | ``` 57 | 58 | 59 | ### Step 3: Train the model 60 | 61 | ```bash 62 | python train.py -save_model data/model/dp/dp -batch_size 64 -enc_layers 4 -dec_layers 2 -rnn_size 800 -word_vec_size 100 -feat_vec_size 100 -pre_word_vecs_enc data/temp/dp/en_embeddings.enc.pt -data data/temp/dp/dp -encoder_type brnn -gpuid 0 -position_encoding -bridge -global_attention mlp -optim adam -learning_rate 0.001 -tensorboard -tensorboard_log_dir logs -elmo -elmo_size 500 -elmo_options data/pretrain/elmo_2x4096_512_2048cnn_2xhighway_options.json -elmo_weight data/pretrain/elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5 -subword_elmo -subword_elmo_size 500 -subword_elmo_options data/pretrain/subword_elmo_options.json -subword_weight data/pretrain/en.wiki.bpe.op10000.d50.w2v.txt -subword_spm_model data/pretrain/en.wiki.bpe.op10000.model 63 | ``` 64 | 65 | - elmo_2x4096_512_2048cnn_2xhighway_options.json 66 | - elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5 67 | - subword_elmo_options.json 68 | - en.wiki.bpe.op10000.d50.w2v.txt 69 | - en.wiki.bpe.op10000.model 70 | 71 | You can download these files from [here](https://drive.google.com/drive/folders/1ug6ab14fpM22ed_vomOTjjUB8Awh66VM?usp=sharing). 72 | 73 | 74 | ### Step 3: Translate 75 | 76 | ```bash 77 | python translate.py -model data/model/dp/xxx.pt -src data/input/dp/src_ptb_sd_test.input -tgt data/input/dp/tgt_ptb_sd_test.input -output data/results/dp/tgt_ptb_sd_test.pred -replace_unk -verbose -gpu 0 -beam_size 64 -constraint_length 8 -alpha_c 0.8 -alpha_p 0.8 78 | ``` 79 | 80 | Now you have a model which you can use to predict on new data. We do this by running beam search where `constraint_length`, `alpha_c`, `alpha_p` are parameters used in tree constraints. 81 | 82 | # Notes 83 | You can refer to our paper for more details. Thank you! 84 | 85 | ## Citation 86 | 87 | [Seq2seq Dependency Parsing](./resources/seq2seq4dp.pdf) 88 | 89 | ``` 90 | @inproceedings{li2018seq2seq, 91 | title={Seq2seq dependency parsing}, 92 | author={Li, Zuchao and He, Shexia and Zhao, Hai}, 93 | booktitle={Proceedings of the 27th International Conference on Computational Linguistics (COLING 2018)}, 94 | year={2018} 95 | } 96 | ``` 97 | -------------------------------------------------------------------------------- /data/scripts/eval_subroot.py: -------------------------------------------------------------------------------- 1 | import os 2 | from collections import Counter 3 | 4 | def load_data(path): 5 | with open(path, 'r') as f: 6 | data = f.readlines() 7 | 8 | data = [line.strip().split() for line in data if len(line.strip())>0] 9 | 10 | return data 11 | 12 | 13 | def f1(target, predict): 14 | TP = 0 15 | TN = 0 16 | FP = 0 17 | FN = 0 18 | total = 0 19 | correct = 0 20 | assert len(target) == len(predict) 21 | for i in range(len(target)): 22 | assert len(target[i]) == len(predict[i]) 23 | for j in range(len(target[i])): 24 | total += 1 25 | if target[i][j] == predict[i][j]: 26 | correct += 1 27 | assert predict[i][j] == '0' or predict[i][j] == '1' 28 | if target[i][j] == '1' and target[i][j] == predict[i][j]: 29 | TP += 1 30 | if target[i][j] == '0' and target[i][j] == predict[i][j]: 31 | TN += 1 32 | if target[i][j] == '0' and target[i][j] != predict[i][j]: 33 | FP += 1 34 | if target[i][j] == '1' and target[i][j] != predict[i][j]: 35 | FN += 1 36 | P = TP / (TP + FP) 37 | R = TP / (TP + FN) 38 | F1 = 2 * P * R / (P + R) 39 | 40 | print('eval Acc:{:.2f} P:{:.2f} R:{:.2f} F1:{:.2f}'.format(correct/total*100, P * 100, R * 100, F1 * 100)) 41 | 42 | if __name__ == '__main__': 43 | f1(load_data(os.path.join(os.path.dirname(__file__), '../input/subroot/tgt_ptb_sd_subroot_train.input')), 44 | load_data(os.path.join(os.path.dirname(__file__), '../results/subroot/tgt_ptb_sd_subroot_train.pred'))) 45 | 46 | f1(load_data(os.path.join(os.path.dirname(__file__), '../input/subroot/tgt_ptb_sd_subroot_dev.input')), 47 | load_data(os.path.join(os.path.dirname(__file__), '../results/subroot/tgt_ptb_sd_subroot_dev.pred'))) 48 | 49 | f1(load_data(os.path.join(os.path.dirname(__file__), '../input/subroot/tgt_ptb_sd_subroot_test.input')), 50 | load_data(os.path.join(os.path.dirname(__file__), '../results/subroot/tgt_ptb_sd_subroot_test.pred'))) 51 | -------------------------------------------------------------------------------- /data/scripts/make_dp_dataset.py: -------------------------------------------------------------------------------- 1 | import os 2 | import tqdm 3 | 4 | # def is_scientific_notation(s): 5 | # s = str(s) 6 | # if s.count(',')>=1: 7 | # sl = s.split(',') 8 | # for item in sl: 9 | # if not item.isdigit(): 10 | # return False 11 | # return True 12 | # return False 13 | 14 | # def is_float(s): 15 | # s = str(s) 16 | # if s.count('.')==1: 17 | # sl = s.split('.') 18 | # left = sl[0] 19 | # right = sl[1] 20 | # if left.startswith('-') and left.count('-')==1 and right.isdigit(): 21 | # lleft = left.split('-')[1] 22 | # if lleft.isdigit() or is_scientific_notation(lleft): 23 | # return True 24 | # elif (left.isdigit() or is_scientific_notation(left)) and right.isdigit(): 25 | # return True 26 | # return False 27 | 28 | # def is_fraction(s): 29 | # s = str(s) 30 | # if s.count('\/')==1: 31 | # sl = s.split('\/') 32 | # if len(sl)== 2 and sl[0].isdigit() and sl[1].isdigit(): 33 | # return True 34 | # if s.count('/')==1: 35 | # sl = s.split('/') 36 | # if len(sl)== 2 and sl[0].isdigit() and sl[1].isdigit(): 37 | # return True 38 | # if s[-1]=='%' and len(s)>1: 39 | # return True 40 | # return False 41 | 42 | # def is_number(s): 43 | # s = str(s) 44 | # if s.isdigit() or is_float(s) or is_fraction(s) or is_scientific_notation(s): 45 | # return True 46 | # else: 47 | # return False 48 | 49 | def make_input(file_name, src_path, tgt_path): 50 | with open(file_name, 'r') as f: 51 | data = f.readlines() 52 | 53 | origin_data = [] 54 | sentence = [] 55 | 56 | for i in range(len(data)): 57 | if len(data[i].strip()) > 0: 58 | sentence.append(data[i].strip().split('\t')) 59 | else: 60 | origin_data.append(sentence) 61 | sentence = [] 62 | 63 | if len(sentence) > 0: 64 | origin_data.append(sentence) 65 | 66 | src_data = [] 67 | tgt_data = [] 68 | for sentence in origin_data: 69 | src_line = [] 70 | tgt_line = [] 71 | for line in sentence: 72 | dep_ind = int(line[0]) 73 | head_ind = int(line[6]) 74 | if dep_ind > head_ind: 75 | tag = 'L' + str(abs(dep_ind - head_ind)) 76 | else: 77 | tag = 'R' + str(abs(dep_ind - head_ind)) 78 | # word = ''.join([c if not c.isdigit() else '0' for c in line[1].lower()]) 79 | is_number = False 80 | word = line[1].lower() 81 | for c in word: 82 | if c.isdigit(): 83 | is_number = True 84 | break 85 | if is_number: 86 | word = 'number' 87 | src_line.append([word, line[4]]) 88 | tgt_line.append(tag) 89 | if len(src_line) >= 1: 90 | src_data.append(src_line) 91 | tgt_data.append(tgt_line) 92 | 93 | with open(src_path, 'w') as f: 94 | for line in src_data: 95 | f.write(' '.join(['|'.join(item) for item in line])) 96 | f.write('\n') 97 | 98 | 99 | with open(tgt_path, 'w') as f: 100 | for line in tgt_data: 101 | f.write(' '.join(line)) 102 | f.write('\n') 103 | 104 | if __name__ == '__main__': 105 | train_file = os.path.join(os.path.dirname(__file__), '../ptb-sd/train_pro_wsd.conll') 106 | dev_file = os.path.join(os.path.dirname(__file__), '../ptb-sd/dev_pro.conll') 107 | test_file = os.path.join(os.path.dirname(__file__), '../ptb-sd/test_pro.conll') 108 | 109 | make_input(train_file, os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_train.input'), 110 | os.path.join(os.path.dirname(__file__), '../input/dp/tgt_ptb_sd_train.input')) 111 | make_input(dev_file, os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_dev.input'), 112 | os.path.join(os.path.dirname(__file__), '../input/dp/tgt_ptb_sd_dev.input')) 113 | make_input(test_file, os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_test.input'), 114 | os.path.join(os.path.dirname(__file__), '../input/dp/tgt_ptb_sd_test.input')) 115 | -------------------------------------------------------------------------------- /data/scripts/make_subroot_dataset.py: -------------------------------------------------------------------------------- 1 | import os 2 | import tqdm 3 | 4 | def make_input(file_name, src_path, tgt_path): 5 | with open(file_name, 'r') as f: 6 | data = f.readlines() 7 | 8 | origin_data = [] 9 | sentence = [] 10 | 11 | for i in range(len(data)): 12 | if len(data[i].strip()) > 0: 13 | sentence.append(data[i].strip().split('\t')) 14 | else: 15 | origin_data.append(sentence) 16 | sentence = [] 17 | 18 | if len(sentence) > 0: 19 | origin_data.append(sentence) 20 | 21 | src_data = [] 22 | tgt_data = [] 23 | for sentence in origin_data: 24 | src_line = [] 25 | tgt_line = [] 26 | for line in sentence: 27 | dep_ind = int(line[0]) 28 | head_ind = int(line[6]) 29 | if head_ind == 0: 30 | tag = '1' 31 | else: 32 | tag = '0' 33 | # word = ''.join([c if not c.isdigit() else '0' for c in line[1].lower()]) 34 | is_number = False 35 | word = line[1].lower() 36 | for c in word: 37 | if c.isdigit(): 38 | is_number = True 39 | break 40 | if is_number: 41 | word = 'number' 42 | src_line.append([word, line[4]]) 43 | tgt_line.append(tag) 44 | if len(src_line) > 1: 45 | src_data.append(src_line) 46 | tgt_data.append(tgt_line) 47 | 48 | with open(src_path, 'w') as f: 49 | for line in src_data: 50 | f.write(' '.join(['|'.join(item) for item in line])) 51 | f.write('\n') 52 | 53 | 54 | with open(tgt_path, 'w') as f: 55 | for line in tgt_data: 56 | f.write(' '.join(line)) 57 | f.write('\n') 58 | 59 | if __name__ == '__main__': 60 | train_file = os.path.join(os.path.dirname(__file__), '../ptb-sd/train_pro_wsd.conll') 61 | dev_file = os.path.join(os.path.dirname(__file__), '../ptb-sd/dev_pro.conll') 62 | test_file = os.path.join(os.path.dirname(__file__), '../ptb-sd/test_pro.conll') 63 | 64 | make_input(train_file, 65 | os.path.join(os.path.dirname(__file__), '../input/subroot/src_ptb_sd_subroot_train.input'), 66 | os.path.join(os.path.dirname(__file__), '../input/subroot/tgt_ptb_sd_subroot_train.input')) 67 | 68 | make_input(dev_file, 69 | os.path.join(os.path.dirname(__file__), '../input/subroot/src_ptb_sd_subroot_dev.input'), 70 | os.path.join(os.path.dirname(__file__), '../input/subroot/tgt_ptb_sd_subroot_dev.input')) 71 | 72 | make_input(test_file, 73 | os.path.join(os.path.dirname(__file__), '../input/subroot/src_ptb_sd_subroot_test.input'), 74 | os.path.join(os.path.dirname(__file__), '../input/subroot/tgt_ptb_sd_subroot_test.input')) 75 | -------------------------------------------------------------------------------- /data/scripts/make_wsd_pred.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | def make_wsd_pred(pred_file, map_file, output_file): 4 | with open(pred_file, 'r') as f: 5 | pred_data = f.readlines() 6 | 7 | pred_data = [line.split() for line in pred_data if len(line.strip())>0] 8 | 9 | 10 | with open(map_file, 'r') as f: 11 | map_data = f.readlines() 12 | 13 | map_data = [line.strip() for line in map_data if len(line.strip())>0] 14 | 15 | output_data = [] 16 | 17 | assert len(map_data) == len(pred_data) 18 | 19 | sent_len = len(map_data) 20 | sent_line = [] 21 | for i in range(sent_len): 22 | if len(sent_line) == 0: 23 | sent_line = pred_data[i] 24 | else: 25 | if map_data[i] == map_data[i-1]: 26 | sent_line[-1] = '' 27 | sent_line += pred_data[i][1:] 28 | else: 29 | output_data.append(sent_line) 30 | sent_line = pred_data[i] 31 | 32 | if len(sent_line)>0: 33 | output_data.append(sent_line) 34 | 35 | with open(output_file, 'w') as f: 36 | for i in range(len(output_data)): 37 | for j in range(len(output_data[i])): 38 | if output_data[i][j] == '': 39 | output_data[i][j] = 'L'+str(j+1) 40 | f.write(' '.join(output_data[i])) 41 | f.write('\n') 42 | 43 | 44 | if __name__ == '__main__': 45 | # make_wsd_pred(os.path.join(os.path.dirname(__file__), '../results/dp/tgt_ptb_sd_dev_wsd_30.pred'), 46 | # os.path.join(os.path.dirname(__file__), '../input/dp/tgt_ptb_sd_dev_wsd_30_map.input')) 47 | 48 | make_wsd_pred(os.path.join(os.path.dirname(__file__), '../results/dp/tgt_ptb_sd_test_wsd_40.pred'), 49 | os.path.join(os.path.dirname(__file__), '../input/dp/tgt_ptb_sd_test_wsd_40_map.input'), 50 | os.path.join(os.path.dirname(__file__), '../results/dp/tgt_ptb_sd_test_wsd_40_org.pred')) 51 | 52 | -------------------------------------------------------------------------------- /data/scripts/mergy_subroot_feature.py: -------------------------------------------------------------------------------- 1 | # we merge the golden subroot feature into train dataset for training 2 | # and we use the predict subroot (by BiLSTM+CRF) feature into dev/train dataset 3 | 4 | import os 5 | 6 | def merge_train(input_file, origin_file, output_file): 7 | with open(input_file, 'r') as f: 8 | input_data = f.readlines() 9 | 10 | input_data = [line.split() for line in input_data if len(line.strip())>0] 11 | 12 | with open(origin_file, 'r') as f: 13 | data = f.readlines() 14 | 15 | origin_data = [] 16 | sentence = [] 17 | 18 | for i in range(len(data)): 19 | if len(data[i].strip()) > 0: 20 | sentence.append(data[i].strip().split('\t')) 21 | else: 22 | origin_data.append(sentence) 23 | sentence = [] 24 | 25 | if len(sentence) > 0: 26 | origin_data.append(sentence) 27 | 28 | assert len(input_data) == len(origin_data) 29 | 30 | with open(output_file, 'w') as f: 31 | for i in range(len(input_data)): 32 | assert len(input_data[i]) == len(origin_data[i]) 33 | line = [] 34 | for j in range(len(input_data[i])): 35 | if int(origin_data[i][j][6]) == 0: 36 | line.append(input_data[i][j]+'|1') 37 | else: 38 | line.append(input_data[i][j]+'|0') 39 | f.write(' '.join(line)) 40 | f.write('\n') 41 | 42 | 43 | def merge_pred(input_file, subroot_pred_file, output_file): 44 | with open(input_file, 'r') as f: 45 | input_data = f.readlines() 46 | 47 | input_data = [line.split() for line in input_data if len(line.strip())>0] 48 | 49 | with open(subroot_pred_file, 'r') as f: 50 | data = f.readlines() 51 | 52 | pred_data = [] 53 | sentence = [] 54 | 55 | for i in range(len(data)): 56 | if len(data[i].strip()) > 0: 57 | sentence.append(data[i].strip().split('\t')) 58 | else: 59 | pred_data.append(sentence) 60 | sentence = [] 61 | 62 | if len(sentence) > 0: 63 | pred_data.append(sentence) 64 | 65 | assert len(input_data) == len(pred_data) 66 | 67 | with open(output_file, 'w') as f: 68 | for i in range(len(input_data)): 69 | assert len(input_data[i]) == len(pred_data[i]) 70 | line = [] 71 | for j in range(len(input_data[i])): 72 | line.append(input_data[i][j]+'|'+pred_data[i][j][1]) 73 | f.write(' '.join(line)) 74 | f.write('\n') 75 | 76 | 77 | if __name__ == '__main__': 78 | merge_train(os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_train.input'), 79 | os.path.join(os.path.dirname(__file__), '../ptb-sd/train_pro.conll'), 80 | os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_train_ws.input')) 81 | 82 | merge_pred(os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_dev.input'), 83 | os.path.join(os.path.dirname(__file__), '../../subroot/result/dev_predicate_95.94.pred'), 84 | os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_dev_ws.input')) 85 | 86 | merge_pred(os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_test.input'), 87 | os.path.join(os.path.dirname(__file__), '../../subroot/result/test_predicate_95.16.pred'), 88 | os.path.join(os.path.dirname(__file__), '../input/dp/src_ptb_sd_test_ws.input')) -------------------------------------------------------------------------------- /data/scripts/recover_dataset.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | 4 | def recover_data(file_name, pred_data, output_path): 5 | with open(file_name, 'r') as f: 6 | data = f.readlines() 7 | 8 | 9 | golden_data = [] 10 | sentence = [] 11 | 12 | for i in range(len(data)): 13 | if len(data[i].strip()) > 0: 14 | sentence.append(data[i].strip().split('\t')) 15 | else: 16 | golden_data.append(sentence) 17 | sentence = [] 18 | 19 | if len(sentence) > 0: 20 | golden_data.append(sentence) 21 | 22 | with open(pred_data, 'r') as f: 23 | data = f.readlines() 24 | 25 | pred_data = [item.strip().split() for item in data if len(item.strip()) > 0] 26 | 27 | pred_index = 0 28 | for i in range(len(golden_data)): 29 | predicate_idx = 0 30 | for j in range(len(golden_data[i])): 31 | if golden_data[i][j][12] == 'Y': 32 | predicate_idx += 1 33 | for k in range(len(golden_data[i])): 34 | golden_data[i][k][13 + predicate_idx] = pred_data[pred_index][k] 35 | pred_index += 1 36 | 37 | with open(output_path, 'w') as f: 38 | for sentence in golden_data: 39 | for line in sentence: 40 | f.write('\t'.join(line)) 41 | f.write('\n') 42 | f.write('\n') 43 | 44 | if __name__ == '__main__': 45 | recover_data(os.path.join(os.path.dirname(__file__), 'conll09-english/conll09_test.dataset'), 46 | os.path.join(os.path.dirname(__file__), 'tgt_conll09_en_test.pred'), 47 | os.path.join(os.path.dirname(__file__), 'conll09_en_test.dataset.pred')) 48 | -------------------------------------------------------------------------------- /data/scripts/subroot_replace_unk.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | def replace_unk(input_file): 4 | with open(input_file, 'r') as f: 5 | input_data = f.readlines() 6 | 7 | input_data = [line.split() for line in input_data if len(line.strip())>0] 8 | 9 | with open(input_file, 'w') as f: 10 | for i in range(len(input_data)): 11 | line = [] 12 | for j in range(len(input_data[i])): 13 | if input_data[i][j] == '0' or input_data[i][j]=='1': 14 | line.append(input_data[i][j]) 15 | else: 16 | line.append('0') 17 | f.write(' '.join(line)) 18 | f.write('\n') 19 | 20 | 21 | if __name__ == '__main__': 22 | replace_unk(os.path.join(os.path.dirname(__file__), '../results/subroot/tgt_ptb_sd_subroot_train.pred')) 23 | 24 | replace_unk(os.path.join(os.path.dirname(__file__), '../results/subroot/tgt_ptb_sd_subroot_dev.pred')) 25 | 26 | replace_unk(os.path.join(os.path.dirname(__file__), '../results/subroot/tgt_ptb_sd_subroot_test.pred')) -------------------------------------------------------------------------------- /node2vec/train.py: -------------------------------------------------------------------------------- 1 | import networkx as nx 2 | from node2vec import Node2Vec 3 | 4 | # FILES 5 | EMBEDDING_FILENAME = './node2vec_en.emb' 6 | EMBEDDING_MODEL_FILENAME = './node2vec_en.model' 7 | 8 | # Create a graph 9 | # graph = nx.fast_gnp_random_graph(n=100, p=0.5) 10 | graph = nx.Graph() 11 | 12 | raw_train_file = '../data/ptb-sd/train_pro.conll' 13 | 14 | with open(raw_train_file, 'r') as f: 15 | data = f.readlines() 16 | 17 | # read data 18 | train_data = [] 19 | sentence = [] 20 | for line in data: 21 | if len(line.strip()) > 0: 22 | line = line.strip().split('\t') 23 | sentence.append(line) 24 | else: 25 | train_data.append(sentence) 26 | sentence = [] 27 | if len(sentence)>0: 28 | train_data.append(sentence) 29 | sentence = [] 30 | 31 | for sentence in train_data: 32 | for line in sentence: 33 | head_idx = int(line[6])-1 34 | if head_idx == -1: 35 | is_number = False 36 | word = line[1].lower() 37 | for c in word: 38 | if c.isdigit(): 39 | is_number = True 40 | break 41 | if is_number: 42 | word = 'number' 43 | graph.add_edge('', word, weight=1) 44 | else: 45 | hw = sentence[head_idx][1].lower() 46 | is_number = False 47 | for c in hw: 48 | if c.isdigit(): 49 | is_number = True 50 | break 51 | if is_number: 52 | hw = 'number' 53 | w = line[1].lower() 54 | is_number = False 55 | for c in w: 56 | if c.isdigit(): 57 | is_number = True 58 | break 59 | if is_number: 60 | w = 'number' 61 | graph.add_edge(hw, w, weight=0.5) 62 | 63 | # Precompute probabilities and generate walks 64 | node2vec = Node2Vec(graph, dimensions=100, walk_length=100, num_walks=18, workers=1) 65 | 66 | # Embed 67 | model = node2vec.fit(window=16, min_count=1, batch_words=64) # Any keywords acceptable by gensim.Word2Vec can be passed, `diemnsions` and `workers` are automatically passed (from the Node2Vec constructor) 68 | 69 | # Look for most similar nodes 70 | model.wv.most_similar('') # Output node names are always strings 71 | 72 | # Save embeddings for later use 73 | model.wv.save_word2vec_format(EMBEDDING_FILENAME) 74 | 75 | # Save model for later use 76 | model.save(EMBEDDING_MODEL_FILENAME) -------------------------------------------------------------------------------- /onmt/Highway.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | class HighwayMLP(nn.Module): 6 | 7 | def __init__(self, 8 | input_size, 9 | gate_bias=-2, 10 | activation_function=nn.functional.relu, 11 | gate_activation=nn.functional.softmax): 12 | 13 | super(HighwayMLP, self).__init__() 14 | 15 | self.activation_function = activation_function 16 | self.gate_activation = gate_activation 17 | 18 | self.normal_layer = nn.Linear(input_size, input_size) 19 | 20 | self.gate_layer = nn.Linear(input_size, input_size) 21 | self.gate_layer.bias.data.fill_(gate_bias) 22 | 23 | def forward(self, x): 24 | 25 | normal_layer_result = self.activation_function(self.normal_layer(x)) 26 | gate_layer_result = self.gate_activation(self.gate_layer(x),dim=0) 27 | 28 | multiplyed_gate_and_normal = torch.mul(normal_layer_result, gate_layer_result) 29 | multiplyed_gate_and_input = torch.mul((1 - gate_layer_result), x) 30 | 31 | return torch.add(multiplyed_gate_and_normal, 32 | multiplyed_gate_and_input) 33 | 34 | 35 | class HighwayCNN(nn.Module): 36 | def __init__(self, 37 | input_size, 38 | gate_bias=-1, 39 | activation_function=nn.functional.relu, 40 | gate_activation=nn.functional.softmax): 41 | 42 | super(HighwayCNN, self).__init__() 43 | 44 | self.activation_function = activation_function 45 | self.gate_activation = gate_activation 46 | 47 | self.normal_layer = nn.Linear(input_size, input_size) 48 | 49 | self.gate_layer = nn.Linear(input_size, input_size) 50 | self.gate_layer.bias.data.fill_(gate_bias) 51 | 52 | def forward(self, x): 53 | 54 | normal_layer_result = self.activation_function(self.normal_layer(x)) 55 | gate_layer_result = self.gate_activation(self.gate_layer(x)) 56 | 57 | multiplyed_gate_and_normal = torch.mul(normal_layer_result, gate_layer_result) 58 | multiplyed_gate_and_input = torch.mul((1 - gate_layer_result), x) 59 | 60 | return torch.add(multiplyed_gate_and_normal, 61 | multiplyed_gate_and_input) -------------------------------------------------------------------------------- /onmt/Optim.py: -------------------------------------------------------------------------------- 1 | import torch.optim as optim 2 | from torch.nn.utils import clip_grad_norm 3 | 4 | 5 | class MultipleOptimizer(object): 6 | def __init__(self, op): 7 | self.optimizers = op 8 | 9 | def zero_grad(self): 10 | for op in self.optimizers: 11 | op.zero_grad() 12 | 13 | def step(self): 14 | for op in self.optimizers: 15 | op.step() 16 | 17 | 18 | class Optim(object): 19 | """ 20 | Controller class for optimization. Mostly a thin 21 | wrapper for `optim`, but also useful for implementing 22 | rate scheduling beyond what is currently available. 23 | Also implements necessary methods for training RNNs such 24 | as grad manipulations. 25 | 26 | Args: 27 | method (:obj:`str`): one of [sgd, adagrad, adadelta, adam] 28 | lr (float): learning rate 29 | lr_decay (float, optional): learning rate decay multiplier 30 | start_decay_at (int, optional): epoch to start learning rate decay 31 | beta1, beta2 (float, optional): parameters for adam 32 | adagrad_accum (float, optional): initialization parameter for adagrad 33 | decay_method (str, option): custom decay options 34 | warmup_steps (int, option): parameter for `noam` decay 35 | model_size (int, option): parameter for `noam` decay 36 | """ 37 | # We use the default parameters for Adam that are suggested by 38 | # the original paper https://arxiv.org/pdf/1412.6980.pdf 39 | # These values are also used by other established implementations, 40 | # e.g. https://www.tensorflow.org/api_docs/python/tf/train/AdamOptimizer 41 | # https://keras.io/optimizers/ 42 | # Recently there are slightly different values used in the paper 43 | # "Attention is all you need" 44 | # https://arxiv.org/pdf/1706.03762.pdf, particularly the value beta2=0.98 45 | # was used there however, beta2=0.999 is still arguably the more 46 | # established value, so we use that here as well 47 | def __init__(self, method, lr, max_grad_norm, 48 | lr_decay=1, start_decay_at=None, 49 | beta1=0.9, beta2=0.999, 50 | adagrad_accum=0.0, 51 | decay_method=None, 52 | warmup_steps=4000, 53 | model_size=None): 54 | self.last_ppl = None 55 | self.lr = lr 56 | self.original_lr = lr 57 | self.max_grad_norm = max_grad_norm 58 | self.method = method 59 | self.lr_decay = lr_decay 60 | self.start_decay_at = start_decay_at 61 | self.start_decay = False 62 | self._step = 0 63 | self.betas = [beta1, beta2] 64 | self.adagrad_accum = adagrad_accum 65 | self.decay_method = decay_method 66 | self.warmup_steps = warmup_steps 67 | self.model_size = model_size 68 | 69 | def set_parameters(self, params): 70 | self.params = [] 71 | self.sparse_params = [] 72 | for k, p in params: 73 | if p.requires_grad: 74 | if self.method != 'sparseadam' or "embed" not in k: 75 | self.params.append(p) 76 | else: 77 | self.sparse_params.append(p) 78 | if self.method == 'sgd': 79 | self.optimizer = optim.SGD(self.params, lr=self.lr) 80 | elif self.method == 'adagrad': 81 | self.optimizer = optim.Adagrad(self.params, lr=self.lr) 82 | for group in self.optimizer.param_groups: 83 | for p in group['params']: 84 | self.optimizer.state[p]['sum'] = self.optimizer\ 85 | .state[p]['sum'].fill_(self.adagrad_accum) 86 | elif self.method == 'adadelta': 87 | self.optimizer = optim.Adadelta(self.params, lr=self.lr) 88 | elif self.method == 'adam': 89 | self.optimizer = optim.Adam(self.params, lr=self.lr, 90 | betas=self.betas, eps=1e-9) 91 | elif self.method == 'sparseadam': 92 | self.optimizer = MultipleOptimizer( 93 | [optim.Adam(self.params, lr=self.lr, 94 | betas=self.betas, eps=1e-8), 95 | optim.SparseAdam(self.sparse_params, lr=self.lr, 96 | betas=self.betas, eps=1e-8)]) 97 | else: 98 | raise RuntimeError("Invalid optim method: " + self.method) 99 | 100 | def _set_rate(self, lr): 101 | self.lr = lr 102 | if self.method != 'sparseadam': 103 | self.optimizer.param_groups[0]['lr'] = self.lr 104 | else: 105 | for op in self.optimizer.optimizers: 106 | op.param_groups[0]['lr'] = self.lr 107 | 108 | def step(self): 109 | """Update the model parameters based on current gradients. 110 | 111 | Optionally, will employ gradient modification or update learning 112 | rate. 113 | """ 114 | self._step += 1 115 | 116 | # Decay method used in tensor2tensor. 117 | if self.decay_method == "noam": 118 | self._set_rate( 119 | self.original_lr * 120 | (self.model_size ** (-0.5) * 121 | min(self._step ** (-0.5), 122 | self._step * self.warmup_steps**(-1.5)))) 123 | 124 | if self.max_grad_norm: 125 | clip_grad_norm(self.params, self.max_grad_norm) 126 | self.optimizer.step() 127 | 128 | def update_learning_rate(self, ppl, epoch): 129 | """ 130 | Decay learning rate if val perf does not improve 131 | or we hit the start_decay_at limit. 132 | """ 133 | 134 | if self.start_decay_at is not None and epoch >= self.start_decay_at: 135 | self.start_decay = True 136 | if self.last_ppl is not None and ppl > self.last_ppl: 137 | self.start_decay = True 138 | 139 | if self.start_decay: 140 | self.lr = self.lr * self.lr_decay 141 | print("Decaying learning rate to %g" % self.lr) 142 | 143 | self.last_ppl = ppl 144 | if self.method != 'sparseadam': 145 | self.optimizer.param_groups[0]['lr'] = self.lr 146 | -------------------------------------------------------------------------------- /onmt/Utils.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | def aeq(*args): 5 | """ 6 | Assert all arguments have the same value 7 | """ 8 | arguments = (arg for arg in args) 9 | first = next(arguments) 10 | assert all(arg == first for arg in arguments), \ 11 | "Not all arguments have the same value: " + str(args) 12 | 13 | 14 | def sequence_mask(lengths, max_len=None): 15 | """ 16 | Creates a boolean mask from sequence lengths. 17 | """ 18 | batch_size = lengths.numel() 19 | max_len = max_len or lengths.max() 20 | return (torch.arange(0, max_len) 21 | .type_as(lengths) 22 | .repeat(batch_size, 1) 23 | .lt(lengths.unsqueeze(1))) 24 | 25 | 26 | def use_gpu(opt): 27 | return (hasattr(opt, 'gpuid') and len(opt.gpuid) > 0) or \ 28 | (hasattr(opt, 'gpu') and opt.gpu > -1) 29 | -------------------------------------------------------------------------------- /onmt/__init__.py: -------------------------------------------------------------------------------- 1 | import onmt.io 2 | import onmt.Models 3 | import onmt.Loss 4 | import onmt.translate 5 | import onmt.opts 6 | from onmt.Trainer import Trainer, Statistics 7 | from onmt.Optim import Optim 8 | 9 | # For flake8 compatibility 10 | __all__ = [onmt.Loss, onmt.Models, onmt.opts, 11 | Trainer, Optim, Statistics, onmt.io, onmt.translate] 12 | -------------------------------------------------------------------------------- /onmt/__init__.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/__init__.pyc -------------------------------------------------------------------------------- /onmt/io/DatasetBase.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | 3 | from itertools import chain 4 | import torchtext 5 | 6 | 7 | PAD_WORD = '' 8 | UNK_WORD = '' 9 | UNK = 0 10 | BOS_WORD = '' 11 | EOS_WORD = '' 12 | 13 | 14 | class ONMTDatasetBase(torchtext.data.Dataset): 15 | """ 16 | A dataset basically supports iteration over all the examples 17 | it contains. We currently have 3 datasets inheriting this base 18 | for 3 types of corpus respectively: "text", "img", "audio". 19 | 20 | Internally it initializes an `torchtext.data.Dataset` object with 21 | the following attributes: 22 | 23 | `examples`: a sequence of `torchtext.data.Example` objects. 24 | `fields`: a dictionary associating str keys with `torchtext.data.Field` 25 | objects, and not necessarily having the same keys as the input fields. 26 | """ 27 | def __getstate__(self): 28 | return self.__dict__ 29 | 30 | def __setstate__(self, d): 31 | self.__dict__.update(d) 32 | 33 | def __reduce_ex__(self, proto): 34 | "This is a hack. Something is broken with torch pickle." 35 | return super(ONMTDatasetBase, self).__reduce_ex__() 36 | 37 | def load_fields(self, vocab_dict): 38 | """ Load fields from vocab.pt, and set the `fields` attribute. 39 | 40 | Args: 41 | vocab_dict (dict): a dict of loaded vocab from vocab.pt file. 42 | """ 43 | from onmt.io.IO import load_fields_from_vocab 44 | 45 | fields = load_fields_from_vocab(vocab_dict.items(), self.data_type) 46 | self.fields = dict([(k, f) for (k, f) in fields.items() 47 | if k in self.examples[0].__dict__]) 48 | 49 | @staticmethod 50 | def extract_text_features(tokens): 51 | """ 52 | Args: 53 | tokens: A list of tokens, where each token consists of a word, 54 | optionally followed by u"￨"-delimited features. 55 | Returns: 56 | A sequence of words, a sequence of features, and num of features. 57 | """ 58 | if not tokens: 59 | return [], [], -1 60 | 61 | split_tokens = [token.split('|') for token in tokens] # u"￨" 62 | split_tokens = [token for token in split_tokens if token[0]] 63 | token_size = len(split_tokens[0]) 64 | 65 | assert all(len(token) == token_size for token in split_tokens), \ 66 | "all words must have the same number of features" 67 | words_and_features = list(zip(*split_tokens)) 68 | words = words_and_features[0] 69 | features = words_and_features[1:] 70 | 71 | return words, features, token_size - 1 72 | 73 | # Below are helper functions for intra-class use only. 74 | 75 | def _join_dicts(self, *args): 76 | """ 77 | Args: 78 | dictionaries with disjoint keys. 79 | 80 | Returns: 81 | a single dictionary that has the union of these keys. 82 | """ 83 | return dict(chain(*[d.items() for d in args])) 84 | 85 | def _peek(self, seq): 86 | """ 87 | Args: 88 | seq: an iterator. 89 | 90 | Returns: 91 | the first thing returned by calling next() on the iterator 92 | and an iterator created by re-chaining that value to the beginning 93 | of the iterator. 94 | """ 95 | first = next(seq) 96 | return first, chain([first], seq) 97 | 98 | def _construct_example_fromlist(self, data, fields): 99 | """ 100 | Args: 101 | data: the data to be set as the value of the attributes of 102 | the to-be-created `Example`, associating with respective 103 | `Field` objects with same key. 104 | fields: a dict of `torchtext.data.Field` objects. The keys 105 | are attributes of the to-be-created `Example`. 106 | 107 | Returns: 108 | the created `Example` object. 109 | """ 110 | ex = torchtext.data.Example() 111 | for (name, field), val in zip(fields, data): 112 | if field is not None: 113 | setattr(ex, name, field.preprocess(val)) 114 | else: 115 | setattr(ex, name, val) 116 | return ex 117 | -------------------------------------------------------------------------------- /onmt/io/IO.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/io/IO.pyc -------------------------------------------------------------------------------- /onmt/io/__init__.py: -------------------------------------------------------------------------------- 1 | from onmt.io.IO import collect_feature_vocabs, make_features, \ 2 | collect_features, get_num_features, \ 3 | load_fields_from_vocab, get_fields, \ 4 | save_fields_to_vocab, build_dataset, \ 5 | build_vocab, merge_vocabs, OrderedIterator 6 | from onmt.io.DatasetBase import ONMTDatasetBase, PAD_WORD, BOS_WORD, \ 7 | EOS_WORD, UNK 8 | from onmt.io.TextDataset import TextDataset, ShardedTextCorpusIterator 9 | from onmt.io.ImageDataset import ImageDataset 10 | from onmt.io.AudioDataset import AudioDataset 11 | 12 | 13 | __all__ = [PAD_WORD, BOS_WORD, EOS_WORD, UNK, ONMTDatasetBase, 14 | collect_feature_vocabs, make_features, 15 | collect_features, get_num_features, 16 | load_fields_from_vocab, get_fields, 17 | save_fields_to_vocab, build_dataset, 18 | build_vocab, merge_vocabs, OrderedIterator, 19 | TextDataset, ImageDataset, AudioDataset, 20 | ShardedTextCorpusIterator] 21 | -------------------------------------------------------------------------------- /onmt/io/__init__.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/io/__init__.pyc -------------------------------------------------------------------------------- /onmt/io/__pycache__/AudioDataset.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/io/__pycache__/AudioDataset.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/io/__pycache__/DatasetBase.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/io/__pycache__/DatasetBase.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/io/__pycache__/IO.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/io/__pycache__/IO.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/io/__pycache__/ImageDataset.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/io/__pycache__/ImageDataset.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/io/__pycache__/TextDataset.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/io/__pycache__/TextDataset.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/io/__pycache__/__init__.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/io/__pycache__/__init__.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/AudioEncoder.py: -------------------------------------------------------------------------------- 1 | import math 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | 6 | class AudioEncoder(nn.Module): 7 | """ 8 | A simple encoder convolutional -> recurrent neural network for 9 | audio input. 10 | 11 | Args: 12 | num_layers (int): number of encoder layers. 13 | bidirectional (bool): bidirectional encoder. 14 | rnn_size (int): size of hidden states of the rnn. 15 | dropout (float): dropout probablity. 16 | sample_rate (float): input spec 17 | window_size (int): input spec 18 | 19 | """ 20 | def __init__(self, num_layers, bidirectional, rnn_size, dropout, 21 | sample_rate, window_size): 22 | super(AudioEncoder, self).__init__() 23 | self.num_layers = num_layers 24 | self.num_directions = 2 if bidirectional else 1 25 | self.hidden_size = rnn_size 26 | 27 | self.layer1 = nn.Conv2d(1, 32, kernel_size=(41, 11), 28 | padding=(0, 10), stride=(2, 2)) 29 | self.batch_norm1 = nn.BatchNorm2d(32) 30 | self.layer2 = nn.Conv2d(32, 32, kernel_size=(21, 11), 31 | padding=(0, 0), stride=(2, 1)) 32 | self.batch_norm2 = nn.BatchNorm2d(32) 33 | 34 | input_size = int(math.floor((sample_rate * window_size) / 2) + 1) 35 | input_size = int(math.floor(input_size - 41) / 2 + 1) 36 | input_size = int(math.floor(input_size - 21) / 2 + 1) 37 | input_size *= 32 38 | self.rnn = nn.LSTM(input_size, rnn_size, 39 | num_layers=num_layers, 40 | dropout=dropout, 41 | bidirectional=bidirectional) 42 | 43 | def load_pretrained_vectors(self, opt): 44 | # Pass in needed options only when modify function definition. 45 | pass 46 | 47 | def forward(self, input, lengths=None): 48 | "See :obj:`onmt.modules.EncoderBase.forward()`" 49 | # (batch_size, 1, nfft, t) 50 | # layer 1 51 | input = self.batch_norm1(self.layer1(input[:, :, :, :])) 52 | 53 | # (batch_size, 32, nfft/2, t/2) 54 | input = F.hardtanh(input, 0, 20, inplace=True) 55 | 56 | # (batch_size, 32, nfft/2/2, t/2) 57 | # layer 2 58 | input = self.batch_norm2(self.layer2(input)) 59 | 60 | # (batch_size, 32, nfft/2/2, t/2) 61 | input = F.hardtanh(input, 0, 20, inplace=True) 62 | 63 | batch_size = input.size(0) 64 | length = input.size(3) 65 | input = input.view(batch_size, -1, length) 66 | input = input.transpose(0, 2).transpose(1, 2) 67 | 68 | output, hidden = self.rnn(input) 69 | 70 | return hidden, output 71 | -------------------------------------------------------------------------------- /onmt/modules/ConvMultiStepAttention.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from onmt.Utils import aeq 5 | 6 | 7 | SCALE_WEIGHT = 0.5 ** 0.5 8 | 9 | 10 | def seq_linear(linear, x): 11 | # linear transform for 3-d tensor 12 | batch, hidden_size, length, _ = x.size() 13 | h = linear(torch.transpose(x, 1, 2).contiguous().view( 14 | batch * length, hidden_size)) 15 | return torch.transpose(h.view(batch, length, hidden_size, 1), 1, 2) 16 | 17 | 18 | class ConvMultiStepAttention(nn.Module): 19 | """ 20 | 21 | Conv attention takes a key matrix, a value matrix and a query vector. 22 | Attention weight is calculated by key matrix with the query vector 23 | and sum on the value matrix. And the same operation is applied 24 | in each decode conv layer. 25 | 26 | """ 27 | 28 | def __init__(self, input_size): 29 | super(ConvMultiStepAttention, self).__init__() 30 | self.linear_in = nn.Linear(input_size, input_size) 31 | self.mask = None 32 | 33 | def apply_mask(self, mask): 34 | self.mask = mask 35 | 36 | def forward(self, base_target_emb, input, encoder_out_top, 37 | encoder_out_combine): 38 | """ 39 | Args: 40 | base_target_emb: target emb tensor 41 | input: output of decode conv 42 | encoder_out_t: the key matrix for calculation of attetion weight, 43 | which is the top output of encode conv 44 | encoder_out_combine: 45 | the value matrix for the attention-weighted sum, 46 | which is the combination of base emb and top output of encode 47 | 48 | """ 49 | # checks 50 | batch, channel, height, width = base_target_emb.size() 51 | batch_, channel_, height_, width_ = input.size() 52 | aeq(batch, batch_) 53 | aeq(height, height_) 54 | 55 | enc_batch, enc_channel, enc_height = encoder_out_top.size() 56 | enc_batch_, enc_channel_, enc_height_ = encoder_out_combine.size() 57 | 58 | aeq(enc_batch, enc_batch_) 59 | aeq(enc_height, enc_height_) 60 | 61 | preatt = seq_linear(self.linear_in, input) 62 | target = (base_target_emb + preatt) * SCALE_WEIGHT 63 | target = torch.squeeze(target, 3) 64 | target = torch.transpose(target, 1, 2) 65 | pre_attn = torch.bmm(target, encoder_out_top) 66 | 67 | if self.mask is not None: 68 | pre_attn.data.masked_fill_(self.mask, -float('inf')) 69 | 70 | pre_attn = pre_attn.transpose(0, 2) 71 | attn = F.softmax(pre_attn) 72 | attn = attn.transpose(0, 2).contiguous() 73 | context_output = torch.bmm( 74 | attn, torch.transpose(encoder_out_combine, 1, 2)) 75 | context_output = torch.transpose( 76 | torch.unsqueeze(context_output, 3), 1, 2) 77 | return context_output, attn 78 | -------------------------------------------------------------------------------- /onmt/modules/Gate.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | def context_gate_factory(type, embeddings_size, decoder_size, 6 | attention_size, output_size): 7 | """Returns the correct ContextGate class""" 8 | 9 | gate_types = {'source': SourceContextGate, 10 | 'target': TargetContextGate, 11 | 'both': BothContextGate} 12 | 13 | assert type in gate_types, "Not valid ContextGate type: {0}".format(type) 14 | return gate_types[type](embeddings_size, decoder_size, attention_size, 15 | output_size) 16 | 17 | 18 | class ContextGate(nn.Module): 19 | """ 20 | Context gate is a decoder module that takes as input the previous word 21 | embedding, the current decoder state and the attention state, and 22 | produces a gate. 23 | The gate can be used to select the input from the target side context 24 | (decoder state), from the source context (attention state) or both. 25 | """ 26 | def __init__(self, embeddings_size, decoder_size, 27 | attention_size, output_size): 28 | super(ContextGate, self).__init__() 29 | input_size = embeddings_size + decoder_size + attention_size 30 | self.gate = nn.Linear(input_size, output_size, bias=True) 31 | self.sig = nn.Sigmoid() 32 | self.source_proj = nn.Linear(attention_size, output_size) 33 | self.target_proj = nn.Linear(embeddings_size + decoder_size, 34 | output_size) 35 | 36 | def forward(self, prev_emb, dec_state, attn_state): 37 | input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1) 38 | z = self.sig(self.gate(input_tensor)) 39 | proj_source = self.source_proj(attn_state) 40 | proj_target = self.target_proj( 41 | torch.cat((prev_emb, dec_state), dim=1)) 42 | return z, proj_source, proj_target 43 | 44 | 45 | class SourceContextGate(nn.Module): 46 | """Apply the context gate only to the source context""" 47 | 48 | def __init__(self, embeddings_size, decoder_size, 49 | attention_size, output_size): 50 | super(SourceContextGate, self).__init__() 51 | self.context_gate = ContextGate(embeddings_size, decoder_size, 52 | attention_size, output_size) 53 | self.tanh = nn.Tanh() 54 | 55 | def forward(self, prev_emb, dec_state, attn_state): 56 | z, source, target = self.context_gate( 57 | prev_emb, dec_state, attn_state) 58 | return self.tanh(target + z * source) 59 | 60 | 61 | class TargetContextGate(nn.Module): 62 | """Apply the context gate only to the target context""" 63 | 64 | def __init__(self, embeddings_size, decoder_size, 65 | attention_size, output_size): 66 | super(TargetContextGate, self).__init__() 67 | self.context_gate = ContextGate(embeddings_size, decoder_size, 68 | attention_size, output_size) 69 | self.tanh = nn.Tanh() 70 | 71 | def forward(self, prev_emb, dec_state, attn_state): 72 | z, source, target = self.context_gate(prev_emb, dec_state, attn_state) 73 | return self.tanh(z * target + source) 74 | 75 | 76 | class BothContextGate(nn.Module): 77 | """Apply the context gate to both contexts""" 78 | 79 | def __init__(self, embeddings_size, decoder_size, 80 | attention_size, output_size): 81 | super(BothContextGate, self).__init__() 82 | self.context_gate = ContextGate(embeddings_size, decoder_size, 83 | attention_size, output_size) 84 | self.tanh = nn.Tanh() 85 | 86 | def forward(self, prev_emb, dec_state, attn_state): 87 | z, source, target = self.context_gate(prev_emb, dec_state, attn_state) 88 | return self.tanh((1. - z) * target + z * source) 89 | -------------------------------------------------------------------------------- /onmt/modules/GlobalAttention.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | from onmt.Utils import aeq, sequence_mask 5 | 6 | 7 | class GlobalAttention(nn.Module): 8 | """ 9 | Global attention takes a matrix and a query vector. It 10 | then computes a parameterized convex combination of the matrix 11 | based on the input query. 12 | 13 | Constructs a unit mapping a query `q` of size `dim` 14 | and a source matrix `H` of size `n x dim`, to an output 15 | of size `dim`. 16 | 17 | 18 | .. mermaid:: 19 | 20 | graph BT 21 | A[Query] 22 | subgraph RNN 23 | C[H 1] 24 | D[H 2] 25 | E[H N] 26 | end 27 | F[Attn] 28 | G[Output] 29 | A --> F 30 | C --> F 31 | D --> F 32 | E --> F 33 | C -.-> G 34 | D -.-> G 35 | E -.-> G 36 | F --> G 37 | 38 | All models compute the output as 39 | :math:`c = \sum_{j=1}^{SeqLength} a_j H_j` where 40 | :math:`a_j` is the softmax of a score function. 41 | Then then apply a projection layer to [q, c]. 42 | 43 | However they 44 | differ on how they compute the attention score. 45 | 46 | * Luong Attention (dot, general): 47 | * dot: :math:`score(H_j,q) = H_j^T q` 48 | * general: :math:`score(H_j, q) = H_j^T W_a q` 49 | 50 | 51 | * Bahdanau Attention (mlp): 52 | * :math:`score(H_j, q) = v_a^T tanh(W_a q + U_a h_j)` 53 | 54 | 55 | Args: 56 | dim (int): dimensionality of query and key 57 | coverage (bool): use coverage term 58 | attn_type (str): type of attention to use, options [dot,general,mlp] 59 | 60 | """ 61 | def __init__(self, dim, coverage=False, attn_type="dot"): 62 | super(GlobalAttention, self).__init__() 63 | 64 | self.dim = dim 65 | self.attn_type = attn_type 66 | assert (self.attn_type in ["dot", "general", "mlp"]), ( 67 | "Please select a valid attention type.") 68 | 69 | if self.attn_type == "general": 70 | self.linear_in = nn.Linear(dim, dim, bias=False) 71 | elif self.attn_type == "mlp": 72 | self.linear_context = nn.Linear(dim, dim, bias=False) 73 | self.linear_query = nn.Linear(dim, dim, bias=True) 74 | self.v = nn.Linear(dim, 1, bias=False) 75 | # mlp wants it with bias 76 | out_bias = self.attn_type == "mlp" 77 | self.linear_out = nn.Linear(dim*2, dim, bias=out_bias) 78 | 79 | self.sm = nn.Softmax(dim=-1) 80 | self.tanh = nn.Tanh() 81 | 82 | if coverage: 83 | self.linear_cover = nn.Linear(1, dim, bias=False) 84 | 85 | def score(self, h_t, h_s): 86 | """ 87 | Args: 88 | h_t (`FloatTensor`): sequence of queries `[batch x tgt_len x dim]` 89 | h_s (`FloatTensor`): sequence of sources `[batch x src_len x dim]` 90 | 91 | Returns: 92 | :obj:`FloatTensor`: 93 | raw attention scores (unnormalized) for each src index 94 | `[batch x tgt_len x src_len]` 95 | 96 | """ 97 | 98 | # Check input sizes 99 | src_batch, src_len, src_dim = h_s.size() 100 | tgt_batch, tgt_len, tgt_dim = h_t.size() 101 | aeq(src_batch, tgt_batch) 102 | aeq(src_dim, tgt_dim) 103 | aeq(self.dim, src_dim) 104 | 105 | if self.attn_type in ["general", "dot"]: 106 | if self.attn_type == "general": 107 | h_t_ = h_t.view(tgt_batch*tgt_len, tgt_dim) 108 | h_t_ = self.linear_in(h_t_) 109 | h_t = h_t_.view(tgt_batch, tgt_len, tgt_dim) 110 | h_s_ = h_s.transpose(1, 2) 111 | # (batch, t_len, d) x (batch, d, s_len) --> (batch, t_len, s_len) 112 | return torch.bmm(h_t, h_s_) 113 | else: 114 | dim = self.dim 115 | wq = self.linear_query(h_t.view(-1, dim)) 116 | wq = wq.view(tgt_batch, tgt_len, 1, dim) 117 | wq = wq.expand(tgt_batch, tgt_len, src_len, dim) 118 | 119 | uh = self.linear_context(h_s.contiguous().view(-1, dim)) 120 | uh = uh.view(src_batch, 1, src_len, dim) 121 | uh = uh.expand(src_batch, tgt_len, src_len, dim) 122 | 123 | # (batch, t_len, s_len, d) 124 | wquh = self.tanh(wq + uh) 125 | 126 | return self.v(wquh.view(-1, dim)).view(tgt_batch, tgt_len, src_len) 127 | 128 | def forward(self, input, memory_bank, memory_lengths=None, coverage=None): 129 | """ 130 | 131 | Args: 132 | input (`FloatTensor`): query vectors `[batch x tgt_len x dim]` 133 | memory_bank (`FloatTensor`): source vectors `[batch x src_len x dim]` 134 | memory_lengths (`LongTensor`): the source context lengths `[batch]` 135 | coverage (`FloatTensor`): None (not supported yet) 136 | 137 | Returns: 138 | (`FloatTensor`, `FloatTensor`): 139 | 140 | * Computed vector `[tgt_len x batch x dim]` 141 | * Attention distribtutions for each query 142 | `[tgt_len x batch x src_len]` 143 | """ 144 | 145 | # one step input 146 | if input.dim() == 2: 147 | one_step = True 148 | input = input.unsqueeze(1) 149 | else: 150 | one_step = False 151 | 152 | batch, sourceL, dim = memory_bank.size() 153 | batch_, targetL, dim_ = input.size() 154 | aeq(batch, batch_) 155 | aeq(dim, dim_) 156 | aeq(self.dim, dim) 157 | if coverage is not None: 158 | batch_, sourceL_ = coverage.size() 159 | aeq(batch, batch_) 160 | aeq(sourceL, sourceL_) 161 | 162 | if coverage is not None: 163 | cover = coverage.view(-1).unsqueeze(1) 164 | memory_bank += self.linear_cover(cover).view_as(memory_bank) 165 | memory_bank = self.tanh(memory_bank) 166 | 167 | # compute attention scores, as in Luong et al. 168 | align = self.score(input, memory_bank) 169 | 170 | if memory_lengths is not None: 171 | mask = sequence_mask(memory_lengths) 172 | mask = mask.unsqueeze(1) # Make it broadcastable. 173 | align.data.masked_fill_(1 - mask, -float('inf')) 174 | 175 | # Softmax to normalize attention weights 176 | align_vectors = self.sm(align.view(batch*targetL, sourceL)) 177 | align_vectors = align_vectors.view(batch, targetL, sourceL) 178 | 179 | # each context vector c_t is the weighted average 180 | # over all the source hidden states 181 | c = torch.bmm(align_vectors, memory_bank) 182 | 183 | # concatenate 184 | concat_c = torch.cat([c, input], 2).view(batch*targetL, dim*2) 185 | attn_h = self.linear_out(concat_c).view(batch, targetL, dim) 186 | if self.attn_type in ["general", "dot"]: 187 | attn_h = self.tanh(attn_h) 188 | 189 | if one_step: 190 | attn_h = attn_h.squeeze(1) 191 | align_vectors = align_vectors.squeeze(1) 192 | 193 | # Check output sizes 194 | batch_, dim_ = attn_h.size() 195 | aeq(batch, batch_) 196 | aeq(dim, dim_) 197 | batch_, sourceL_ = align_vectors.size() 198 | aeq(batch, batch_) 199 | aeq(sourceL, sourceL_) 200 | else: 201 | attn_h = attn_h.transpose(0, 1).contiguous() 202 | align_vectors = align_vectors.transpose(0, 1).contiguous() 203 | 204 | # Check output sizes 205 | targetL_, batch_, dim_ = attn_h.size() 206 | aeq(targetL, targetL_) 207 | aeq(batch, batch_) 208 | aeq(dim, dim_) 209 | targetL_, batch_, sourceL_ = align_vectors.size() 210 | aeq(targetL, targetL_) 211 | aeq(batch, batch_) 212 | aeq(sourceL, sourceL_) 213 | 214 | return attn_h, align_vectors 215 | -------------------------------------------------------------------------------- /onmt/modules/ImageEncoder.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | import torch 4 | from torch.autograd import Variable 5 | 6 | 7 | class ImageEncoder(nn.Module): 8 | """ 9 | A simple encoder convolutional -> recurrent neural network for 10 | image input. 11 | 12 | Args: 13 | num_layers (int): number of encoder layers. 14 | bidirectional (bool): bidirectional encoder. 15 | rnn_size (int): size of hidden states of the rnn. 16 | dropout (float): dropout probablity. 17 | """ 18 | def __init__(self, num_layers, bidirectional, rnn_size, dropout): 19 | super(ImageEncoder, self).__init__() 20 | self.num_layers = num_layers 21 | self.num_directions = 2 if bidirectional else 1 22 | self.hidden_size = rnn_size 23 | 24 | self.layer1 = nn.Conv2d(3, 64, kernel_size=(3, 3), 25 | padding=(1, 1), stride=(1, 1)) 26 | self.layer2 = nn.Conv2d(64, 128, kernel_size=(3, 3), 27 | padding=(1, 1), stride=(1, 1)) 28 | self.layer3 = nn.Conv2d(128, 256, kernel_size=(3, 3), 29 | padding=(1, 1), stride=(1, 1)) 30 | self.layer4 = nn.Conv2d(256, 256, kernel_size=(3, 3), 31 | padding=(1, 1), stride=(1, 1)) 32 | self.layer5 = nn.Conv2d(256, 512, kernel_size=(3, 3), 33 | padding=(1, 1), stride=(1, 1)) 34 | self.layer6 = nn.Conv2d(512, 512, kernel_size=(3, 3), 35 | padding=(1, 1), stride=(1, 1)) 36 | 37 | self.batch_norm1 = nn.BatchNorm2d(256) 38 | self.batch_norm2 = nn.BatchNorm2d(512) 39 | self.batch_norm3 = nn.BatchNorm2d(512) 40 | 41 | input_size = 512 42 | self.rnn = nn.LSTM(input_size, rnn_size, 43 | num_layers=num_layers, 44 | dropout=dropout, 45 | bidirectional=bidirectional) 46 | self.pos_lut = nn.Embedding(1000, input_size) 47 | 48 | def load_pretrained_vectors(self, opt): 49 | # Pass in needed options only when modify function definition. 50 | pass 51 | 52 | def forward(self, input, lengths=None): 53 | "See :obj:`onmt.modules.EncoderBase.forward()`" 54 | 55 | batch_size = input.size(0) 56 | # (batch_size, 64, imgH, imgW) 57 | # layer 1 58 | input = F.relu(self.layer1(input[:, :, :, :]-0.5), True) 59 | 60 | # (batch_size, 64, imgH/2, imgW/2) 61 | input = F.max_pool2d(input, kernel_size=(2, 2), stride=(2, 2)) 62 | 63 | # (batch_size, 128, imgH/2, imgW/2) 64 | # layer 2 65 | input = F.relu(self.layer2(input), True) 66 | 67 | # (batch_size, 128, imgH/2/2, imgW/2/2) 68 | input = F.max_pool2d(input, kernel_size=(2, 2), stride=(2, 2)) 69 | 70 | # (batch_size, 256, imgH/2/2, imgW/2/2) 71 | # layer 3 72 | # batch norm 1 73 | input = F.relu(self.batch_norm1(self.layer3(input)), True) 74 | 75 | # (batch_size, 256, imgH/2/2, imgW/2/2) 76 | # layer4 77 | input = F.relu(self.layer4(input), True) 78 | 79 | # (batch_size, 256, imgH/2/2/2, imgW/2/2) 80 | input = F.max_pool2d(input, kernel_size=(1, 2), stride=(1, 2)) 81 | 82 | # (batch_size, 512, imgH/2/2/2, imgW/2/2) 83 | # layer 5 84 | # batch norm 2 85 | input = F.relu(self.batch_norm2(self.layer5(input)), True) 86 | 87 | # (batch_size, 512, imgH/2/2/2, imgW/2/2/2) 88 | input = F.max_pool2d(input, kernel_size=(2, 1), stride=(2, 1)) 89 | 90 | # (batch_size, 512, imgH/2/2/2, imgW/2/2/2) 91 | input = F.relu(self.batch_norm3(self.layer6(input)), True) 92 | 93 | # # (batch_size, 512, H, W) 94 | all_outputs = [] 95 | for row in range(input.size(2)): 96 | inp = input[:, :, row, :].transpose(0, 2)\ 97 | .transpose(1, 2) 98 | row_vec = torch.Tensor(batch_size).type_as(inp.data)\ 99 | .long().fill_(row) 100 | pos_emb = self.pos_lut(Variable(row_vec)) 101 | with_pos = torch.cat( 102 | (pos_emb.view(1, pos_emb.size(0), pos_emb.size(1)), inp), 0) 103 | outputs, hidden_t = self.rnn(with_pos) 104 | all_outputs.append(outputs) 105 | out = torch.cat(all_outputs, 0) 106 | 107 | return hidden_t, out 108 | -------------------------------------------------------------------------------- /onmt/modules/MultiHeadedAttn.py: -------------------------------------------------------------------------------- 1 | import math 2 | import torch 3 | import torch.nn as nn 4 | from torch.autograd import Variable 5 | 6 | from onmt.Utils import aeq 7 | 8 | 9 | class MultiHeadedAttention(nn.Module): 10 | """ 11 | Multi-Head Attention module from 12 | "Attention is All You Need" 13 | :cite:`DBLP:journals/corr/VaswaniSPUJGKP17`. 14 | 15 | Similar to standard `dot` attention but uses 16 | multiple attention distributions simulataneously 17 | to select relevant items. 18 | 19 | .. mermaid:: 20 | 21 | graph BT 22 | A[key] 23 | B[value] 24 | C[query] 25 | O[output] 26 | subgraph Attn 27 | D[Attn 1] 28 | E[Attn 2] 29 | F[Attn N] 30 | end 31 | A --> D 32 | C --> D 33 | A --> E 34 | C --> E 35 | A --> F 36 | C --> F 37 | D --> O 38 | E --> O 39 | F --> O 40 | B --> O 41 | 42 | Also includes several additional tricks. 43 | 44 | Args: 45 | head_count (int): number of parallel heads 46 | model_dim (int): the dimension of keys/values/queries, 47 | must be divisible by head_count 48 | dropout (float): dropout parameter 49 | """ 50 | def __init__(self, head_count, model_dim, dropout=0.1): 51 | assert model_dim % head_count == 0 52 | self.dim_per_head = model_dim // head_count 53 | self.model_dim = model_dim 54 | 55 | super(MultiHeadedAttention, self).__init__() 56 | self.head_count = head_count 57 | 58 | self.linear_keys = nn.Linear(model_dim, 59 | head_count * self.dim_per_head) 60 | self.linear_values = nn.Linear(model_dim, 61 | head_count * self.dim_per_head) 62 | self.linear_query = nn.Linear(model_dim, 63 | head_count * self.dim_per_head) 64 | self.sm = nn.Softmax(dim=-1) 65 | self.dropout = nn.Dropout(dropout) 66 | self.final_linear = nn.Linear(model_dim, model_dim) 67 | 68 | def forward(self, key, value, query, mask=None): 69 | """ 70 | Compute the context vector and the attention vectors. 71 | 72 | Args: 73 | key (`FloatTensor`): set of `key_len` 74 | key vectors `[batch, key_len, dim]` 75 | value (`FloatTensor`): set of `key_len` 76 | value vectors `[batch, key_len, dim]` 77 | query (`FloatTensor`): set of `query_len` 78 | query vectors `[batch, query_len, dim]` 79 | mask: binary mask indicating which keys have 80 | non-zero attention `[batch, query_len, key_len]` 81 | Returns: 82 | (`FloatTensor`, `FloatTensor`) : 83 | 84 | * output context vectors `[batch, query_len, dim]` 85 | * one of the attention vectors `[batch, query_len, key_len]` 86 | """ 87 | 88 | # CHECKS 89 | batch, k_len, d = key.size() 90 | batch_, k_len_, d_ = value.size() 91 | aeq(batch, batch_) 92 | aeq(k_len, k_len_) 93 | aeq(d, d_) 94 | batch_, q_len, d_ = query.size() 95 | aeq(batch, batch_) 96 | aeq(d, d_) 97 | aeq(self.model_dim % 8, 0) 98 | if mask is not None: 99 | batch_, q_len_, k_len_ = mask.size() 100 | aeq(batch_, batch) 101 | aeq(k_len_, k_len) 102 | aeq(q_len_ == q_len) 103 | # END CHECKS 104 | 105 | batch_size = key.size(0) 106 | dim_per_head = self.dim_per_head 107 | head_count = self.head_count 108 | key_len = key.size(1) 109 | query_len = query.size(1) 110 | 111 | def shape(x): 112 | return x.view(batch_size, -1, head_count, dim_per_head) \ 113 | .transpose(1, 2) 114 | 115 | def unshape(x): 116 | return x.transpose(1, 2).contiguous() \ 117 | .view(batch_size, -1, head_count * dim_per_head) 118 | 119 | # 1) Project key, value, and query. 120 | key_up = shape(self.linear_keys(key)) 121 | value_up = shape(self.linear_values(value)) 122 | query_up = shape(self.linear_query(query)) 123 | 124 | # 2) Calculate and scale scores. 125 | query_up = query_up / math.sqrt(dim_per_head) 126 | scores = torch.matmul(query_up, key_up.transpose(2, 3)) 127 | 128 | if mask is not None: 129 | mask = mask.unsqueeze(1).expand_as(scores) 130 | scores = scores.masked_fill(Variable(mask), -1e18) 131 | 132 | # 3) Apply attention dropout and compute context vectors. 133 | attn = self.sm(scores) 134 | drop_attn = self.dropout(attn) 135 | context = unshape(torch.matmul(drop_attn, value_up)) 136 | 137 | output = self.final_linear(context) 138 | # CHECK 139 | batch_, q_len_, d_ = output.size() 140 | aeq(q_len, q_len_) 141 | aeq(batch, batch_) 142 | aeq(d, d_) 143 | 144 | # Return one attn 145 | top_attn = attn \ 146 | .view(batch_size, head_count, 147 | query_len, key_len)[:, 0, :, :] \ 148 | .contiguous() 149 | # END CHECK 150 | return output, top_attn 151 | -------------------------------------------------------------------------------- /onmt/modules/StackedRNN.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | class StackedLSTM(nn.Module): 6 | """ 7 | Our own implementation of stacked LSTM. 8 | Needed for the decoder, because we do input feeding. 9 | """ 10 | def __init__(self, num_layers, input_size, rnn_size, dropout): 11 | super(StackedLSTM, self).__init__() 12 | self.dropout = nn.Dropout(dropout) 13 | self.num_layers = num_layers 14 | self.layers = nn.ModuleList() 15 | 16 | for i in range(num_layers): 17 | self.layers.append(nn.LSTMCell(input_size, rnn_size)) 18 | input_size = rnn_size 19 | 20 | def forward(self, input, hidden): 21 | h_0, c_0 = hidden 22 | h_1, c_1 = [], [] 23 | for i, layer in enumerate(self.layers): 24 | h_1_i, c_1_i = layer(input, (h_0[i], c_0[i])) 25 | input = h_1_i 26 | if i + 1 != self.num_layers: 27 | input = self.dropout(input) 28 | h_1 += [h_1_i] 29 | c_1 += [c_1_i] 30 | 31 | h_1 = torch.stack(h_1) 32 | c_1 = torch.stack(c_1) 33 | 34 | return input, (h_1, c_1) 35 | 36 | 37 | class StackedGRU(nn.Module): 38 | 39 | def __init__(self, num_layers, input_size, rnn_size, dropout): 40 | super(StackedGRU, self).__init__() 41 | self.dropout = nn.Dropout(dropout) 42 | self.num_layers = num_layers 43 | self.layers = nn.ModuleList() 44 | 45 | for i in range(num_layers): 46 | self.layers.append(nn.GRUCell(input_size, rnn_size)) 47 | input_size = rnn_size 48 | 49 | def forward(self, input, hidden): 50 | h_1 = [] 51 | for i, layer in enumerate(self.layers): 52 | h_1_i = layer(input, hidden[0][i]) 53 | input = h_1_i 54 | if i + 1 != self.num_layers: 55 | input = self.dropout(input) 56 | h_1 += [h_1_i] 57 | 58 | h_1 = torch.stack(h_1) 59 | return input, (h_1,) 60 | -------------------------------------------------------------------------------- /onmt/modules/StructuredAttention.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch 3 | import torch.cuda 4 | from torch.autograd import Variable 5 | 6 | 7 | class MatrixTree(nn.Module): 8 | """Implementation of the matrix-tree theorem for computing marginals 9 | of non-projective dependency parsing. This attention layer is used 10 | in the paper "Learning Structured Text Representations." 11 | 12 | 13 | :cite:`DBLP:journals/corr/LiuL17d` 14 | """ 15 | def __init__(self, eps=1e-5): 16 | self.eps = eps 17 | super(MatrixTree, self).__init__() 18 | 19 | def forward(self, input): 20 | laplacian = input.exp() + self.eps 21 | output = input.clone() 22 | for b in range(input.size(0)): 23 | lap = laplacian[b].masked_fill( 24 | Variable(torch.eye(input.size(1)).cuda().ne(0)), 0) 25 | lap = -lap + torch.diag(lap.sum(0)) 26 | # store roots on diagonal 27 | lap[0] = input[b].diag().exp() 28 | inv_laplacian = lap.inverse() 29 | 30 | factor = inv_laplacian.diag().unsqueeze(1)\ 31 | .expand_as(input[b]).transpose(0, 1) 32 | term1 = input[b].exp().mul(factor).clone() 33 | term2 = input[b].exp().mul(inv_laplacian.transpose(0, 1)).clone() 34 | term1[:, 0] = 0 35 | term2[0] = 0 36 | output[b] = term1 - term2 37 | roots_output = input[b].diag().exp().mul( 38 | inv_laplacian.transpose(0, 1)[0]) 39 | output[b] = output[b] + torch.diag(roots_output) 40 | return output 41 | 42 | 43 | if __name__ == "__main__": 44 | dtree = MatrixTree() 45 | q = torch.rand(1, 5, 5).cuda() 46 | marg = dtree.forward(Variable(q)) 47 | print(marg.sum(1)) 48 | -------------------------------------------------------------------------------- /onmt/modules/UtilClass.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | class LayerNorm(nn.Module): 6 | def __init__(self, features, eps=1e-6): 7 | super(LayerNorm, self).__init__() 8 | self.a_2 = nn.Parameter(torch.ones(features)) 9 | self.b_2 = nn.Parameter(torch.zeros(features)) 10 | self.eps = eps 11 | 12 | def forward(self, x): 13 | mean = x.mean(-1, keepdim=True) 14 | std = x.std(-1, keepdim=True) 15 | return self.a_2 * (x - mean) / (std + self.eps) + self.b_2 16 | 17 | 18 | class Elementwise(nn.ModuleList): 19 | """ 20 | A simple network container. 21 | Parameters are a list of modules. 22 | Inputs are a 3d Variable whose last dimension is the same length 23 | as the list. 24 | Outputs are the result of applying modules to inputs elementwise. 25 | An optional merge parameter allows the outputs to be reduced to a 26 | single Variable. 27 | """ 28 | 29 | def __init__(self, merge=None, *args): 30 | assert merge in [None, 'first', 'concat', 'sum', 'mlp'] 31 | self.merge = merge 32 | super(Elementwise, self).__init__(*args) 33 | 34 | def forward(self, input): 35 | inputs = [feat.squeeze(2) for feat in input.split(1, dim=2)] 36 | assert len(self) == len(inputs) 37 | outputs = [f(x) for f, x in zip(self, inputs)] 38 | if self.merge == 'first': 39 | return outputs[0] 40 | elif self.merge == 'concat' or self.merge == 'mlp': 41 | return torch.cat(outputs, 2) 42 | elif self.merge == 'sum': 43 | return sum(outputs) 44 | else: 45 | return outputs 46 | -------------------------------------------------------------------------------- /onmt/modules/__init__.py: -------------------------------------------------------------------------------- 1 | from onmt.modules.UtilClass import LayerNorm, Elementwise 2 | from onmt.modules.Gate import context_gate_factory, ContextGate 3 | from onmt.modules.GlobalAttention import GlobalAttention 4 | from onmt.modules.ConvMultiStepAttention import ConvMultiStepAttention 5 | from onmt.modules.ImageEncoder import ImageEncoder 6 | from onmt.modules.AudioEncoder import AudioEncoder 7 | from onmt.modules.CopyGenerator import CopyGenerator, CopyGeneratorLossCompute 8 | from onmt.modules.StructuredAttention import MatrixTree 9 | from onmt.modules.Transformer import \ 10 | TransformerEncoder, TransformerDecoder, PositionwiseFeedForward 11 | from onmt.modules.Conv2Conv import CNNEncoder, CNNDecoder 12 | from onmt.modules.MultiHeadedAttn import MultiHeadedAttention 13 | from onmt.modules.StackedRNN import StackedLSTM, StackedGRU 14 | from onmt.modules.Embeddings import Embeddings, PositionalEncoding 15 | from onmt.modules.WeightNorm import WeightNormConv2d 16 | 17 | from onmt.Models import EncoderBase, MeanEncoder, StdRNNDecoder, \ 18 | RNNDecoderBase, InputFeedRNNDecoder, RNNEncoder, NMTModel#, HighwayMLPDecoder 19 | 20 | from onmt.modules.SRU import check_sru_requirement 21 | can_use_sru = check_sru_requirement() 22 | if can_use_sru: 23 | from onmt.modules.SRU import SRU 24 | 25 | 26 | # For flake8 compatibility. 27 | __all__ = [EncoderBase, MeanEncoder, RNNDecoderBase, InputFeedRNNDecoder, 28 | RNNEncoder, NMTModel, #HighwayMLPDecoder, 29 | StdRNNDecoder, ContextGate, GlobalAttention, ImageEncoder, 30 | PositionwiseFeedForward, PositionalEncoding, 31 | CopyGenerator, MultiHeadedAttention, 32 | LayerNorm, 33 | TransformerEncoder, TransformerDecoder, Embeddings, Elementwise, 34 | MatrixTree, WeightNormConv2d, ConvMultiStepAttention, 35 | CNNEncoder, CNNDecoder, StackedLSTM, StackedGRU, 36 | context_gate_factory, CopyGeneratorLossCompute, AudioEncoder] 37 | 38 | if can_use_sru: 39 | __all__.extend([SRU, check_sru_requirement]) 40 | -------------------------------------------------------------------------------- /onmt/modules/__pycache__/AudioEncoder.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/AudioEncoder.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/Conv2Conv.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/Conv2Conv.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/ConvMultiStepAttention.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/ConvMultiStepAttention.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/CopyGenerator.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/CopyGenerator.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/Embeddings.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/Embeddings.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/Gate.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/Gate.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/GlobalAttention.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/GlobalAttention.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/ImageEncoder.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/ImageEncoder.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/MultiHeadedAttn.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/MultiHeadedAttn.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/SRU.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/SRU.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/StackedRNN.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/StackedRNN.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/StructuredAttention.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/StructuredAttention.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/Transformer.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/Transformer.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/UtilClass.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/UtilClass.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/WeightNorm.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/WeightNorm.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/modules/__pycache__/__init__.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/modules/__pycache__/__init__.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/translate/Penalties.py: -------------------------------------------------------------------------------- 1 | from __future__ import division 2 | import torch 3 | 4 | 5 | class PenaltyBuilder(object): 6 | """ 7 | Returns the Length and Coverage Penalty function for Beam Search. 8 | 9 | Args: 10 | length_pen (str): option name of length pen 11 | cov_pen (str): option name of cov pen 12 | """ 13 | def __init__(self, cov_pen, length_pen): 14 | self.length_pen = length_pen 15 | self.cov_pen = cov_pen 16 | 17 | def coverage_penalty(self): 18 | if self.cov_pen == "wu": 19 | return self.coverage_wu 20 | elif self.cov_pen == "summary": 21 | return self.coverage_summary 22 | else: 23 | return self.coverage_none 24 | 25 | def length_penalty(self): 26 | if self.length_pen == "wu": 27 | return self.length_wu 28 | elif self.length_pen == "avg": 29 | return self.length_average 30 | else: 31 | return self.length_none 32 | 33 | """ 34 | Below are all the different penalty terms implemented so far 35 | """ 36 | 37 | def coverage_wu(self, beam, cov, beta=0.): 38 | """ 39 | NMT coverage re-ranking score from 40 | "Google's Neural Machine Translation System" :cite:`wu2016google`. 41 | """ 42 | penalty = -torch.min(cov, cov.clone().fill_(1.0)).log().sum(1) 43 | return beta * penalty 44 | 45 | def coverage_summary(self, beam, cov, beta=0.): 46 | """ 47 | Our summary penalty. 48 | """ 49 | penalty = torch.max(cov, cov.clone().fill_(1.0)).sum(1) 50 | penalty -= cov.size(1) 51 | return beta * penalty 52 | 53 | def coverage_none(self, beam, cov, beta=0.): 54 | """ 55 | returns zero as penalty 56 | """ 57 | return beam.scores.clone().fill_(0.0) 58 | 59 | def length_wu(self, beam, logprobs, alpha=0.): 60 | """ 61 | NMT length re-ranking score from 62 | "Google's Neural Machine Translation System" :cite:`wu2016google`. 63 | """ 64 | 65 | modifier = (((5 + len(beam.next_ys)) ** alpha) / 66 | ((5 + 1) ** alpha)) 67 | return (logprobs / modifier) 68 | 69 | def length_average(self, beam, logprobs, alpha=0.): 70 | """ 71 | Returns the average probability of tokens in a sequence. 72 | """ 73 | return logprobs / len(beam.next_ys) 74 | 75 | def length_none(self, beam, logprobs, alpha=0., beta=0.): 76 | """ 77 | Returns unmodified scores. 78 | """ 79 | return logprobs 80 | -------------------------------------------------------------------------------- /onmt/translate/Translation.py: -------------------------------------------------------------------------------- 1 | from __future__ import division, unicode_literals 2 | 3 | import torch 4 | import onmt.io 5 | 6 | 7 | class TranslationBuilder(object): 8 | """ 9 | Build a word-based translation from the batch output 10 | of translator and the underlying dictionaries. 11 | 12 | Replacement based on "Addressing the Rare Word 13 | Problem in Neural Machine Translation" :cite:`Luong2015b` 14 | 15 | Args: 16 | data (DataSet): 17 | fields (dict of Fields): data fields 18 | n_best (int): number of translations produced 19 | replace_unk (bool): replace unknown words using attention 20 | has_tgt (bool): will the batch have gold targets 21 | """ 22 | def __init__(self, data, fields, n_best=1, replace_unk=False, 23 | has_tgt=False): 24 | self.data = data 25 | self.fields = fields 26 | self.n_best = n_best 27 | self.replace_unk = replace_unk 28 | self.has_tgt = has_tgt 29 | 30 | def _build_target_tokens(self, src, src_vocab, src_raw, pred, attn): 31 | vocab = self.fields["tgt"].vocab 32 | tokens = [] 33 | for tok in pred: 34 | if tok < len(vocab): 35 | tokens.append(vocab.itos[tok]) 36 | else: 37 | tokens.append(src_vocab.itos[tok - len(vocab)]) 38 | if tokens[-1] == onmt.io.EOS_WORD: 39 | tokens = tokens[:-1] 40 | break 41 | if self.replace_unk and (attn is not None) and (src is not None): 42 | for i in range(len(tokens)): 43 | if tokens[i] == vocab.itos[onmt.io.UNK]: 44 | _, maxIndex = attn[i].max(0) 45 | tokens[i] = src_raw[maxIndex[0]] 46 | return tokens 47 | 48 | def from_batch(self, translation_batch): 49 | batch = translation_batch["batch"] 50 | assert(len(translation_batch["gold_score"]) == 51 | len(translation_batch["predictions"])) 52 | batch_size = batch.batch_size 53 | 54 | preds, pred_score, attn, gold_score, indices = list(zip( 55 | *sorted(zip(translation_batch["predictions"], 56 | translation_batch["scores"], 57 | translation_batch["attention"], 58 | translation_batch["gold_score"], 59 | batch.indices.data), 60 | key=lambda x: x[-1]))) 61 | 62 | # Sorting 63 | inds, perm = torch.sort(batch.indices.data) 64 | data_type = self.data.data_type 65 | if data_type == 'text': 66 | src = batch.src[0].data.index_select(1, perm) 67 | else: 68 | src = None 69 | 70 | if self.has_tgt: 71 | tgt = batch.tgt.data.index_select(1, perm) 72 | else: 73 | tgt = None 74 | 75 | translations = [] 76 | for b in range(batch_size): 77 | if data_type == 'text': 78 | src_vocab = self.data.src_vocabs[inds[b]] \ 79 | if self.data.src_vocabs else None 80 | src_raw = self.data.examples[inds[b]].src 81 | else: 82 | src_vocab = None 83 | src_raw = None 84 | pred_sents = [self._build_target_tokens( 85 | src[:, b] if src is not None else None, 86 | src_vocab, src_raw, 87 | preds[b][n], attn[b][n]) 88 | for n in range(self.n_best)] 89 | gold_sent = None 90 | if tgt is not None: 91 | gold_sent = self._build_target_tokens( 92 | src[:, b] if src is not None else None, 93 | src_vocab, src_raw, 94 | tgt[1:, b] if tgt is not None else None, None) 95 | 96 | translation = Translation(src[:, b] if src is not None else None, 97 | src_raw, pred_sents, 98 | attn[b], pred_score[b], gold_sent, 99 | gold_score[b]) 100 | translations.append(translation) 101 | 102 | return translations 103 | 104 | 105 | class Translation(object): 106 | """ 107 | Container for a translated sentence. 108 | 109 | Attributes: 110 | src (`LongTensor`): src word ids 111 | src_raw ([str]): raw src words 112 | 113 | pred_sents ([[str]]): words from the n-best translations 114 | pred_scores ([[float]]): log-probs of n-best translations 115 | attns ([`FloatTensor`]) : attention dist for each translation 116 | gold_sent ([str]): words from gold translation 117 | gold_score ([float]): log-prob of gold translation 118 | 119 | """ 120 | def __init__(self, src, src_raw, pred_sents, 121 | attn, pred_scores, tgt_sent, gold_score): 122 | self.src = src 123 | self.src_raw = src_raw 124 | self.pred_sents = pred_sents 125 | self.attns = attn 126 | self.pred_scores = pred_scores 127 | self.gold_sent = tgt_sent 128 | self.gold_score = gold_score 129 | 130 | def log(self, sent_number): 131 | """ 132 | Log translation to stdout. 133 | """ 134 | output = '\nSENT {}: {}\n'.format(sent_number, self.src_raw) 135 | 136 | best_pred = self.pred_sents[0] 137 | best_score = self.pred_scores[0] 138 | pred_sent = ' '.join(best_pred) 139 | output += 'PRED {}: {}\n'.format(sent_number, pred_sent) 140 | print("PRED SCORE: {:.4f}".format(best_score)) 141 | 142 | if self.gold_sent is not None: 143 | tgt_sent = ' '.join(self.gold_sent) 144 | output += 'GOLD {}: {}\n'.format(sent_number, tgt_sent) 145 | # output += ("GOLD SCORE: {:.4f}".format(self.gold_score)) 146 | print("GOLD SCORE: {:.4f}".format(self.gold_score)) 147 | if len(self.pred_sents) > 1: 148 | print('\nBEST HYP:') 149 | for score, sent in zip(self.pred_scores, self.pred_sents): 150 | output += "[{:.4f}] {}\n".format(score, sent) 151 | 152 | return output 153 | -------------------------------------------------------------------------------- /onmt/translate/__init__.py: -------------------------------------------------------------------------------- 1 | from onmt.translate.Translator import Translator 2 | from onmt.translate.Translation import Translation, TranslationBuilder 3 | from onmt.translate.Beam import Beam, GNMTGlobalScorer 4 | from onmt.translate.Penalties import PenaltyBuilder 5 | from onmt.translate.TranslationServer import TranslationServer, \ 6 | ServerModelError 7 | 8 | __all__ = [Translator, Translation, Beam, 9 | GNMTGlobalScorer, TranslationBuilder, 10 | PenaltyBuilder, TranslationServer, ServerModelError] 11 | -------------------------------------------------------------------------------- /onmt/translate/__pycache__/Beam.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/translate/__pycache__/Beam.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/translate/__pycache__/Penalties.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/translate/__pycache__/Penalties.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/translate/__pycache__/Translation.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/translate/__pycache__/Translation.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/translate/__pycache__/TranslationServer.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/translate/__pycache__/TranslationServer.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/translate/__pycache__/Translator.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/translate/__pycache__/Translator.cpython-36.pyc -------------------------------------------------------------------------------- /onmt/translate/__pycache__/__init__.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/onmt/translate/__pycache__/__init__.cpython-36.pyc -------------------------------------------------------------------------------- /preprocess.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | import argparse 5 | import os 6 | import glob 7 | import sys 8 | 9 | import torch 10 | 11 | import onmt.io 12 | import onmt.opts 13 | 14 | 15 | def check_existing_pt_files(opt): 16 | # We will use glob.glob() to find sharded {train|valid}.[0-9]*.pt 17 | # when training, so check to avoid tampering with existing pt files 18 | # or mixing them up. 19 | for t in ['train', 'valid', 'vocab']: 20 | pattern = opt.save_data + '.' + t + '*.pt' 21 | if glob.glob(pattern): 22 | sys.stderr.write("Please backup exisiting pt file: %s, " 23 | "to avoid tampering!\n" % pattern) 24 | sys.exit(1) 25 | 26 | 27 | def parse_args(): 28 | parser = argparse.ArgumentParser( 29 | description='preprocess.py', 30 | formatter_class=argparse.ArgumentDefaultsHelpFormatter) 31 | 32 | onmt.opts.add_md_help_argument(parser) 33 | onmt.opts.preprocess_opts(parser) 34 | 35 | opt = parser.parse_args() 36 | torch.manual_seed(opt.seed) 37 | 38 | check_existing_pt_files(opt) 39 | 40 | return opt 41 | 42 | 43 | def build_save_text_dataset_in_shards(src_corpus, tgt_corpus, fields, 44 | corpus_type, opt): 45 | ''' 46 | Divide the big corpus into shards, and build dataset separately. 47 | This is currently only for data_type=='text'. 48 | 49 | The reason we do this is to avoid taking up too much memory due 50 | to sucking in a huge corpus file. 51 | 52 | To tackle this, we only read in part of the corpus file of size 53 | `max_shard_size`(actually it is multiples of 64 bytes that equals 54 | or is slightly larger than this size), and process it into dataset, 55 | then write it to disk along the way. By doing this, we only focus on 56 | part of the corpus at any moment, thus effectively reducing memory use. 57 | According to test, this method can reduce memory footprint by ~50%. 58 | 59 | Note! As we process along the shards, previous shards might still 60 | stay in memory, but since we are done with them, and no more 61 | reference to them, if there is memory tight situation, the OS could 62 | easily reclaim these memory. 63 | 64 | If `max_shard_size` is 0 or is larger than the corpus size, it is 65 | effectively preprocessed into one dataset, i.e. no sharding. 66 | 67 | NOTE! `max_shard_size` is measuring the input corpus size, not the 68 | output pt file size. So a shard pt file consists of examples of size 69 | 2 * `max_shard_size`(source + target). 70 | ''' 71 | 72 | corpus_size = os.path.getsize(src_corpus) 73 | if corpus_size > 10 * (1024**2) and opt.max_shard_size == 0: 74 | print("Warning. The corpus %s is larger than 10M bytes, you can " 75 | "set '-max_shard_size' to process it by small shards " 76 | "to use less memory." % src_corpus) 77 | 78 | if opt.max_shard_size != 0: 79 | print(' * divide corpus into shards and build dataset separately' 80 | '(shard_size = %d bytes).' % opt.max_shard_size) 81 | 82 | ret_list = [] 83 | src_iter = onmt.io.ShardedTextCorpusIterator( 84 | src_corpus, opt.src_seq_length_trunc, 85 | "src", opt.max_shard_size) 86 | tgt_iter = onmt.io.ShardedTextCorpusIterator( 87 | tgt_corpus, opt.tgt_seq_length_trunc, 88 | "tgt", opt.max_shard_size, 89 | assoc_iter=src_iter) 90 | 91 | index = 0 92 | while not src_iter.hit_end(): 93 | index += 1 94 | dataset = onmt.io.TextDataset( 95 | fields, src_iter, tgt_iter, 96 | src_iter.num_feats, tgt_iter.num_feats, 97 | src_seq_length=opt.src_seq_length, 98 | tgt_seq_length=opt.tgt_seq_length, 99 | dynamic_dict=opt.dynamic_dict) 100 | 101 | # We save fields in vocab.pt seperately, so make it empty. 102 | dataset.fields = [] 103 | 104 | pt_file = "{:s}.{:s}.{:d}.pt".format( 105 | opt.save_data, corpus_type, index) 106 | print(" * saving %s data shard to %s." % (corpus_type, pt_file)) 107 | torch.save(dataset, pt_file) 108 | 109 | ret_list.append(pt_file) 110 | 111 | return ret_list 112 | 113 | 114 | def build_save_dataset(corpus_type, fields, opt): 115 | assert corpus_type in ['train', 'valid'] 116 | 117 | if corpus_type == 'train': 118 | src_corpus = opt.train_src 119 | tgt_corpus = opt.train_tgt 120 | else: 121 | src_corpus = opt.valid_src 122 | tgt_corpus = opt.valid_tgt 123 | 124 | # Currently we only do preprocess sharding for corpus: data_type=='text'. 125 | if opt.data_type == 'text': 126 | return build_save_text_dataset_in_shards( 127 | src_corpus, tgt_corpus, fields, 128 | corpus_type, opt) 129 | 130 | # For data_type == 'img' or 'audio', currently we don't do 131 | # preprocess sharding. We only build a monolithic dataset. 132 | # But since the interfaces are uniform, it would be not hard 133 | # to do this should users need this feature. 134 | dataset = onmt.io.build_dataset( 135 | fields, opt.data_type, src_corpus, tgt_corpus, 136 | src_dir=opt.src_dir, 137 | src_seq_length=opt.src_seq_length, 138 | tgt_seq_length=opt.tgt_seq_length, 139 | src_seq_length_trunc=opt.src_seq_length_trunc, 140 | tgt_seq_length_trunc=opt.tgt_seq_length_trunc, 141 | dynamic_dict=opt.dynamic_dict, 142 | sample_rate=opt.sample_rate, 143 | window_size=opt.window_size, 144 | window_stride=opt.window_stride, 145 | window=opt.window) 146 | 147 | # We save fields in vocab.pt seperately, so make it empty. 148 | dataset.fields = [] 149 | 150 | pt_file = "{:s}.{:s}.pt".format(opt.save_data, corpus_type) 151 | print(" * saving %s dataset to %s." % (corpus_type, pt_file)) 152 | torch.save(dataset, pt_file) 153 | 154 | return [pt_file] 155 | 156 | 157 | def build_save_vocab(train_dataset, fields, opt): 158 | fields = onmt.io.build_vocab(train_dataset, fields, opt.data_type, 159 | opt.share_vocab, 160 | opt.src_vocab, 161 | opt.src_vocab_size, 162 | opt.src_words_min_frequency, 163 | opt.tgt_vocab, 164 | opt.tgt_vocab_size, 165 | opt.tgt_words_min_frequency) 166 | 167 | # Can't save fields, so remove/reconstruct at training time. 168 | vocab_file = opt.save_data + '.vocab.pt' 169 | torch.save(onmt.io.save_fields_to_vocab(fields), vocab_file) 170 | 171 | 172 | def main(): 173 | opt = parse_args() 174 | 175 | print("Extracting features...") 176 | src_nfeats = onmt.io.get_num_features(opt.data_type, opt.train_src, 'src') 177 | tgt_nfeats = onmt.io.get_num_features(opt.data_type, opt.train_tgt, 'tgt') 178 | print(" * number of source features: %d." % src_nfeats) 179 | print(" * number of target features: %d." % tgt_nfeats) 180 | 181 | print("Building `Fields` object...") 182 | fields = onmt.io.get_fields(opt.data_type, src_nfeats, tgt_nfeats) 183 | 184 | print("Building & saving training data...") 185 | train_dataset_files = build_save_dataset('train', fields, opt) 186 | 187 | print("Building & saving vocabulary...") 188 | build_save_vocab(train_dataset_files, fields, opt) 189 | 190 | print("Building & saving validation data...") 191 | build_save_dataset('valid', fields, opt) 192 | 193 | 194 | if __name__ == "__main__": 195 | main() 196 | -------------------------------------------------------------------------------- /requirements.opt.txt: -------------------------------------------------------------------------------- 1 | cffi 2 | torchvision==0.1.8 3 | librosa 4 | Pillow 5 | git+https://github.com/pytorch/audio 6 | pyrouge 7 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | six 2 | tqdm 3 | pytorch=0.3.1 4 | torchtext>=0.2.1 5 | future 6 | allennlp 7 | sentencepiece 8 | gensim -------------------------------------------------------------------------------- /resources/seq2seq4dp.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/resources/seq2seq4dp.pdf -------------------------------------------------------------------------------- /screenshots/seq2seq_model.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bcmi220/seq2seq_parser/4143c2f9b3164c0fe8b8374f6bcca747184193d9/screenshots/seq2seq_model.png -------------------------------------------------------------------------------- /server.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | import argparse 3 | 4 | from flask import Flask, jsonify, request 5 | from onmt.translate import TranslationServer, ServerModelError 6 | 7 | STATUS_OK = "ok" 8 | STATUS_ERROR = "error" 9 | 10 | 11 | def start(config_file, 12 | url_root="./translator", 13 | host="0.0.0.0", 14 | port=5000, 15 | debug=True): 16 | def prefix_route(route_function, prefix='', mask='{0}{1}'): 17 | def newroute(route, *args, **kwargs): 18 | return route_function(mask.format(prefix, route), *args, **kwargs) 19 | return newroute 20 | 21 | app = Flask(__name__) 22 | app.route = prefix_route(app.route, url_root) 23 | translation_server = TranslationServer() 24 | translation_server.start(config_file) 25 | 26 | @app.route('/models', methods=['GET']) 27 | def get_models(): 28 | out = translation_server.list_models() 29 | return jsonify(out) 30 | 31 | @app.route('/clone_model/', methods=['POST']) 32 | def clone_model(model_id): 33 | out = {} 34 | data = request.get_json(force=True) 35 | timeout = -1 36 | if 'timeout' in data: 37 | timeout = data['timeout'] 38 | del data['timeout'] 39 | 40 | opt = data.get('opt', None) 41 | try: 42 | model_id, load_time = translation_server.clone_model( 43 | model_id, opt, timeout) 44 | except ServerModelError as e: 45 | out['status'] = STATUS_ERROR 46 | out['error'] = str(e) 47 | else: 48 | out['status'] = STATUS_OK 49 | out['model_id'] = model_id 50 | out['load_time'] = load_time 51 | 52 | return jsonify(out) 53 | 54 | @app.route('/unload_model/', methods=['GET']) 55 | def unload_model(model_id): 56 | out = {"model_id": model_id} 57 | 58 | try: 59 | translation_server.unload_model(model_id) 60 | out['status'] = STATUS_OK 61 | except Exception as e: 62 | out['status'] = STATUS_ERROR 63 | out['error'] = str(e) 64 | 65 | return jsonify(out) 66 | 67 | @app.route('/translate', methods=['POST']) 68 | def translate(): 69 | inputs = request.get_json(force=True) 70 | out = {} 71 | try: 72 | translation, scores, n_best, times = translation_server.run(inputs) 73 | assert len(translation) == len(inputs) 74 | assert len(scores) == len(inputs) 75 | 76 | out = [[{"src": inputs[i]['src'], "tgt": translation[i], 77 | "n_best": n_best, 78 | "pred_score": scores[i]} 79 | for i in range(len(translation))]] 80 | except ServerModelError as e: 81 | out['error'] = str(e) 82 | out['status'] = STATUS_ERROR 83 | 84 | return jsonify(out) 85 | 86 | @app.route('/to_cpu/', methods=['GET']) 87 | def to_cpu(model_id): 88 | out = {'model_id': model_id} 89 | translation_server.models[model_id].to_cpu() 90 | 91 | out['status'] = STATUS_OK 92 | return jsonify(out) 93 | 94 | @app.route('/to_gpu/', methods=['GET']) 95 | def to_gpu(model_id): 96 | out = {'model_id': model_id} 97 | translation_server.models[model_id].to_gpu() 98 | 99 | out['status'] = STATUS_OK 100 | return jsonify(out) 101 | 102 | app.run(debug=debug, host=host, port=port, use_reloader=False) 103 | 104 | 105 | if __name__ == '__main__': 106 | parser = argparse.ArgumentParser(description="OpenNMT-py REST Server") 107 | parser.add_argument("--ip", type=str, default="0.0.0.0") 108 | parser.add_argument("--port", type=int, default="5000") 109 | parser.add_argument("--url_root", type=str, default="/translator") 110 | parser.add_argument("--debug", "-d", action="store_true") 111 | parser.add_argument("--config", "-c", type=str, 112 | default="./available_models/conf.json") 113 | args = parser.parse_args() 114 | start(args.config, url_root=args.url_root, host=args.ip, port=args.port, 115 | debug=args.debug) 116 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | 3 | from setuptools import setup 4 | 5 | setup(name='OpenNMT-py', 6 | description='A python implementation of OpenNMT', 7 | version='0.1', 8 | packages=['onmt', 'onmt.io', 'onmt.translate', 'onmt.modules']) 9 | -------------------------------------------------------------------------------- /subroot/README.md: -------------------------------------------------------------------------------- 1 | # DNN Pytorch 2 | 3 | This repository includes a name tagger implemented with bidirectional LSTMs CRF network. It has an interface for external features. 4 | 5 | 6 | ## Model 7 | 8 | ![Alt Text](https://blender04.cs.rpi.edu/~zhangb8/public_misc/ijcnlp17_model.png) 9 | 10 | ## Requirements 11 | 12 | Python3, Pytorch 13 | 14 | ## Data Format 15 | 16 | * Label format 17 | 18 | The name tagger follows *BIO* or *BIOES* scheme: 19 | 20 | ![Alt Text](https://blender04.cs.rpi.edu/~zhangb8/public_misc/bio_scheme_example.png) 21 | 22 | * Sentence format 23 | 24 | Document is segmented into sentences. Each sentence is tokenized into multiple tokens. 25 | 26 | In the training file, sentences are separated by an empty line. Tokens are separated by linebreak. For each token, label should be always at the end. Token and label are separated by space. 27 | 28 | CRF style features can be added between token and labels. 29 | 30 | Example: 31 | ``` 32 | George B-PER 33 | W. I-PER 34 | Bush I-PER 35 | went O 36 | to O 37 | Germany B-GPE 38 | yesterday O 39 | . O 40 | 41 | New B-ORG 42 | York I-ORG 43 | Times I-ORG 44 | ``` 45 | 46 | A real example of a bio file: `example/seq_labeling/data/eng.train.bio` 47 | 48 | A real example of a bio file with features: `example/seq_labeling/data/eng.train.feat.bio` 49 | 50 | 51 | ## Usage 52 | 53 | Training and testing examples are provided in `example/seq_labeling/`. 54 | 55 | ## Citation 56 | 57 | [1] Boliang Zhang, Di Lu, Xiaoman Pan, Ying Lin, Halidanmu Abudukelimu, Heng Ji, Kevin Knight. [Embracing Non-Traditional Linguistic Resources for Low-resource Language Name Tagging](http://aclweb.org/anthology/I17-1037), Proc. IJCNLP, 2017 58 | 59 | [2] Boliang Zhang, Xiaoman Pan, Tianlu Wang, Ashish Vaswani, Heng Ji, Kevin Knight, and Daniel Marcu. [Name Tagging for Low-Resource Incident Languages Based on Expectation-Driven Learning](http://nlp.cs.rpi.edu/paper/expectation2016.pdf), Proc. NAACL, 2016 60 | 61 | 62 | -------------------------------------------------------------------------------- /subroot/RUN.md: -------------------------------------------------------------------------------- 1 | python3 ./srl_pytorch/train.py --train ./srl/data/predicate-train.txt --dev ./srl/data/predicate-dev.txt --test ./srl/data/predicate-test.txt --model_dp ./model/ --tag_scheme classification --lower 1 --zeros 1 --char_dim 0 --char_lstm_dim 0 --char_conv_channel 0 --word_dim 100 --word_lstm_dim 512 --word_lstm_layer 4 --pre_emb '' --all_emb 0 --cap_dim 0 --feat_dim 100 --feat_column 2 --crf 1 --dropout 0.1 --lr_method sgd-init_lr=.001-lr_decay_epoch=100 --batch_size 64 --gpu 1 2 | -------------------------------------------------------------------------------- /subroot/dnn_pytorch/dnn_utils.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | from argparse import ArgumentParser 4 | 5 | parser = ArgumentParser(add_help=False) 6 | parser.add_argument( 7 | "--gpu", default="0", 8 | type=int, help="set 1 to use gpu." 9 | ) 10 | args = parser.parse_known_args() 11 | 12 | # set global torch tensor variables. default is using cpu 13 | if args[0].gpu == 1: 14 | FloatTensor = torch.cuda.FloatTensor 15 | LongTensor = torch.cuda.LongTensor 16 | else: 17 | FloatTensor = torch.FloatTensor 18 | LongTensor = torch.LongTensor 19 | 20 | 21 | def exp_lr_scheduler(optimizer, epoch, init_lr=0.01, lr_decay_epoch=7): 22 | """Decay learning rate by a factor of 0.1 every lr_decay_epoch epochs.""" 23 | lr = init_lr * (0.1**(epoch // lr_decay_epoch)) 24 | 25 | if epoch % lr_decay_epoch == 0: 26 | print('LR is set to {}'.format(lr)) 27 | 28 | for param_group in optimizer.param_groups: 29 | param_group['lr'] = lr 30 | 31 | return optimizer 32 | 33 | 34 | def init_variable(shape): 35 | if len(shape) == 1: 36 | value = np.zeros(shape) # bias are initialized with zeros 37 | else: 38 | drange = np.sqrt(6. / (np.sum(shape))) 39 | value = drange * np.random.uniform(low=-1.0, high=1.0, size=shape) 40 | return value 41 | 42 | 43 | def init_param(layer): 44 | """ 45 | randomly initialize parameters of the given layer 46 | """ 47 | for p in layer.parameters(): 48 | p.data = torch.from_numpy(init_variable(p.size())).type(FloatTensor) 49 | 50 | return layer 51 | 52 | 53 | def log_sum_exp(x, dim=None): 54 | """ 55 | Sum probabilities in the log-space. 56 | """ 57 | xmax, _ = x.max(dim=dim, keepdim=True) 58 | xmax_, _ = x.max(dim=dim) 59 | # return xmax_ 60 | return xmax_ + torch.log(torch.exp(x - xmax).sum(dim=dim)) 61 | 62 | 63 | def sequence_mask(batch_len, max_len=None): 64 | if not max_len: 65 | max_len = np.max(batch_len) 66 | 67 | mask = np.zeros((len(batch_len), max_len)) 68 | for i in range(len(batch_len)): 69 | mask[i, range(batch_len[i])] = 1 70 | 71 | return mask 72 | -------------------------------------------------------------------------------- /subroot/dnn_pytorch/tag.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import time 3 | import torch 4 | 5 | from nn import SeqLabeling 6 | from utils import create_input, iobes_iob 7 | from loader import prepare_dataset, load_sentences 8 | 9 | 10 | # Read parameters from command line 11 | parser = argparse.ArgumentParser() 12 | parser.add_argument( 13 | "--model", default="", 14 | help="Model location" 15 | ) 16 | parser.add_argument( 17 | "--input", default="", 18 | help="Input bio file location" 19 | ) 20 | parser.add_argument( 21 | "--output", default="", 22 | help="Output bio file location" 23 | ) 24 | parser.add_argument( 25 | "--batch_size", default="50", 26 | type=int, help="batch size" 27 | ) 28 | parser.add_argument( 29 | "--gpu", default="0", 30 | type=int, help="default is 0. set 1 to use gpu." 31 | ) 32 | args = parser.parse_args() 33 | 34 | print('loading model from:', args.model) 35 | if args.gpu: 36 | state = torch.load(args.model) 37 | else: 38 | state = torch.load(args.model, map_location=lambda storage, loc: storage) 39 | 40 | parameters = state['parameters'] 41 | mappings = state['mappings'] 42 | 43 | # Load reverse mappings 44 | word_to_id, char_to_id, tag_to_id = [ 45 | {v: k for k, v in x.items()} 46 | for x in [mappings['id_to_word'], mappings['id_to_char'], mappings['id_to_tag']] 47 | ] 48 | feat_to_id_list = [ 49 | {v: k for k, v in id_to_feat.items()} 50 | for id_to_feat in mappings['id_to_feat_list'] 51 | ] 52 | 53 | # eval sentences 54 | eval_sentences = load_sentences( 55 | args.input, 56 | parameters['lower'], 57 | parameters['zeros'] 58 | ) 59 | 60 | eval_dataset = prepare_dataset( 61 | eval_sentences, parameters['feat_column'], 62 | word_to_id, char_to_id, tag_to_id, feat_to_id_list, parameters['lower'], 63 | is_train=False 64 | ) 65 | 66 | print("%i sentences in eval set." % len(eval_dataset)) 67 | 68 | # initialize model 69 | model = SeqLabeling(parameters) 70 | model.load_state_dict(state['state_dict']) 71 | model.train(False) 72 | 73 | since = time.time() 74 | batch_size = args.batch_size 75 | f_output = open(args.output, 'w') 76 | 77 | # Iterate over data. 78 | print('tagging...') 79 | for i in range(0, len(eval_dataset), batch_size): 80 | inputs, seq_index_mapping, char_index_mapping, seq_len, char_len = \ 81 | create_input(eval_dataset[i:i+batch_size], parameters, add_label=False) 82 | 83 | # forward 84 | outputs, loss = model.forward(inputs, seq_len, char_len, char_index_mapping) 85 | if parameters['crf']: 86 | preds = [outputs[seq_index_mapping[j]].data 87 | for j in range(len(outputs))] 88 | else: 89 | _, _preds = torch.max(outputs.data, 2) 90 | 91 | preds = [ 92 | _preds[seq_index_mapping[j]][:seq_len[seq_index_mapping[j]]] 93 | for j in range(len(seq_index_mapping)) 94 | ] 95 | for j, pred in enumerate(preds): 96 | pred = [mappings['id_to_tag'][p] for p in pred] 97 | # Output tags in the IOB2 format 98 | if parameters['tag_scheme'] == 'iobes': 99 | pred = iobes_iob(pred) 100 | # Write tags 101 | assert len(pred) == len(eval_sentences[i+j]) 102 | f_output.write('%s\n\n' % '\n'.join('%s%s%s' % (' '.join(w), ' ', z) 103 | for w, z in zip(eval_sentences[i+j], 104 | pred))) 105 | if (i + j + 1) % 500 == 0: 106 | print(i+j+1) 107 | 108 | end = time.time() # epoch end time 109 | print('time elapssed: %f seconds' % round( 110 | (end - since), 2)) 111 | 112 | -------------------------------------------------------------------------------- /subroot/dnn_pytorch/utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | import re 3 | import io 4 | import itertools 5 | import codecs 6 | import time 7 | import numpy as np 8 | import collections 9 | from torch.autograd import Variable 10 | from dnn_utils import LongTensor,FloatTensor 11 | 12 | try: 13 | import _pickle as cPickle 14 | except ImportError: 15 | import cPickle 16 | 17 | models_path = "./models" 18 | 19 | 20 | def create_dico(item_list): 21 | """ 22 | Create a dictionary of items from a list of list of items. 23 | """ 24 | assert type(item_list) is list 25 | dico = {} 26 | for items in item_list: 27 | for item in items: 28 | if item not in dico: 29 | dico[item] = 1 30 | else: 31 | dico[item] += 1 32 | return dico 33 | 34 | 35 | def create_mapping(dico): 36 | """ 37 | Create a mapping (item to ID / ID to item) from a dictionary. 38 | Items are ordered by decreasing frequency. 39 | """ 40 | sorted_items = sorted(dico.items(), key=lambda x: (-x[1], x[0])) 41 | id_to_item = {i: v[0] for i, v in enumerate(sorted_items)} 42 | item_to_id = {v: k for k, v in id_to_item.items()} 43 | return item_to_id, id_to_item 44 | 45 | 46 | def zero_digits(s): 47 | """ 48 | Replace every digit in a string by a zero. 49 | """ 50 | return re.sub('\d', '0', s) 51 | 52 | def insert_singletons(words, singletons, p=0.5): 53 | """ 54 | Replace singletons by the unknown word with a probability p. 55 | """ 56 | new_words = [] 57 | for word in words: 58 | if word in singletons and np.random.uniform() < p: 59 | new_words.append(0) 60 | else: 61 | new_words.append(word) 62 | return new_words 63 | 64 | 65 | def pad_word(inputs, seq_len): 66 | # get the max sequence length in the batch 67 | max_len = seq_len[0] 68 | 69 | padding = np.zeros_like([inputs[0][0]]).tolist() 70 | 71 | padded_inputs = [] 72 | for item in inputs: 73 | padded_inputs.append(item + padding * (max_len - len(item))) 74 | 75 | return padded_inputs 76 | 77 | 78 | def pad_chars(inputs): 79 | chained_chars = list(itertools.chain.from_iterable(inputs)) 80 | 81 | char_index_mapping, chars = zip( 82 | *[item for item in sorted( 83 | enumerate(chained_chars), key=lambda x: len(x[1]), reverse=True 84 | )] 85 | ) 86 | char_index_mapping = {v: i for i, v in enumerate(char_index_mapping)} 87 | 88 | char_len = [len(c) for c in chars] 89 | 90 | chars = pad_word(chars, char_len) 91 | 92 | # pad chars to length of 25 if max char len less than 25 93 | # char CNN layer requires at least 25 chars 94 | if len(chars[0]) < 25: 95 | chars = [c + [0]*(25-len(c)) for c in chars] 96 | 97 | return chars, char_index_mapping, char_len 98 | 99 | 100 | def create_input(data, parameters, add_label=True): 101 | """ 102 | Take sentence data and return an input for 103 | the training or the evaluation function. 104 | """ 105 | # sort data by sequence length 106 | seq_index_mapping, data = zip(*[item for item in sorted(enumerate(data), key=lambda x: len(x[1]['words']), reverse=True)]) 107 | seq_index_mapping = {v: i for i, v in enumerate(seq_index_mapping)} 108 | 109 | inputs = collections.defaultdict(list) 110 | seq_len = [] 111 | 112 | for d in data: 113 | words = d['words'] 114 | 115 | seq_len.append(len(words)) 116 | 117 | chars = d['chars'] 118 | 119 | if parameters['word_dim']: 120 | inputs['words'].append(words) 121 | if parameters['char_dim']: 122 | inputs['chars'].append(chars) 123 | if parameters['cap_dim']: 124 | caps = d['caps'] 125 | inputs['caps'].append(caps) 126 | 127 | # boliang: add expectation features into input 128 | if d['feats']: 129 | inputs['feats'].append(d['feats']) 130 | 131 | if add_label: 132 | tags = d['tags'] 133 | inputs['tags'].append(tags) 134 | 135 | char_index_mapping = [] 136 | char_len = [] 137 | for k, v in inputs.items(): 138 | if k == 'chars': 139 | padded_chars, char_index_mapping, char_len = pad_chars(v) 140 | inputs[k] = padded_chars 141 | else: 142 | inputs[k] = pad_word(v, seq_len) 143 | 144 | # convert inputs and labels to Variable 145 | for k, v in inputs.items(): 146 | inputs[k] = Variable(LongTensor(v)) 147 | 148 | return inputs, seq_index_mapping, char_index_mapping, seq_len, char_len 149 | 150 | def count_sentence_predicate(sentence): 151 | count = 0 152 | for item in sentence: 153 | if item[-2] == 'Y': 154 | count += 1 155 | return count 156 | 157 | def evaluate(phase, preds, dataset, id_to_tag, eval_out_dir=None): 158 | """ 159 | Evaluate current model using CoNLL script. 160 | """ 161 | n_tags = len(id_to_tag) 162 | 163 | tp = 0 164 | fp = 0 165 | fn = 0 166 | correct = 0 167 | total = 0 168 | 169 | output = [] 170 | for d, p in zip(dataset, preds): 171 | 172 | assert len(d['words']) == len(p) 173 | str_words = d['str_words'] 174 | p_tags = [id_to_tag[y_pred] for y_pred in p] 175 | r_tags = [id_to_tag[y_real] for y_real in d['tags']] 176 | 177 | block = [] 178 | for i in range(len(p_tags)): 179 | if r_tags[i]!='0' and p_tags[i] == r_tags[i]: 180 | tp += 1 181 | if r_tags[i]!='0' and p_tags[i] != r_tags[i]: 182 | fp += 1 183 | if r_tags[i]=='0' and p_tags[i] != r_tags[i]: 184 | fn += 1 185 | if p_tags[i] == r_tags[i]: 186 | correct += 1 187 | total += 1 188 | block.append([r_tags[i],p_tags[i]]) 189 | output.append(block) 190 | 191 | p = tp / (tp + fp + 1e-13) 192 | 193 | r = tp / (tp + fn + 1e-13) 194 | 195 | f1 = 2 * p * r / ( p + r + 1e-13) 196 | 197 | acc = correct / total 198 | 199 | # Global accuracy 200 | print("Acc:%.5f%% P:%.5f R:%.5f F1:%.5f" % ( 201 | acc * 100, p * 100, r * 100, f1 * 100 202 | )) 203 | 204 | if eval_out_dir is not None: 205 | output_file = os.path.join(eval_out_dir,'{}_predicate_{:.2f}.pred'.format(phase,p*100)) 206 | with open(output_file, 'w') as fout: 207 | for block in output: 208 | for line in block: 209 | fout.write('\t'.join(line)) 210 | fout.write('\n') 211 | fout.write('\n') 212 | 213 | return f1, acc 214 | 215 | 216 | ######################################################################################################################## 217 | # temporal script below 218 | # 219 | def load_exp_feats(fp): 220 | bio_feats_fp = fp 221 | res = [] 222 | for sent in io.open(bio_feats_fp, 'r', -1, 'utf-8').read().split('\n\n'): 223 | sent_feats = [] 224 | for line in sent.splitlines(): 225 | feats = line.split('\t')[1:] 226 | sent_feats.append(feats) 227 | res.append(sent_feats) 228 | 229 | return res 230 | 231 | 232 | class Tee(object): 233 | def __init__(self, *files): 234 | self.files = files 235 | 236 | def write(self, obj): 237 | for f in self.files: 238 | f.write(obj) 239 | f.flush() # If you want the output to be visible immediately 240 | 241 | def flush(self) : 242 | for f in self.files: 243 | f.flush() 244 | 245 | 246 | 247 | 248 | -------------------------------------------------------------------------------- /subroot/subroot/preprocess.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | def preprocess(): 4 | raw_train_file = os.path.join(os.path.dirname(__file__),'./data/ptb-sd/train_pro.conll') 5 | raw_dev_file = os.path.join(os.path.dirname(__file__),'./data/ptb-sd/dev_pro.conll') 6 | raw_test_file = os.path.join(os.path.dirname(__file__),'./data/ptb-sd/test_pro.conll') 7 | 8 | predicate_train_file = os.path.join(os.path.dirname(__file__),'./data/subroot-train.txt') 9 | predicate_dev_file = os.path.join(os.path.dirname(__file__),'./data/subroot-dev.txt') 10 | predicate_test_file = os.path.join(os.path.dirname(__file__),'./data/subroot-test.txt') 11 | 12 | with open(raw_train_file, 'r') as f: 13 | with open(predicate_train_file, 'w') as fo: 14 | data = f.readlines() 15 | 16 | # read data 17 | sentence_data = [] 18 | sentence = [] 19 | for line in data: 20 | if len(line.strip()) > 0: 21 | line = line.strip().split('\t') 22 | sentence.append(line) 23 | else: 24 | sentence_data.append(sentence) 25 | sentence = [] 26 | if len(sentence)>0: 27 | sentence_data.append(sentence) 28 | sentence = [] 29 | 30 | # process data 31 | # copy the data by predicate num 32 | train_data = [] 33 | for sentence in sentence_data: 34 | lines = [] 35 | for i in range(len(sentence)): 36 | is_subroot = '0' 37 | if sentence[i][6] == '0': 38 | is_subroot = '1' 39 | word = sentence[i][1].lower() 40 | # is_number = False 41 | # for c in word: 42 | # if c.isdigit(): 43 | # is_number = True 44 | # break 45 | # if is_number: 46 | # word = 'number' 47 | lines.append([word, sentence[i][4], is_subroot]) 48 | train_data.append(lines) 49 | 50 | for sentence in train_data: 51 | fo.write('\n'.join(['\t'.join(line) for line in sentence])) 52 | fo.write('\n\n') 53 | 54 | with open(raw_dev_file, 'r') as f: 55 | with open(predicate_dev_file, 'w') as fo: 56 | data = f.readlines() 57 | 58 | # read data 59 | sentence_data = [] 60 | sentence = [] 61 | for line in data: 62 | if len(line.strip()) > 0: 63 | line = line.strip().split('\t') 64 | sentence.append(line) 65 | else: 66 | sentence_data.append(sentence) 67 | sentence = [] 68 | if len(sentence)>0: 69 | sentence_data.append(sentence) 70 | sentence = [] 71 | 72 | # process data 73 | # copy the data by predicate num 74 | dev_data = [] 75 | for sentence in sentence_data: 76 | lines = [] 77 | for i in range(len(sentence)): 78 | is_subroot = '0' 79 | if sentence[i][6] == '0': 80 | is_subroot = '1' 81 | word = sentence[i][1].lower() 82 | # is_number = False 83 | # for c in word: 84 | # if c.isdigit(): 85 | # is_number = True 86 | # break 87 | # if is_number: 88 | # word = 'number' 89 | lines.append([word, sentence[i][4], is_subroot]) 90 | dev_data.append(lines) 91 | 92 | for sentence in dev_data: 93 | fo.write('\n'.join(['\t'.join(line) for line in sentence])) 94 | fo.write('\n\n') 95 | 96 | with open(raw_test_file, 'r') as f: 97 | with open(predicate_test_file, 'w') as fo: 98 | data = f.readlines() 99 | 100 | # read data 101 | sentence_data = [] 102 | sentence = [] 103 | for line in data: 104 | if len(line.strip()) > 0: 105 | line = line.strip().split('\t') 106 | sentence.append(line) 107 | else: 108 | sentence_data.append(sentence) 109 | sentence = [] 110 | if len(sentence)>0: 111 | sentence_data.append(sentence) 112 | sentence = [] 113 | 114 | # process data 115 | # copy the data by predicate num 116 | test_data = [] 117 | for sentence in sentence_data: 118 | lines = [] 119 | for i in range(len(sentence)): 120 | is_subroot = '0' 121 | if sentence[i][6] == '0': 122 | is_subroot = '1' 123 | word = sentence[i][1].lower() 124 | # is_number = False 125 | # for c in word: 126 | # if c.isdigit(): 127 | # is_number = True 128 | # break 129 | # if is_number: 130 | # word = 'number' 131 | lines.append([word, sentence[i][4], is_subroot]) 132 | test_data.append(lines) 133 | 134 | for sentence in test_data: 135 | fo.write('\n'.join(['\t'.join(line) for line in sentence])) 136 | fo.write('\n\n') 137 | 138 | 139 | if __name__ == '__main__': 140 | preprocess() -------------------------------------------------------------------------------- /subroot/subroot/stat.py: -------------------------------------------------------------------------------- 1 | import math 2 | 3 | def stat_f1(pred_file): 4 | with open(pred_file,'r') as f: 5 | data_lines = f.readlines() 6 | 7 | # split by sentence 8 | pred_data = [] 9 | sentence_data = [] 10 | for line in data_lines: 11 | if len(line.strip()) > 0: 12 | sentence_data.append(line.strip().split("\t")) 13 | else: 14 | pred_data.append(sentence_data) 15 | sentence_data = [] 16 | if len(sentence_data)>0: 17 | pred_data.append(sentence_data) 18 | sentence_data = [] 19 | 20 | tps = [0 for _ in range(7)] 21 | fps = [0 for _ in range(7)] 22 | fns = [0 for _ in range(7)] 23 | f1s = [0 for _ in range(7)] 24 | for sentence in pred_data: 25 | idx = math.ceil(len(sentence)/10)-1 26 | if idx >= 7: 27 | continue 28 | for line in sentence: 29 | if line[0]!='0' and line[1] == line[0]: 30 | tps[idx] += 1 31 | if line[0]!='0' and line[1] != line[0]: 32 | fps[idx] += 1 33 | if line[0]=='0' and line[1] != line[0]: 34 | fns[idx] += 1 35 | 36 | for i in range(7): 37 | p = tps[i] / (tps[i] + fps[i] + 1e-13) 38 | r = tps[i] / (tps[i] + fns[i] + 1e-13) 39 | f1s[i] = 2 * p * r / ( p + r + 1e-13) 40 | 41 | return f1s 42 | 43 | 44 | if __name__ == '__main__': 45 | print('\ndev:') 46 | print(stat_f1('../result/dev_predicate_96.53.pred')) 47 | 48 | print('\ntest:') 49 | print(stat_f1('../result/test_predicate_95.45.pred')) -------------------------------------------------------------------------------- /subroot/subroot/test.py: -------------------------------------------------------------------------------- 1 | import subprocess 2 | import os 3 | 4 | 5 | dnn_tagger_script = '../../dnn_pytorch/seq_labeling/tag.py' 6 | model_dir = 'model/tag_scheme=iobes,zeros=True,char_dim=25,char_lstm_dim=25,char_conv_channel=25,word_dim=100,word_lstm_dim=100,feat_dim=5,feat_column=1,crf=True,dropout=0.5,lr_method=sgd-init_lr=.005-lr_decay_epoch=100,num_epochs=100,batch_size=20/best_model.pth.tar' 7 | 8 | input_file = 'data/eng.testb.bio' 9 | output_file = 'result/eng.test.output' 10 | 11 | cmd = [ 12 | 'python3', 13 | dnn_tagger_script, 14 | '--model', model_dir, 15 | '--input', input_file, 16 | '--output', output_file, 17 | '--batch_size', '50', 18 | '--gpu', '0' 19 | ] 20 | 21 | # set OMP threads to 1 22 | os.environ.update({'OMP_NUM_THREADS': '1'}) 23 | # set which gpu to use if gpu option is turned on 24 | gpu_device = '0' 25 | os.environ.update({'CUDA_VISIBLE_DEVICES': gpu_device}) 26 | 27 | print(' '.join(cmd)) 28 | subprocess.call(cmd) -------------------------------------------------------------------------------- /subroot/subroot/train.py: -------------------------------------------------------------------------------- 1 | import os 2 | import subprocess 3 | 4 | 5 | train = './subroot/data/subroot-train.txt' 6 | dev = './subroot/data/subroot-dev.txt' 7 | test = './subroot/data/subroot-test.txt' 8 | 9 | model_dir = 'model/' 10 | result_dir = 'result/' 11 | 12 | # use word2vec to generate pre-trained embeddings. 13 | # tutorial: https://code.google.com/archive/p/word2vec/ 14 | pre_emb = './subroot/data/glove.6B.100d.txt' 15 | # pre_emb = '/nas/data/m1/zhangb8/ml/data/embeddings/lample_pretrained/eng.Skip100' 16 | 17 | # 18 | # run command 19 | # 20 | script = 'dnn_pytorch/train.py' 21 | cmd = [ 22 | 'python3', 23 | script, 24 | # data settings 25 | '--train', train, 26 | '--dev', dev, 27 | '--test', test, 28 | '--model_dp', model_dir, 29 | '--result_dp', result_dir, 30 | # parameter settings 31 | '--lower', '0', 32 | '--zeros', '1', 33 | '--char_dim', '25', 34 | '--char_lstm_dim', '25', 35 | '--char_conv_channel', '25', 36 | '--word_dim', '100', 37 | '--word_lstm_dim', '100', 38 | '--pre_emb', pre_emb, 39 | '--all_emb', '0', 40 | '--cap_dim', '0', 41 | '--feat_dim', '100', 42 | '--feat_column', '1', 43 | '--crf', '1', 44 | '--dropout', '0.5', 45 | '--lr_method', 'sgd-init_lr=.005-lr_decay_epoch=100', 46 | '--batch_size', '72', 47 | '--gpu', '1', 48 | ] 49 | 50 | # set OMP threads to 1 51 | os.environ.update({'OMP_NUM_THREADS': '1'}) 52 | # set which gpu to use if gpu option is turned on 53 | gpu_device = '0' 54 | os.environ.update({'CUDA_VISIBLE_DEVICES': gpu_device}) 55 | 56 | print(' '.join(cmd)) 57 | #subprocess.call(cmd, env=os.environ) 58 | -------------------------------------------------------------------------------- /tools/README.md: -------------------------------------------------------------------------------- 1 | This directly contains scripts and tools adopted from other open source projects such as Apache Joshua and Moses Decoder. 2 | 3 | TODO: credit the authors and resolve license issues (if any) 4 | -------------------------------------------------------------------------------- /tools/average_models.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | import argparse 3 | import torch 4 | 5 | 6 | def average_models(model_files): 7 | vocab = None 8 | opt = None 9 | epoch = None 10 | avg_model = None 11 | avg_generator = None 12 | 13 | for i, model_file in enumerate(model_files): 14 | m = torch.load(model_file) 15 | model_weights = m['model'] 16 | generator_weights = m['generator'] 17 | 18 | if i == 0: 19 | vocab, opt, epoch = m['vocab'], m['opt'], m['epoch'] 20 | avg_model = model_weights 21 | avg_generator = generator_weights 22 | else: 23 | for (k, v) in avg_model.items(): 24 | avg_model[k].mul_(i).add_(model_weights[k]).div_(i + 1) 25 | 26 | for (k, v) in avg_generator.items(): 27 | avg_generator[k].mul_(i).add_(generator_weights[k]).div_(i + 1) 28 | 29 | final = {"vocab": vocab, "opt": opt, "epoch": epoch, "optim": None, 30 | "generator": avg_generator, "model": avg_model} 31 | return final 32 | 33 | 34 | def main(): 35 | parser = argparse.ArgumentParser(description="") 36 | parser.add_argument("-models", "-m", nargs="+", required=True, 37 | help="List of models") 38 | parser.add_argument("-output", "-o", required=True, 39 | help="Output file") 40 | opt = parser.parse_args() 41 | 42 | final = average_models(opt.models) 43 | torch.save(final, opt.output) 44 | 45 | 46 | if __name__ == "__main__": 47 | main() 48 | -------------------------------------------------------------------------------- /tools/bpe_pipeline.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | # Author : Thamme Gowda 3 | # Created : Nov 06, 2017 4 | 5 | ONMT="$( cd "$( dirname "${BASH_SOURCE[0]}" )/.." && pwd )" 6 | 7 | #======= EXPERIMENT SETUP ====== 8 | # Activate python environment if needed 9 | source ~/.bashrc 10 | # source activate py3 11 | 12 | # update these variables 13 | NAME="run1" 14 | OUT="onmt-runs/$NAME" 15 | 16 | DATA="$ONMT/onmt-runs/data" 17 | TRAIN_SRC=$DATA/*train.src 18 | TRAIN_TGT=$DATA/*train.tgt 19 | VALID_SRC=$DATA/*dev.src 20 | VALID_TGT=$DATA/*dev.tgt 21 | TEST_SRC=$DATA/*test.src 22 | TEST_TGT=$DATA/*test.tgt 23 | 24 | BPE="" # default 25 | BPE="src" # src, tgt, src+tgt 26 | 27 | # applicable only when BPE="src" or "src+tgt" 28 | BPE_SRC_OPS=10000 29 | 30 | # applicable only when BPE="tgt" or "src+tgt" 31 | BPE_TGT_OPS=10000 32 | 33 | GPUARG="" # default 34 | GPUARG="0" 35 | 36 | 37 | #====== EXPERIMENT BEGIN ====== 38 | 39 | # Check if input exists 40 | for f in $TRAIN_SRC $TRAIN_TGT $VALID_SRC $VALID_TGT $TEST_SRC $TEST_TGT; do 41 | if [[ ! -f "$f" ]]; then 42 | echo "Input File $f doesnt exist. Please fix the paths" 43 | exit 1 44 | fi 45 | done 46 | 47 | function lines_check { 48 | l1=`wc -l $1` 49 | l2=`wc -l $2` 50 | if [[ $l1 != $l2 ]]; then 51 | echo "ERROR: Record counts doesnt match between: $1 and $2" 52 | exit 2 53 | fi 54 | } 55 | lines_check $TRAIN_SRC $TRAIN_TGT 56 | lines_check $VALID_SRC $VALID_TGT 57 | lines_check $TEST_SRC $TEST_TGT 58 | 59 | 60 | echo "Output dir = $OUT" 61 | [ -d $OUT ] || mkdir -p $OUT 62 | [ -d $OUT/data ] || mkdir -p $OUT/data 63 | [ -d $OUT/models ] || mkdir $OUT/models 64 | [ -d $OUT/test ] || mkdir -p $OUT/test 65 | 66 | 67 | echo "Step 1a: Preprocess inputs" 68 | if [[ "$BPE" == *"src"* ]]; then 69 | echo "BPE on source" 70 | # Here we could use more monolingual data 71 | $ONMT/tools/learn_bpe.py -s $BPE_SRC_OPS < $TRAIN_SRC > $OUT/data/bpe-codes.src 72 | 73 | $ONMT/tools/apply_bpe.py -c $OUT/data/bpe-codes.src < $TRAIN_SRC > $OUT/data/train.src 74 | $ONMT/tools/apply_bpe.py -c $OUT/data/bpe-codes.src < $VALID_SRC > $OUT/data/valid.src 75 | $ONMT/tools/apply_bpe.py -c $OUT/data/bpe-codes.src < $TEST_SRC > $OUT/data/test.src 76 | else 77 | ln -sf $TRAIN_SRC $OUT/data/train.src 78 | ln -sf $VALID_SRC $OUT/data/valid.src 79 | ln -sf $TEST_SRC $OUT/data/test.src 80 | fi 81 | 82 | 83 | if [[ "$BPE" == *"tgt"* ]]; then 84 | echo "BPE on target" 85 | # Here we could use more monolingual data 86 | $ONMT/tools/learn_bpe.py -s $BPE_SRC_OPS < $TRAIN_TGT > $OUT/data/bpe-codes.tgt 87 | 88 | $ONMT/tools/apply_bpe.py -c $OUT/data/bpe-codes.tgt < $TRAIN_TGT > $OUT/data/train.tgt 89 | $ONMT/tools/apply_bpe.py -c $OUT/data/bpe-codes.tgt < $VALID_TGT > $OUT/data/valid.tgt 90 | #$ONMT/tools/apply_bpe.py -c $OUT/data/bpe-codes.tgt < $TEST_TGT > $OUT/data/test.tgt 91 | # We dont touch the test References, No BPE on them! 92 | ln -sf $TEST_TGT $OUT/data/test.tgt 93 | else 94 | ln -sf $TRAIN_TGT $OUT/data/train.tgt 95 | ln -sf $VALID_TGT $OUT/data/valid.tgt 96 | ln -sf $TEST_TGT $OUT/data/test.tgt 97 | fi 98 | 99 | 100 | #: < maxv) {maxv=score; max=$0}} END{ print max}'` 124 | echo "Chosen Model = $model" 125 | if [[ -z "$model" ]]; then 126 | echo "Model not found. Looked in $OUT/models/" 127 | exit 1 128 | fi 129 | 130 | GPU_OPTS="" 131 | if [ ! -z $GPUARG ]; then 132 | GPU_OPTS="-gpu $GPUARG" 133 | fi 134 | 135 | echo "Step 3a: Translate Test" 136 | python $ONMT/translate.py -model $model \ 137 | -src $OUT/data/test.src \ 138 | -output $OUT/test/test.out \ 139 | -replace_unk -verbose $GPU_OPTS > $OUT/test/test.log 140 | 141 | echo "Step 3b: Translate Dev" 142 | python $ONMT/translate.py -model $model \ 143 | -src $OUT/data/valid.src \ 144 | -output $OUT/test/valid.out \ 145 | -replace_unk -verbose $GPU_OPTS > $OUT/test/valid.log 146 | 147 | if [[ "$BPE" == *"tgt"* ]]; then 148 | echo "BPE decoding/detokenising target to match with references" 149 | mv $OUT/test/test.out{,.bpe} 150 | mv $OUT/test/valid.out{,.bpe} 151 | cat $OUT/test/valid.out.bpe | sed -E 's/(@@ )|(@@ ?$)//g' > $OUT/test/valid.out 152 | cat $OUT/test/test.out.bpe | sed -E 's/(@@ )|(@@ ?$)//g' > $OUT/test/test.out 153 | fi 154 | 155 | echo "Step 4a: Evaluate Test" 156 | $ONMT/tools/multi-bleu-detok.perl $OUT/data/test.tgt < $OUT/test/test.out > $OUT/test/test.tc.bleu 157 | $ONMT/tools/multi-bleu-detok.perl -lc $OUT/data/test.tgt < $OUT/test/test.out > $OUT/test/test.lc.bleu 158 | 159 | echo "Step 4b: Evaluate Dev" 160 | $ONMT/tools/multi-bleu-detok.perl $OUT/data/valid.tgt < $OUT/test/valid.out > $OUT/test/valid.tc.bleu 161 | $ONMT/tools/multi-bleu-detok.perl -lc $OUT/data/valid.tgt < $OUT/test/valid.out > $OUT/test/valid.lc.bleu 162 | 163 | #===== EXPERIMENT END ====== 164 | -------------------------------------------------------------------------------- /tools/detokenize.perl: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env perl 2 | 3 | # Note: retrieved from https://github.com/apache/incubator-joshua/blob/master/scripts/preparation/detokenize.pl 4 | 5 | # Licensed to the Apache Software Foundation (ASF) under one or more 6 | # contributor license agreements. See the NOTICE file distributed with 7 | # this work for additional information regarding copyright ownership. 8 | # The ASF licenses this file to You under the Apache License, Version 2.0 9 | # (the "License"); you may not use this file except in compliance with 10 | # the License. You may obtain a copy of the License at 11 | # 12 | # http://www.apache.org/licenses/LICENSE-2.0 13 | # 14 | # Unless required by applicable law or agreed to in writing, software 15 | # distributed under the License is distributed on an "AS IS" BASIS, 16 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 17 | # See the License for the specific language governing permissions and 18 | # limitations under the License. 19 | 20 | use warnings; 21 | use strict; 22 | 23 | # Sample De-Tokenizer 24 | # written by Josh Schroeder, based on code by Philipp Koehn 25 | # modified later by ByungGyu Ahn, bahn@cs.jhu.edu, Luke Orland 26 | 27 | binmode(STDIN, ":utf8"); 28 | binmode(STDOUT, ":utf8"); 29 | 30 | my $language = "en"; 31 | my $QUIET = 1; 32 | my $HELP = 0; 33 | 34 | while (@ARGV) { 35 | $_ = shift; 36 | /^-l$/ && ($language = shift, next); 37 | /^-v$/ && ($QUIET = 0, next); 38 | /^-h$/ && ($HELP = 1, next); 39 | } 40 | 41 | if ($HELP) { 42 | print "Usage ./detokenizer.perl (-l [en|de|...]) < tokenizedfile > detokenizedfile\n"; 43 | exit; 44 | } 45 | if (!$QUIET) { 46 | print STDERR "Detokenizer Version 1.1\n"; 47 | print STDERR "Language: $language\n"; 48 | } 49 | 50 | while() { 51 | if (/^<.+>$/ || /^\s*$/) { 52 | #don't try to detokenize XML/HTML tag lines 53 | print $_; 54 | } 55 | else { 56 | print &detokenize($_); 57 | } 58 | } 59 | 60 | sub detokenize { 61 | my($text) = @_; 62 | chomp($text); 63 | $text = " $text "; 64 | 65 | # convert curly quotes to ASCII e.g. ‘“”’ 66 | $text =~ s/\x{2018}/'/gs; 67 | $text =~ s/\x{2019}/'/gs; 68 | $text =~ s/\x{201c}/"/gs; 69 | $text =~ s/\x{201d}/"/gs; 70 | $text =~ s/\x{e2}\x{80}\x{98}/'/gs; 71 | $text =~ s/\x{e2}\x{80}\x{99}/'/gs; 72 | $text =~ s/\x{e2}\x{80}\x{9c}/"/gs; 73 | $text =~ s/\x{e2}\x{80}\x{9d}/"/gs; 74 | 75 | $text =~ s/ '\s+' / " /g; 76 | $text =~ s/ ` / ' /g; 77 | $text =~ s/ ' / ' /g; 78 | $text =~ s/ `` / " /g; 79 | $text =~ s/ '' / " /g; 80 | 81 | # replace the pipe character, which is 82 | # a special reserved character in Moses 83 | $text =~ s/ -PIPE- / \| /g; 84 | 85 | $text =~ s/ -LRB- / \( /g; 86 | $text =~ s/ -RRB- / \) /g; 87 | $text =~ s/ -LSB- / \[ /g; 88 | $text =~ s/ -RSB- / \] /g; 89 | $text =~ s/ -LCB- / \{ /g; 90 | $text =~ s/ -RCB- / \} /g; 91 | $text =~ s/ -lrb- / \( /g; 92 | $text =~ s/ -rrb- / \) /g; 93 | $text =~ s/ -lsb- / \[ /g; 94 | $text =~ s/ -rsb- / \] /g; 95 | $text =~ s/ -lcb- / \{ /g; 96 | $text =~ s/ -rcb- / \} /g; 97 | 98 | $text =~ s/ 'll /'ll /g; 99 | $text =~ s/ 're /'re /g; 100 | $text =~ s/ 've /'ve /g; 101 | $text =~ s/ n't /n't /g; 102 | $text =~ s/ 'LL /'LL /g; 103 | $text =~ s/ 'RE /'RE /g; 104 | $text =~ s/ 'VE /'VE /g; 105 | $text =~ s/ N'T /N'T /g; 106 | $text =~ s/ can not / cannot /g; 107 | $text =~ s/ Can not / Cannot /g; 108 | 109 | # just in case the contraction was not properly treated 110 | $text =~ s/ ' ll /'ll /g; 111 | $text =~ s/ ' re /'re /g; 112 | $text =~ s/ ' ve /'ve /g; 113 | $text =~ s/n ' t /n't /g; 114 | $text =~ s/ ' LL /'LL /g; 115 | $text =~ s/ ' RE /'RE /g; 116 | $text =~ s/ ' VE /'VE /g; 117 | $text =~ s/N ' T /N'T /g; 118 | 119 | my $word; 120 | my $i; 121 | my @words = split(/ /,$text); 122 | $text = ""; 123 | my %quoteCount = ("\'"=>0,"\""=>0); 124 | my $prependSpace = " "; 125 | for ($i=0;$i<(scalar(@words));$i++) { 126 | if ($words[$i] =~ /^[\p{IsSc}]+$/) { 127 | #perform shift on currency 128 | if (($i<(scalar(@words)-1)) && ($words[$i+1] =~ /^[0-9]/)) { 129 | $text = $text.$prependSpace.$words[$i]; 130 | $prependSpace = ""; 131 | } else { 132 | $text=$text.$words[$i]; 133 | $prependSpace = " "; 134 | } 135 | } elsif ($words[$i] =~ /^[\(\[\{\¿\¡]+$/) { 136 | #perform right shift on random punctuation items 137 | $text = $text.$prependSpace.$words[$i]; 138 | $prependSpace = ""; 139 | } elsif ($words[$i] =~ /^[\,\.\?\!\:\;\\\%\}\]\)]+$/){ 140 | #perform left shift on punctuation items 141 | $text=$text.$words[$i]; 142 | $prependSpace = " "; 143 | } elsif (($language eq "en") && ($i>0) && ($words[$i] =~ /^[\'][\p{IsAlpha}]/) && ($words[$i-1] =~ /[\p{IsAlnum}]$/)) { 144 | #left-shift the contraction for English 145 | $text=$text.$words[$i]; 146 | $prependSpace = " "; 147 | } elsif (($language eq "en") && ($i>0) && ($i<(scalar(@words)-1)) && ($words[$i] eq "&") && ($words[$i-1] =~ /^[A-Z]$/) && ($words[$i+1] =~ /^[A-Z]$/)) { 148 | #some contraction with an ampersand e.g. "R&D" 149 | $text .= $words[$i]; 150 | $prependSpace = ""; 151 | } elsif (($language eq "fr") && ($i<(scalar(@words)-1)) && ($words[$i] =~ /[\p{IsAlpha}][\']$/) && ($words[$i+1] =~ /^[\p{IsAlpha}]/)) { 152 | #right-shift the contraction for French 153 | $text = $text.$prependSpace.$words[$i]; 154 | $prependSpace = ""; 155 | } elsif ($words[$i] =~ /^[\'\"]+$/) { 156 | #combine punctuation smartly 157 | if (($quoteCount{$words[$i]} % 2) eq 0) { 158 | if(($language eq "en") && ($words[$i] eq "'") && ($i > 0) && ($words[$i-1] =~ /[s]$/)) { 159 | #single quote for posesssives ending in s... "The Jones' house" 160 | #left shift 161 | $text=$text.$words[$i]; 162 | $prependSpace = " "; 163 | } elsif (($language eq "en") && ($words[$i] eq "'") && ($i < (scalar(@words)-1)) && ($words[$i+1] eq "s")) { 164 | #single quote for possessive construction. "John's" 165 | $text .= $words[$i]; 166 | $prependSpace = ""; 167 | } elsif (($quoteCount{$words[$i]} == 0) && 168 | ($language eq "en") && ($words[$i] eq '"') && ($i>1) && ($words[$i-1] =~ /^[,.]$/) && ($words[$i-2] ne "said")) { 169 | #emergency case in which the opening quote is missing 170 | #ending double quote for direct quotes. e.g. Blah," he said. but not like he said, "Blah. 171 | $text .= $words[$i]; 172 | $prependSpace = " "; 173 | } elsif (($language eq "en") && ($words[$i] eq '"') && ($i < (scalar(@words)-1)) && ($words[$i+1] =~ /^[,.]$/)) { 174 | $text .= $words[$i]; 175 | $prependSpace = " "; 176 | } else { 177 | #right shift 178 | $text = $text.$prependSpace.$words[$i]; 179 | $prependSpace = ""; 180 | $quoteCount{$words[$i]} = $quoteCount{$words[$i]} + 1; 181 | 182 | } 183 | } else { 184 | #left shift 185 | $text=$text.$words[$i]; 186 | $prependSpace = " "; 187 | $quoteCount{$words[$i]} = $quoteCount{$words[$i]} + 1; 188 | 189 | } 190 | 191 | } else { 192 | $text=$text.$prependSpace.$words[$i]; 193 | $prependSpace = " "; 194 | } 195 | } 196 | 197 | #clean continuing spaces 198 | $text =~ s/ +/ /g; 199 | 200 | #delete spaces around double angle brackets «» 201 | # Uh-oh. not a good idea. it is not consistent. 202 | $text =~ s/(\x{c2}\x{ab}|\x{ab}) /$1/g; 203 | $text =~ s/ (\x{c2}\x{bb}|\x{bb})/$1/g; 204 | 205 | # delete spaces around all other special characters 206 | # Uh-oh. not a good idea. "Men&Women" 207 | #$text =~ s/ ([^\p{IsAlnum}\s\.\'\`\,\-\"\|]) /$1/g; 208 | $text =~ s/ \/ /\//g; 209 | 210 | # clean up spaces at head and tail of each line as well as any double-spacing 211 | $text =~ s/\n /\n/g; 212 | $text =~ s/ \n/\n/g; 213 | $text =~ s/^ //g; 214 | $text =~ s/ $//g; 215 | 216 | #add trailing break 217 | $text .= "\n" unless $text =~ /\n$/; 218 | 219 | return $text; 220 | } 221 | -------------------------------------------------------------------------------- /tools/embeddings_to_torch.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | from __future__ import print_function 4 | from __future__ import division 5 | import six 6 | import sys 7 | import numpy as np 8 | import argparse 9 | import torch 10 | 11 | 12 | def get_vocabs(dict_file): 13 | vocabs = torch.load(dict_file) 14 | 15 | enc_vocab, dec_vocab = None, None 16 | 17 | # the vocab object is a list of tuple (name, torchtext.Vocab) 18 | # we iterate over this list and associate vocabularies based on the name 19 | for vocab in vocabs: 20 | if vocab[0] == 'src': 21 | enc_vocab = vocab[1] 22 | if vocab[0] == 'tgt': 23 | dec_vocab = vocab[1] 24 | assert type(None) not in [type(enc_vocab), type(dec_vocab)] 25 | 26 | print("From: %s" % dict_file) 27 | print("\t* source vocab: %d words" % len(enc_vocab)) 28 | print("\t* target vocab: %d words" % len(dec_vocab)) 29 | 30 | return enc_vocab, dec_vocab 31 | 32 | 33 | def get_embeddings(file_enc, opt, flag): 34 | embs = dict() 35 | if flag == 'enc': 36 | for (i, l) in enumerate(open(file_enc, 'rb')): 37 | if i < opt.skip_lines: 38 | continue 39 | if not l: 40 | break 41 | if len(l) == 0: 42 | continue 43 | 44 | l_split = l.decode('utf8').strip().split(' ') 45 | if len(l_split) == 2: 46 | continue 47 | embs[l_split[0]] = [float(em) for em in l_split[1:]] 48 | print("Got {} encryption embeddings from {}".format(len(embs), 49 | file_enc)) 50 | else: 51 | 52 | for (i, l) in enumerate(open(file_enc, 'rb')): 53 | if not l: 54 | break 55 | if len(l) == 0: 56 | continue 57 | 58 | l_split = l.decode('utf8').strip().split(' ') 59 | if len(l_split) == 2: 60 | continue 61 | embs[l_split[0]] = [float(em) for em in l_split[1:]] 62 | print("Got {} decryption embeddings from {}".format(len(embs), 63 | file_enc)) 64 | 65 | return embs 66 | 67 | 68 | def match_embeddings(vocab, emb, opt): 69 | dim = len(six.next(six.itervalues(emb))) 70 | filtered_embeddings = np.zeros((len(vocab), dim)) 71 | count = {"match": 0, "miss": 0} 72 | for w, w_id in vocab.stoi.items(): 73 | if w in emb: 74 | filtered_embeddings[w_id] = emb[w] 75 | count['match'] += 1 76 | else: 77 | if opt.verbose: 78 | print(u"not found:\t{}".format(w), file=sys.stderr) 79 | count['miss'] += 1 80 | 81 | return torch.Tensor(filtered_embeddings), count 82 | 83 | 84 | TYPES = ["GloVe", "word2vec"] 85 | 86 | 87 | def main(): 88 | 89 | parser = argparse.ArgumentParser(description='embeddings_to_torch.py') 90 | parser.add_argument('-emb_file_enc', required=True, 91 | help="source Embeddings from this file") 92 | parser.add_argument('-output_file', required=True, 93 | help="Output file for the prepared data") 94 | parser.add_argument('-dict_file', required=True, 95 | help="Dictionary file") 96 | parser.add_argument('-verbose', action="store_true", default=False) 97 | parser.add_argument('-skip_lines', type=int, default=0, 98 | help="Skip first lines of the embedding file") 99 | parser.add_argument('-type', choices=TYPES, default="GloVe") 100 | opt = parser.parse_args() 101 | 102 | enc_vocab, dec_vocab = get_vocabs(opt.dict_file) 103 | if opt.type == "word2vec": 104 | opt.skip_lines = 1 105 | 106 | embeddings_enc = get_embeddings(opt.emb_file_enc, opt, flag='enc') 107 | 108 | 109 | filtered_enc_embeddings, enc_count = match_embeddings(enc_vocab, 110 | embeddings_enc, 111 | opt) 112 | 113 | print("\nMatching: ") 114 | match_percent = [_['match'] / (_['match'] + _['miss']) * 100 115 | for _ in [enc_count]] 116 | print("\t* enc: %d match, %d missing, (%.2f%%)" % (enc_count['match'], 117 | enc_count['miss'], 118 | match_percent[0])) 119 | 120 | print("\nFiltered embeddings:") 121 | print("\t* enc: ", filtered_enc_embeddings.size()) 122 | 123 | enc_output_file = opt.output_file + ".enc.pt" 124 | 125 | print("\nSaving embedding as:\n\t* enc: %s" 126 | % (enc_output_file)) 127 | torch.save(filtered_enc_embeddings, enc_output_file) 128 | 129 | print("\nDone.") 130 | 131 | 132 | if __name__ == "__main__": 133 | main() 134 | -------------------------------------------------------------------------------- /tools/extract_embeddings.py: -------------------------------------------------------------------------------- 1 | from __future__ import division 2 | import torch 3 | import argparse 4 | import onmt 5 | import onmt.ModelConstructor 6 | import onmt.io 7 | import onmt.opts 8 | from onmt.Utils import use_gpu 9 | 10 | parser = argparse.ArgumentParser(description='translate.py') 11 | 12 | parser.add_argument('-model', required=True, 13 | help='Path to model .pt file') 14 | parser.add_argument('-output_dir', default='.', 15 | help="""Path to output the embeddings""") 16 | parser.add_argument('-gpu', type=int, default=-1, 17 | help="Device to run on") 18 | 19 | 20 | def write_embeddings(filename, dict, embeddings): 21 | with open(filename, 'wb') as file: 22 | for i in range(min(len(embeddings), len(dict.itos))): 23 | str = dict.itos[i].encode("utf-8") 24 | for j in range(len(embeddings[0])): 25 | str = str + (" %5f" % (embeddings[i][j])).encode("utf-8") 26 | file.write(str + b"\n") 27 | 28 | 29 | def main(): 30 | dummy_parser = argparse.ArgumentParser(description='train.py') 31 | onmt.opts.model_opts(dummy_parser) 32 | dummy_opt = dummy_parser.parse_known_args([])[0] 33 | opt = parser.parse_args() 34 | opt.cuda = opt.gpu > -1 35 | if opt.cuda: 36 | torch.cuda.set_device(opt.gpu) 37 | 38 | # Add in default model arguments, possibly added since training. 39 | checkpoint = torch.load(opt.model, 40 | map_location=lambda storage, loc: storage) 41 | model_opt = checkpoint['opt'] 42 | src_dict = checkpoint['vocab'][1][1] 43 | tgt_dict = checkpoint['vocab'][0][1] 44 | 45 | fields = onmt.io.load_fields_from_vocab(checkpoint['vocab']) 46 | 47 | model_opt = checkpoint['opt'] 48 | for arg in dummy_opt.__dict__: 49 | if arg not in model_opt: 50 | model_opt.__dict__[arg] = dummy_opt.__dict__[arg] 51 | 52 | model = onmt.ModelConstructor.make_base_model( 53 | model_opt, fields, use_gpu(opt), checkpoint) 54 | encoder = model.encoder 55 | decoder = model.decoder 56 | 57 | encoder_embeddings = encoder.embeddings.word_lut.weight.data.tolist() 58 | decoder_embeddings = decoder.embeddings.word_lut.weight.data.tolist() 59 | 60 | print("Writing source embeddings") 61 | write_embeddings(opt.output_dir + "/src_embeddings.txt", src_dict, 62 | encoder_embeddings) 63 | 64 | print("Writing target embeddings") 65 | write_embeddings(opt.output_dir + "/tgt_embeddings.txt", tgt_dict, 66 | decoder_embeddings) 67 | 68 | print('... done.') 69 | print('Converting model...') 70 | 71 | 72 | if __name__ == "__main__": 73 | main() 74 | -------------------------------------------------------------------------------- /tools/multi-bleu-detok.perl: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env perl 2 | # 3 | # This file is part of moses. Its use is licensed under the GNU Lesser General 4 | # Public License version 2.1 or, at your option, any later version. 5 | 6 | # This file uses the internal tokenization of mteval-v13a.pl, 7 | # giving the exact same (case-sensitive) results on untokenized text. 8 | # Using this script with detokenized output and untokenized references is 9 | # preferrable over multi-bleu.perl, since scores aren't affected by tokenization differences. 10 | # 11 | # like multi-bleu.perl , it supports plain text input and multiple references. 12 | 13 | # This file is retrieved from Moses Decoder :: https://github.com/moses-smt/mosesdecoder 14 | # $Id$ 15 | use warnings; 16 | use strict; 17 | 18 | my $lowercase = 0; 19 | if ($ARGV[0] eq "-lc") { 20 | $lowercase = 1; 21 | shift; 22 | } 23 | 24 | my $stem = $ARGV[0]; 25 | if (!defined $stem) { 26 | print STDERR "usage: multi-bleu-detok.pl [-lc] reference < hypothesis\n"; 27 | print STDERR "Reads the references from reference or reference0, reference1, ...\n"; 28 | exit(1); 29 | } 30 | 31 | $stem .= ".ref" if !-e $stem && !-e $stem."0" && -e $stem.".ref0"; 32 | 33 | my @REF; 34 | my $ref=0; 35 | while(-e "$stem$ref") { 36 | &add_to_ref("$stem$ref",\@REF); 37 | $ref++; 38 | } 39 | &add_to_ref($stem,\@REF) if -e $stem; 40 | die("ERROR: could not find reference file $stem") unless scalar @REF; 41 | 42 | # add additional references explicitly specified on the command line 43 | shift; 44 | foreach my $stem (@ARGV) { 45 | &add_to_ref($stem,\@REF) if -e $stem; 46 | } 47 | 48 | 49 | 50 | sub add_to_ref { 51 | my ($file,$REF) = @_; 52 | my $s=0; 53 | if ($file =~ /.gz$/) { 54 | open(REF,"gzip -dc $file|") or die "Can't read $file"; 55 | } else { 56 | open(REF,$file) or die "Can't read $file"; 57 | } 58 | while() { 59 | chop; 60 | $_ = tokenization($_); 61 | push @{$$REF[$s++]}, $_; 62 | } 63 | close(REF); 64 | } 65 | 66 | my(@CORRECT,@TOTAL,$length_translation,$length_reference); 67 | my $s=0; 68 | while() { 69 | chop; 70 | $_ = lc if $lowercase; 71 | $_ = tokenization($_); 72 | my @WORD = split; 73 | my %REF_NGRAM = (); 74 | my $length_translation_this_sentence = scalar(@WORD); 75 | my ($closest_diff,$closest_length) = (9999,9999); 76 | foreach my $reference (@{$REF[$s]}) { 77 | # print "$s $_ <=> $reference\n"; 78 | $reference = lc($reference) if $lowercase; 79 | my @WORD = split(' ',$reference); 80 | my $length = scalar(@WORD); 81 | my $diff = abs($length_translation_this_sentence-$length); 82 | if ($diff < $closest_diff) { 83 | $closest_diff = $diff; 84 | $closest_length = $length; 85 | # print STDERR "$s: closest diff ".abs($length_translation_this_sentence-$length)." = abs($length_translation_this_sentence-$length), setting len: $closest_length\n"; 86 | } elsif ($diff == $closest_diff) { 87 | $closest_length = $length if $length < $closest_length; 88 | # from two references with the same closeness to me 89 | # take the *shorter* into account, not the "first" one. 90 | } 91 | for(my $n=1;$n<=4;$n++) { 92 | my %REF_NGRAM_N = (); 93 | for(my $start=0;$start<=$#WORD-($n-1);$start++) { 94 | my $ngram = "$n"; 95 | for(my $w=0;$w<$n;$w++) { 96 | $ngram .= " ".$WORD[$start+$w]; 97 | } 98 | $REF_NGRAM_N{$ngram}++; 99 | } 100 | foreach my $ngram (keys %REF_NGRAM_N) { 101 | if (!defined($REF_NGRAM{$ngram}) || 102 | $REF_NGRAM{$ngram} < $REF_NGRAM_N{$ngram}) { 103 | $REF_NGRAM{$ngram} = $REF_NGRAM_N{$ngram}; 104 | # print "$i: REF_NGRAM{$ngram} = $REF_NGRAM{$ngram}
\n"; 105 | } 106 | } 107 | } 108 | } 109 | $length_translation += $length_translation_this_sentence; 110 | $length_reference += $closest_length; 111 | for(my $n=1;$n<=4;$n++) { 112 | my %T_NGRAM = (); 113 | for(my $start=0;$start<=$#WORD-($n-1);$start++) { 114 | my $ngram = "$n"; 115 | for(my $w=0;$w<$n;$w++) { 116 | $ngram .= " ".$WORD[$start+$w]; 117 | } 118 | $T_NGRAM{$ngram}++; 119 | } 120 | foreach my $ngram (keys %T_NGRAM) { 121 | $ngram =~ /^(\d+) /; 122 | my $n = $1; 123 | # my $corr = 0; 124 | # print "$i e $ngram $T_NGRAM{$ngram}
\n"; 125 | $TOTAL[$n] += $T_NGRAM{$ngram}; 126 | if (defined($REF_NGRAM{$ngram})) { 127 | if ($REF_NGRAM{$ngram} >= $T_NGRAM{$ngram}) { 128 | $CORRECT[$n] += $T_NGRAM{$ngram}; 129 | # $corr = $T_NGRAM{$ngram}; 130 | # print "$i e correct1 $T_NGRAM{$ngram}
\n"; 131 | } 132 | else { 133 | $CORRECT[$n] += $REF_NGRAM{$ngram}; 134 | # $corr = $REF_NGRAM{$ngram}; 135 | # print "$i e correct2 $REF_NGRAM{$ngram}
\n"; 136 | } 137 | } 138 | # $REF_NGRAM{$ngram} = 0 if !defined $REF_NGRAM{$ngram}; 139 | # print STDERR "$ngram: {$s, $REF_NGRAM{$ngram}, $T_NGRAM{$ngram}, $corr}\n" 140 | } 141 | } 142 | $s++; 143 | } 144 | my $brevity_penalty = 1; 145 | my $bleu = 0; 146 | 147 | my @bleu=(); 148 | 149 | for(my $n=1;$n<=4;$n++) { 150 | if (defined ($TOTAL[$n])){ 151 | $bleu[$n]=($TOTAL[$n])?$CORRECT[$n]/$TOTAL[$n]:0; 152 | # print STDERR "CORRECT[$n]:$CORRECT[$n] TOTAL[$n]:$TOTAL[$n]\n"; 153 | }else{ 154 | $bleu[$n]=0; 155 | } 156 | } 157 | 158 | if ($length_reference==0){ 159 | printf "BLEU = 0, 0/0/0/0 (BP=0, ratio=0, hyp_len=0, ref_len=0)\n"; 160 | exit(1); 161 | } 162 | 163 | if ($length_translation<$length_reference) { 164 | $brevity_penalty = exp(1-$length_reference/$length_translation); 165 | } 166 | $bleu = $brevity_penalty * exp((my_log( $bleu[1] ) + 167 | my_log( $bleu[2] ) + 168 | my_log( $bleu[3] ) + 169 | my_log( $bleu[4] ) ) / 4) ; 170 | printf "BLEU = %.2f, %.1f/%.1f/%.1f/%.1f (BP=%.3f, ratio=%.3f, hyp_len=%d, ref_len=%d)\n", 171 | 100*$bleu, 172 | 100*$bleu[1], 173 | 100*$bleu[2], 174 | 100*$bleu[3], 175 | 100*$bleu[4], 176 | $brevity_penalty, 177 | $length_translation / $length_reference, 178 | $length_translation, 179 | $length_reference; 180 | 181 | sub my_log { 182 | return -9999999999 unless $_[0]; 183 | return log($_[0]); 184 | } 185 | 186 | 187 | 188 | sub tokenization 189 | { 190 | my ($norm_text) = @_; 191 | 192 | # language-independent part: 193 | $norm_text =~ s///g; # strip "skipped" tags 194 | $norm_text =~ s/-\n//g; # strip end-of-line hyphenation and join lines 195 | $norm_text =~ s/\n/ /g; # join lines 196 | $norm_text =~ s/"/"/g; # convert SGML tag for quote to " 197 | $norm_text =~ s/&/&/g; # convert SGML tag for ampersand to & 198 | $norm_text =~ s/</ 199 | $norm_text =~ s/>/>/g; # convert SGML tag for greater-than to < 200 | 201 | # language-dependent part (assuming Western languages): 202 | $norm_text = " $norm_text "; 203 | $norm_text =~ s/([\{-\~\[-\` -\&\(-\+\:-\@\/])/ $1 /g; # tokenize punctuation 204 | $norm_text =~ s/([^0-9])([\.,])/$1 $2 /g; # tokenize period and comma unless preceded by a digit 205 | $norm_text =~ s/([\.,])([^0-9])/ $1 $2/g; # tokenize period and comma unless followed by a digit 206 | $norm_text =~ s/([0-9])(-)/$1 $2 /g; # tokenize dash when preceded by a digit 207 | $norm_text =~ s/\s+/ /g; # one space only between words 208 | $norm_text =~ s/^\s+//; # no leading space 209 | $norm_text =~ s/\s+$//; # no trailing space 210 | 211 | return $norm_text; 212 | } 213 | -------------------------------------------------------------------------------- /tools/multi-bleu.perl: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env perl 2 | # 3 | # This file is part of moses. Its use is licensed under the GNU Lesser General 4 | # Public License version 2.1 or, at your option, any later version. 5 | 6 | # $Id$ 7 | use warnings; 8 | use strict; 9 | 10 | my $lowercase = 0; 11 | if ($ARGV[0] eq "-lc") { 12 | $lowercase = 1; 13 | shift; 14 | } 15 | 16 | my $stem = $ARGV[0]; 17 | if (!defined $stem) { 18 | print STDERR "usage: multi-bleu.pl [-lc] reference < hypothesis\n"; 19 | print STDERR "Reads the references from reference or reference0, reference1, ...\n"; 20 | exit(1); 21 | } 22 | 23 | $stem .= ".ref" if !-e $stem && !-e $stem."0" && -e $stem.".ref0"; 24 | 25 | my @REF; 26 | my $ref=0; 27 | while(-e "$stem$ref") { 28 | &add_to_ref("$stem$ref",\@REF); 29 | $ref++; 30 | } 31 | &add_to_ref($stem,\@REF) if -e $stem; 32 | die("ERROR: could not find reference file $stem") unless scalar @REF; 33 | 34 | # add additional references explicitly specified on the command line 35 | shift; 36 | foreach my $stem (@ARGV) { 37 | &add_to_ref($stem,\@REF) if -e $stem; 38 | } 39 | 40 | 41 | 42 | sub add_to_ref { 43 | my ($file,$REF) = @_; 44 | my $s=0; 45 | if ($file =~ /.gz$/) { 46 | open(REF,"gzip -dc $file|") or die "Can't read $file"; 47 | } else { 48 | open(REF,$file) or die "Can't read $file"; 49 | } 50 | while() { 51 | chop; 52 | push @{$$REF[$s++]}, $_; 53 | } 54 | close(REF); 55 | } 56 | 57 | my(@CORRECT,@TOTAL,$length_translation,$length_reference); 58 | my $s=0; 59 | while() { 60 | chop; 61 | $_ = lc if $lowercase; 62 | my @WORD = split; 63 | my %REF_NGRAM = (); 64 | my $length_translation_this_sentence = scalar(@WORD); 65 | my ($closest_diff,$closest_length) = (9999,9999); 66 | foreach my $reference (@{$REF[$s]}) { 67 | # print "$s $_ <=> $reference\n"; 68 | $reference = lc($reference) if $lowercase; 69 | my @WORD = split(' ',$reference); 70 | my $length = scalar(@WORD); 71 | my $diff = abs($length_translation_this_sentence-$length); 72 | if ($diff < $closest_diff) { 73 | $closest_diff = $diff; 74 | $closest_length = $length; 75 | # print STDERR "$s: closest diff ".abs($length_translation_this_sentence-$length)." = abs($length_translation_this_sentence-$length), setting len: $closest_length\n"; 76 | } elsif ($diff == $closest_diff) { 77 | $closest_length = $length if $length < $closest_length; 78 | # from two references with the same closeness to me 79 | # take the *shorter* into account, not the "first" one. 80 | } 81 | for(my $n=1;$n<=4;$n++) { 82 | my %REF_NGRAM_N = (); 83 | for(my $start=0;$start<=$#WORD-($n-1);$start++) { 84 | my $ngram = "$n"; 85 | for(my $w=0;$w<$n;$w++) { 86 | $ngram .= " ".$WORD[$start+$w]; 87 | } 88 | $REF_NGRAM_N{$ngram}++; 89 | } 90 | foreach my $ngram (keys %REF_NGRAM_N) { 91 | if (!defined($REF_NGRAM{$ngram}) || 92 | $REF_NGRAM{$ngram} < $REF_NGRAM_N{$ngram}) { 93 | $REF_NGRAM{$ngram} = $REF_NGRAM_N{$ngram}; 94 | # print "$i: REF_NGRAM{$ngram} = $REF_NGRAM{$ngram}
\n"; 95 | } 96 | } 97 | } 98 | } 99 | $length_translation += $length_translation_this_sentence; 100 | $length_reference += $closest_length; 101 | for(my $n=1;$n<=4;$n++) { 102 | my %T_NGRAM = (); 103 | for(my $start=0;$start<=$#WORD-($n-1);$start++) { 104 | my $ngram = "$n"; 105 | for(my $w=0;$w<$n;$w++) { 106 | $ngram .= " ".$WORD[$start+$w]; 107 | } 108 | $T_NGRAM{$ngram}++; 109 | } 110 | foreach my $ngram (keys %T_NGRAM) { 111 | $ngram =~ /^(\d+) /; 112 | my $n = $1; 113 | # my $corr = 0; 114 | # print "$i e $ngram $T_NGRAM{$ngram}
\n"; 115 | $TOTAL[$n] += $T_NGRAM{$ngram}; 116 | if (defined($REF_NGRAM{$ngram})) { 117 | if ($REF_NGRAM{$ngram} >= $T_NGRAM{$ngram}) { 118 | $CORRECT[$n] += $T_NGRAM{$ngram}; 119 | # $corr = $T_NGRAM{$ngram}; 120 | # print "$i e correct1 $T_NGRAM{$ngram}
\n"; 121 | } 122 | else { 123 | $CORRECT[$n] += $REF_NGRAM{$ngram}; 124 | # $corr = $REF_NGRAM{$ngram}; 125 | # print "$i e correct2 $REF_NGRAM{$ngram}
\n"; 126 | } 127 | } 128 | # $REF_NGRAM{$ngram} = 0 if !defined $REF_NGRAM{$ngram}; 129 | # print STDERR "$ngram: {$s, $REF_NGRAM{$ngram}, $T_NGRAM{$ngram}, $corr}\n" 130 | } 131 | } 132 | $s++; 133 | } 134 | my $brevity_penalty = 1; 135 | my $bleu = 0; 136 | 137 | my @bleu=(); 138 | 139 | for(my $n=1;$n<=4;$n++) { 140 | if (defined ($TOTAL[$n])){ 141 | $bleu[$n]=($TOTAL[$n])?$CORRECT[$n]/$TOTAL[$n]:0; 142 | # print STDERR "CORRECT[$n]:$CORRECT[$n] TOTAL[$n]:$TOTAL[$n]\n"; 143 | }else{ 144 | $bleu[$n]=0; 145 | } 146 | } 147 | 148 | if ($length_reference==0){ 149 | printf "BLEU = 0, 0/0/0/0 (BP=0, ratio=0, hyp_len=0, ref_len=0)\n"; 150 | exit(1); 151 | } 152 | 153 | if ($length_translation<$length_reference) { 154 | $brevity_penalty = exp(1-$length_reference/$length_translation); 155 | } 156 | $bleu = $brevity_penalty * exp((my_log( $bleu[1] ) + 157 | my_log( $bleu[2] ) + 158 | my_log( $bleu[3] ) + 159 | my_log( $bleu[4] ) ) / 4) ; 160 | printf "BLEU = %.2f, %.1f/%.1f/%.1f/%.1f (BP=%.3f, ratio=%.3f, hyp_len=%d, ref_len=%d)\n", 161 | 100*$bleu, 162 | 100*$bleu[1], 163 | 100*$bleu[2], 164 | 100*$bleu[3], 165 | 100*$bleu[4], 166 | $brevity_penalty, 167 | $length_translation / $length_reference, 168 | $length_translation, 169 | $length_reference; 170 | 171 | sub my_log { 172 | return -9999999999 unless $_[0]; 173 | return log($_[0]); 174 | } 175 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/README.txt: -------------------------------------------------------------------------------- 1 | The language suffix can be found here: 2 | 3 | http://www.loc.gov/standards/iso639-2/php/code_list.php 4 | 5 | This code includes data from Daniel Naber's Language Tools (czech abbreviations). 6 | This code includes data from czech wiktionary (also czech abbreviations). 7 | 8 | 9 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.ca: -------------------------------------------------------------------------------- 1 | Dr 2 | Dra 3 | pàg 4 | p 5 | c 6 | av 7 | Sr 8 | Sra 9 | adm 10 | esq 11 | Prof 12 | S.A 13 | S.L 14 | p.e 15 | ptes 16 | Sta 17 | St 18 | pl 19 | màx 20 | cast 21 | dir 22 | nre 23 | fra 24 | admdora 25 | Emm 26 | Excma 27 | espf 28 | dc 29 | admdor 30 | tel 31 | angl 32 | aprox 33 | ca 34 | dept 35 | dj 36 | dl 37 | dt 38 | ds 39 | dg 40 | dv 41 | ed 42 | entl 43 | al 44 | i.e 45 | maj 46 | smin 47 | n 48 | núm 49 | pta 50 | A 51 | B 52 | C 53 | D 54 | E 55 | F 56 | G 57 | H 58 | I 59 | J 60 | K 61 | L 62 | M 63 | N 64 | O 65 | P 66 | Q 67 | R 68 | S 69 | T 70 | U 71 | V 72 | W 73 | X 74 | Y 75 | Z 76 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.cs: -------------------------------------------------------------------------------- 1 | Bc 2 | BcA 3 | Ing 4 | Ing.arch 5 | MUDr 6 | MVDr 7 | MgA 8 | Mgr 9 | JUDr 10 | PhDr 11 | RNDr 12 | PharmDr 13 | ThLic 14 | ThDr 15 | Ph.D 16 | Th.D 17 | prof 18 | doc 19 | CSc 20 | DrSc 21 | dr. h. c 22 | PaedDr 23 | Dr 24 | PhMr 25 | DiS 26 | abt 27 | ad 28 | a.i 29 | aj 30 | angl 31 | anon 32 | apod 33 | atd 34 | atp 35 | aut 36 | bd 37 | biogr 38 | b.m 39 | b.p 40 | b.r 41 | cca 42 | cit 43 | cizojaz 44 | c.k 45 | col 46 | čes 47 | čín 48 | čj 49 | ed 50 | facs 51 | fasc 52 | fol 53 | fot 54 | franc 55 | h.c 56 | hist 57 | hl 58 | hrsg 59 | ibid 60 | il 61 | ind 62 | inv.č 63 | jap 64 | jhdt 65 | jv 66 | koed 67 | kol 68 | korej 69 | kl 70 | krit 71 | lat 72 | lit 73 | m.a 74 | maď 75 | mj 76 | mp 77 | násl 78 | např 79 | nepubl 80 | něm 81 | no 82 | nr 83 | n.s 84 | okr 85 | odd 86 | odp 87 | obr 88 | opr 89 | orig 90 | phil 91 | pl 92 | pokrač 93 | pol 94 | port 95 | pozn 96 | př.kr 97 | př.n.l 98 | přel 99 | přeprac 100 | příl 101 | pseud 102 | pt 103 | red 104 | repr 105 | resp 106 | revid 107 | rkp 108 | roč 109 | roz 110 | rozš 111 | samost 112 | sect 113 | sest 114 | seš 115 | sign 116 | sl 117 | srv 118 | stol 119 | sv 120 | šk 121 | šk.ro 122 | špan 123 | tab 124 | t.č 125 | tis 126 | tj 127 | tř 128 | tzv 129 | univ 130 | uspoř 131 | vol 132 | vl.jm 133 | vs 134 | vyd 135 | vyobr 136 | zal 137 | zejm 138 | zkr 139 | zprac 140 | zvl 141 | n.p 142 | např 143 | než 144 | MUDr 145 | abl 146 | absol 147 | adj 148 | adv 149 | ak 150 | ak. sl 151 | akt 152 | alch 153 | amer 154 | anat 155 | angl 156 | anglosas 157 | arab 158 | arch 159 | archit 160 | arg 161 | astr 162 | astrol 163 | att 164 | bás 165 | belg 166 | bibl 167 | biol 168 | boh 169 | bot 170 | bulh 171 | círk 172 | csl 173 | č 174 | čas 175 | čes 176 | dat 177 | děj 178 | dep 179 | dět 180 | dial 181 | dór 182 | dopr 183 | dosl 184 | ekon 185 | epic 186 | etnonym 187 | eufem 188 | f 189 | fam 190 | fem 191 | fil 192 | film 193 | form 194 | fot 195 | fr 196 | fut 197 | fyz 198 | gen 199 | geogr 200 | geol 201 | geom 202 | germ 203 | gram 204 | hebr 205 | herald 206 | hist 207 | hl 208 | hovor 209 | hud 210 | hut 211 | chcsl 212 | chem 213 | ie 214 | imp 215 | impf 216 | ind 217 | indoevr 218 | inf 219 | instr 220 | interj 221 | ión 222 | iron 223 | it 224 | kanad 225 | katalán 226 | klas 227 | kniž 228 | komp 229 | konj 230 | 231 | konkr 232 | kř 233 | kuch 234 | lat 235 | lék 236 | les 237 | lid 238 | lit 239 | liturg 240 | lok 241 | log 242 | m 243 | mat 244 | meteor 245 | metr 246 | mod 247 | ms 248 | mysl 249 | n 250 | náb 251 | námoř 252 | neklas 253 | něm 254 | nesklon 255 | nom 256 | ob 257 | obch 258 | obyč 259 | ojed 260 | opt 261 | part 262 | pas 263 | pejor 264 | pers 265 | pf 266 | pl 267 | plpf 268 | 269 | práv 270 | prep 271 | předl 272 | přivl 273 | r 274 | rcsl 275 | refl 276 | reg 277 | rkp 278 | ř 279 | řec 280 | s 281 | samohl 282 | sg 283 | sl 284 | souhl 285 | spec 286 | srov 287 | stfr 288 | střv 289 | stsl 290 | subj 291 | subst 292 | superl 293 | sv 294 | sz 295 | táz 296 | tech 297 | telev 298 | teol 299 | trans 300 | typogr 301 | var 302 | vedl 303 | verb 304 | vl. jm 305 | voj 306 | vok 307 | vůb 308 | vulg 309 | výtv 310 | vztaž 311 | zahr 312 | zájm 313 | zast 314 | zejm 315 | 316 | zeměd 317 | zkr 318 | zř 319 | mj 320 | dl 321 | atp 322 | sport 323 | Mgr 324 | horn 325 | MVDr 326 | JUDr 327 | RSDr 328 | Bc 329 | PhDr 330 | ThDr 331 | Ing 332 | aj 333 | apod 334 | PharmDr 335 | pomn 336 | ev 337 | slang 338 | nprap 339 | odp 340 | dop 341 | pol 342 | st 343 | stol 344 | p. n. l 345 | před n. l 346 | n. l 347 | př. Kr 348 | po Kr 349 | př. n. l 350 | odd 351 | RNDr 352 | tzv 353 | atd 354 | tzn 355 | resp 356 | tj 357 | p 358 | br 359 | č. j 360 | čj 361 | č. p 362 | čp 363 | a. s 364 | s. r. o 365 | spol. s r. o 366 | p. o 367 | s. p 368 | v. o. s 369 | k. s 370 | o. p. s 371 | o. s 372 | v. r 373 | v z 374 | ml 375 | vč 376 | kr 377 | mld 378 | hod 379 | popř 380 | ap 381 | event 382 | rus 383 | slov 384 | rum 385 | švýc 386 | P. T 387 | zvl 388 | hor 389 | dol 390 | S.O.S -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.de: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | 4 | #any single upper case letter followed by a period is not a sentence ender (excluding I occasionally, but we leave it in) 5 | #usually upper case letters are initials in a name 6 | #no german words end in single lower-case letters, so we throw those in too. 7 | A 8 | B 9 | C 10 | D 11 | E 12 | F 13 | G 14 | H 15 | I 16 | J 17 | K 18 | L 19 | M 20 | N 21 | O 22 | P 23 | Q 24 | R 25 | S 26 | T 27 | U 28 | V 29 | W 30 | X 31 | Y 32 | Z 33 | a 34 | b 35 | c 36 | d 37 | e 38 | f 39 | g 40 | h 41 | i 42 | j 43 | k 44 | l 45 | m 46 | n 47 | o 48 | p 49 | q 50 | r 51 | s 52 | t 53 | u 54 | v 55 | w 56 | x 57 | y 58 | z 59 | 60 | 61 | #Roman Numerals. A dot after one of these is not a sentence break in German. 62 | I 63 | II 64 | III 65 | IV 66 | V 67 | VI 68 | VII 69 | VIII 70 | IX 71 | X 72 | XI 73 | XII 74 | XIII 75 | XIV 76 | XV 77 | XVI 78 | XVII 79 | XVIII 80 | XIX 81 | XX 82 | i 83 | ii 84 | iii 85 | iv 86 | v 87 | vi 88 | vii 89 | viii 90 | ix 91 | x 92 | xi 93 | xii 94 | xiii 95 | xiv 96 | xv 97 | xvi 98 | xvii 99 | xviii 100 | xix 101 | xx 102 | 103 | #Titles and Honorifics 104 | Adj 105 | Adm 106 | Adv 107 | Asst 108 | Bart 109 | Bldg 110 | Brig 111 | Bros 112 | Capt 113 | Cmdr 114 | Col 115 | Comdr 116 | Con 117 | Corp 118 | Cpl 119 | DR 120 | Dr 121 | Ens 122 | Gen 123 | Gov 124 | Hon 125 | Hosp 126 | Insp 127 | Lt 128 | MM 129 | MR 130 | MRS 131 | MS 132 | Maj 133 | Messrs 134 | Mlle 135 | Mme 136 | Mr 137 | Mrs 138 | Ms 139 | Msgr 140 | Op 141 | Ord 142 | Pfc 143 | Ph 144 | Prof 145 | Pvt 146 | Rep 147 | Reps 148 | Res 149 | Rev 150 | Rt 151 | Sen 152 | Sens 153 | Sfc 154 | Sgt 155 | Sr 156 | St 157 | Supt 158 | Surg 159 | 160 | #Misc symbols 161 | Mio 162 | Mrd 163 | bzw 164 | v 165 | vs 166 | usw 167 | d.h 168 | z.B 169 | u.a 170 | etc 171 | Mrd 172 | MwSt 173 | ggf 174 | d.J 175 | D.h 176 | m.E 177 | vgl 178 | I.F 179 | z.T 180 | sogen 181 | ff 182 | u.E 183 | g.U 184 | g.g.A 185 | c.-à-d 186 | Buchst 187 | u.s.w 188 | sog 189 | u.ä 190 | Std 191 | evtl 192 | Zt 193 | Chr 194 | u.U 195 | o.ä 196 | Ltd 197 | b.A 198 | z.Zt 199 | spp 200 | sen 201 | SA 202 | k.o 203 | jun 204 | i.H.v 205 | dgl 206 | dergl 207 | Co 208 | zzt 209 | usf 210 | s.p.a 211 | Dkr 212 | Corp 213 | bzgl 214 | BSE 215 | 216 | #Number indicators 217 | # add #NUMERIC_ONLY# after the word if it should ONLY be non-breaking when a 0-9 digit follows it 218 | No 219 | Nos 220 | Art 221 | Nr 222 | pp 223 | ca 224 | Ca 225 | 226 | #Ordinals are done with . in German - "1." = "1st" in English 227 | 1 228 | 2 229 | 3 230 | 4 231 | 5 232 | 6 233 | 7 234 | 8 235 | 9 236 | 10 237 | 11 238 | 12 239 | 13 240 | 14 241 | 15 242 | 16 243 | 17 244 | 18 245 | 19 246 | 20 247 | 21 248 | 22 249 | 23 250 | 24 251 | 25 252 | 26 253 | 27 254 | 28 255 | 29 256 | 30 257 | 31 258 | 32 259 | 33 260 | 34 261 | 35 262 | 36 263 | 37 264 | 38 265 | 39 266 | 40 267 | 41 268 | 42 269 | 43 270 | 44 271 | 45 272 | 46 273 | 47 274 | 48 275 | 49 276 | 50 277 | 51 278 | 52 279 | 53 280 | 54 281 | 55 282 | 56 283 | 57 284 | 58 285 | 59 286 | 60 287 | 61 288 | 62 289 | 63 290 | 64 291 | 65 292 | 66 293 | 67 294 | 68 295 | 69 296 | 70 297 | 71 298 | 72 299 | 73 300 | 74 301 | 75 302 | 76 303 | 77 304 | 78 305 | 79 306 | 80 307 | 81 308 | 82 309 | 83 310 | 84 311 | 85 312 | 86 313 | 87 314 | 88 315 | 89 316 | 90 317 | 91 318 | 92 319 | 93 320 | 94 321 | 95 322 | 96 323 | 97 324 | 98 325 | 99 326 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.en: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | 4 | #any single upper case letter followed by a period is not a sentence ender (excluding I occasionally, but we leave it in) 5 | #usually upper case letters are initials in a name 6 | A 7 | B 8 | C 9 | D 10 | E 11 | F 12 | G 13 | H 14 | I 15 | J 16 | K 17 | L 18 | M 19 | N 20 | O 21 | P 22 | Q 23 | R 24 | S 25 | T 26 | U 27 | V 28 | W 29 | X 30 | Y 31 | Z 32 | 33 | #List of titles. These are often followed by upper-case names, but do not indicate sentence breaks 34 | Adj 35 | Adm 36 | Adv 37 | Asst 38 | Bart 39 | Bldg 40 | Brig 41 | Bros 42 | Capt 43 | Cmdr 44 | Col 45 | Comdr 46 | Con 47 | Corp 48 | Cpl 49 | DR 50 | Dr 51 | Drs 52 | Ens 53 | Gen 54 | Gov 55 | Hon 56 | Hr 57 | Hosp 58 | Insp 59 | Lt 60 | MM 61 | MR 62 | MRS 63 | MS 64 | Maj 65 | Messrs 66 | Mlle 67 | Mme 68 | Mr 69 | Mrs 70 | Ms 71 | Msgr 72 | Op 73 | Ord 74 | Pfc 75 | Ph 76 | Prof 77 | Pvt 78 | Rep 79 | Reps 80 | Res 81 | Rev 82 | Rt 83 | Sen 84 | Sens 85 | Sfc 86 | Sgt 87 | Sr 88 | St 89 | Supt 90 | Surg 91 | 92 | #misc - odd period-ending items that NEVER indicate breaks (p.m. does NOT fall into this category - it sometimes ends a sentence) 93 | v 94 | vs 95 | i.e 96 | rev 97 | e.g 98 | 99 | #Numbers only. These should only induce breaks when followed by a numeric sequence 100 | # add NUMERIC_ONLY after the word for this function 101 | #This case is mostly for the english "No." which can either be a sentence of its own, or 102 | #if followed by a number, a non-breaking prefix 103 | No #NUMERIC_ONLY# 104 | Nos 105 | Art #NUMERIC_ONLY# 106 | Nr 107 | pp #NUMERIC_ONLY# 108 | 109 | #month abbreviations 110 | Jan 111 | Feb 112 | Mar 113 | Apr 114 | #May is a full word 115 | Jun 116 | Jul 117 | Aug 118 | Sep 119 | Oct 120 | Nov 121 | Dec 122 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.es: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | 4 | #any single upper case letter followed by a period is not a sentence ender 5 | #usually upper case letters are initials in a name 6 | A 7 | B 8 | C 9 | D 10 | E 11 | F 12 | G 13 | H 14 | I 15 | J 16 | K 17 | L 18 | M 19 | N 20 | O 21 | P 22 | Q 23 | R 24 | S 25 | T 26 | U 27 | V 28 | W 29 | X 30 | Y 31 | Z 32 | 33 | # Period-final abbreviation list from http://www.ctspanish.com/words/abbreviations.htm 34 | 35 | A.C 36 | Apdo 37 | Av 38 | Bco 39 | CC.AA 40 | Da 41 | Dep 42 | Dn 43 | Dr 44 | Dra 45 | EE.UU 46 | Excmo 47 | FF.CC 48 | Fil 49 | Gral 50 | J.C 51 | Let 52 | Lic 53 | N.B 54 | P.D 55 | P.V.P 56 | Prof 57 | Pts 58 | Rte 59 | S.A 60 | S.A.R 61 | S.E 62 | S.L 63 | S.R.C 64 | Sr 65 | Sra 66 | Srta 67 | Sta 68 | Sto 69 | T.V.E 70 | Tel 71 | Ud 72 | Uds 73 | V.B 74 | V.E 75 | Vd 76 | Vds 77 | a/c 78 | adj 79 | admón 80 | afmo 81 | apdo 82 | av 83 | c 84 | c.f 85 | c.g 86 | cap 87 | cm 88 | cta 89 | dcha 90 | doc 91 | ej 92 | entlo 93 | esq 94 | etc 95 | f.c 96 | gr 97 | grs 98 | izq 99 | kg 100 | km 101 | mg 102 | mm 103 | núm 104 | núm 105 | p 106 | p.a 107 | p.ej 108 | ptas 109 | pág 110 | págs 111 | pág 112 | págs 113 | q.e.g.e 114 | q.e.s.m 115 | s 116 | s.s.s 117 | vid 118 | vol 119 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.fi: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT 2 | #indicate an end-of-sentence marker. Special cases are included for prefixes 3 | #that ONLY appear before 0-9 numbers. 4 | 5 | #This list is compiled from omorfi database 6 | #by Tommi A Pirinen. 7 | 8 | 9 | #any single upper case letter followed by a period is not a sentence ender 10 | A 11 | B 12 | C 13 | D 14 | E 15 | F 16 | G 17 | H 18 | I 19 | J 20 | K 21 | L 22 | M 23 | N 24 | O 25 | P 26 | Q 27 | R 28 | S 29 | T 30 | U 31 | V 32 | W 33 | X 34 | Y 35 | Z 36 | Å 37 | Ä 38 | Ö 39 | 40 | #List of titles. These are often followed by upper-case names, but do not indicate sentence breaks 41 | alik 42 | alil 43 | amir 44 | apul 45 | apul.prof 46 | arkkit 47 | ass 48 | assist 49 | dipl 50 | dipl.arkkit 51 | dipl.ekon 52 | dipl.ins 53 | dipl.kielenk 54 | dipl.kirjeenv 55 | dipl.kosm 56 | dipl.urk 57 | dos 58 | erikoiseläinl 59 | erikoishammasl 60 | erikoisl 61 | erikoist 62 | ev.luutn 63 | evp 64 | fil 65 | ft 66 | hallinton 67 | hallintot 68 | hammaslääket 69 | jatk 70 | jääk 71 | kansaned 72 | kapt 73 | kapt.luutn 74 | kenr 75 | kenr.luutn 76 | kenr.maj 77 | kers 78 | kirjeenv 79 | kom 80 | kom.kapt 81 | komm 82 | konst 83 | korpr 84 | luutn 85 | maist 86 | maj 87 | Mr 88 | Mrs 89 | Ms 90 | M.Sc 91 | neuv 92 | nimim 93 | Ph.D 94 | prof 95 | puh.joht 96 | pääll 97 | res 98 | san 99 | siht 100 | suom 101 | sähköp 102 | säv 103 | toht 104 | toim 105 | toim.apul 106 | toim.joht 107 | toim.siht 108 | tuom 109 | ups 110 | vänr 111 | vääp 112 | ye.ups 113 | ylik 114 | ylil 115 | ylim 116 | ylimatr 117 | yliop 118 | yliopp 119 | ylip 120 | yliv 121 | 122 | #misc - odd period-ending items that NEVER indicate breaks (p.m. does NOT fall 123 | #into this category - it sometimes ends a sentence) 124 | e.g 125 | ent 126 | esim 127 | huom 128 | i.e 129 | ilm 130 | l 131 | mm 132 | myöh 133 | nk 134 | nyk 135 | par 136 | po 137 | t 138 | v 139 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.fr: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | # 4 | #any single upper case letter followed by a period is not a sentence ender 5 | #usually upper case letters are initials in a name 6 | #no French words end in single lower-case letters, so we throw those in too? 7 | A 8 | B 9 | C 10 | D 11 | E 12 | F 13 | G 14 | H 15 | I 16 | J 17 | K 18 | L 19 | M 20 | N 21 | O 22 | P 23 | Q 24 | R 25 | S 26 | T 27 | U 28 | V 29 | W 30 | X 31 | Y 32 | Z 33 | #a 34 | b 35 | c 36 | d 37 | e 38 | f 39 | g 40 | h 41 | i 42 | j 43 | k 44 | l 45 | m 46 | n 47 | o 48 | p 49 | q 50 | r 51 | s 52 | t 53 | u 54 | v 55 | w 56 | x 57 | y 58 | z 59 | 60 | # Period-final abbreviation list for French 61 | A.C.N 62 | A.M 63 | art 64 | ann 65 | apr 66 | av 67 | auj 68 | lib 69 | B.P 70 | boul 71 | ca 72 | c.-à-d 73 | cf 74 | ch.-l 75 | chap 76 | contr 77 | C.P.I 78 | C.Q.F.D 79 | C.N 80 | C.N.S 81 | C.S 82 | dir 83 | éd 84 | e.g 85 | env 86 | al 87 | etc 88 | E.V 89 | ex 90 | fasc 91 | fém 92 | fig 93 | fr 94 | hab 95 | ibid 96 | id 97 | i.e 98 | inf 99 | LL.AA 100 | LL.AA.II 101 | LL.AA.RR 102 | LL.AA.SS 103 | L.D 104 | LL.EE 105 | LL.MM 106 | LL.MM.II.RR 107 | loc.cit 108 | masc 109 | MM 110 | ms 111 | N.B 112 | N.D.A 113 | N.D.L.R 114 | N.D.T 115 | n/réf 116 | NN.SS 117 | N.S 118 | N.D 119 | N.P.A.I 120 | p.c.c 121 | pl 122 | pp 123 | p.ex 124 | p.j 125 | P.S 126 | R.A.S 127 | R.-V 128 | R.P 129 | R.I.P 130 | SS 131 | S.S 132 | S.A 133 | S.A.I 134 | S.A.R 135 | S.A.S 136 | S.E 137 | sec 138 | sect 139 | sing 140 | S.M 141 | S.M.I.R 142 | sq 143 | sqq 144 | suiv 145 | sup 146 | suppl 147 | tél 148 | T.S.V.P 149 | vb 150 | vol 151 | vs 152 | X.O 153 | Z.I 154 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.ga: -------------------------------------------------------------------------------- 1 | 2 | A 3 | B 4 | C 5 | D 6 | E 7 | F 8 | G 9 | H 10 | I 11 | J 12 | K 13 | L 14 | M 15 | N 16 | O 17 | P 18 | Q 19 | R 20 | S 21 | T 22 | U 23 | V 24 | W 25 | X 26 | Y 27 | Z 28 | Á 29 | É 30 | Í 31 | Ó 32 | Ú 33 | 34 | Uacht 35 | Dr 36 | B.Arch 37 | 38 | m.sh 39 | .i 40 | Co 41 | Cf 42 | cf 43 | i.e 44 | r 45 | Chr 46 | lch #NUMERIC_ONLY# 47 | lgh #NUMERIC_ONLY# 48 | uimh #NUMERIC_ONLY# 49 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.hu: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | 4 | #any single upper case letter followed by a period is not a sentence ender (excluding I occasionally, but we leave it in) 5 | #usually upper case letters are initials in a name 6 | A 7 | B 8 | C 9 | D 10 | E 11 | F 12 | G 13 | H 14 | I 15 | J 16 | K 17 | L 18 | M 19 | N 20 | O 21 | P 22 | Q 23 | R 24 | S 25 | T 26 | U 27 | V 28 | W 29 | X 30 | Y 31 | Z 32 | Á 33 | É 34 | Í 35 | Ó 36 | Ö 37 | Ő 38 | Ú 39 | Ü 40 | Ű 41 | 42 | #List of titles. These are often followed by upper-case names, but do not indicate sentence breaks 43 | Dr 44 | dr 45 | kb 46 | Kb 47 | vö 48 | Vö 49 | pl 50 | Pl 51 | ca 52 | Ca 53 | min 54 | Min 55 | max 56 | Max 57 | ún 58 | Ún 59 | prof 60 | Prof 61 | de 62 | De 63 | du 64 | Du 65 | Szt 66 | St 67 | 68 | #Numbers only. These should only induce breaks when followed by a numeric sequence 69 | # add NUMERIC_ONLY after the word for this function 70 | #This case is mostly for the english "No." which can either be a sentence of its own, or 71 | #if followed by a number, a non-breaking prefix 72 | 73 | # Month name abbreviations 74 | jan #NUMERIC_ONLY# 75 | Jan #NUMERIC_ONLY# 76 | Feb #NUMERIC_ONLY# 77 | feb #NUMERIC_ONLY# 78 | márc #NUMERIC_ONLY# 79 | Márc #NUMERIC_ONLY# 80 | ápr #NUMERIC_ONLY# 81 | Ápr #NUMERIC_ONLY# 82 | máj #NUMERIC_ONLY# 83 | Máj #NUMERIC_ONLY# 84 | jún #NUMERIC_ONLY# 85 | Jún #NUMERIC_ONLY# 86 | Júl #NUMERIC_ONLY# 87 | júl #NUMERIC_ONLY# 88 | aug #NUMERIC_ONLY# 89 | Aug #NUMERIC_ONLY# 90 | Szept #NUMERIC_ONLY# 91 | szept #NUMERIC_ONLY# 92 | okt #NUMERIC_ONLY# 93 | Okt #NUMERIC_ONLY# 94 | nov #NUMERIC_ONLY# 95 | Nov #NUMERIC_ONLY# 96 | dec #NUMERIC_ONLY# 97 | Dec #NUMERIC_ONLY# 98 | 99 | # Other abbreviations 100 | tel #NUMERIC_ONLY# 101 | Tel #NUMERIC_ONLY# 102 | Fax #NUMERIC_ONLY# 103 | fax #NUMERIC_ONLY# 104 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.is: -------------------------------------------------------------------------------- 1 | no #NUMERIC_ONLY# 2 | No #NUMERIC_ONLY# 3 | nr #NUMERIC_ONLY# 4 | Nr #NUMERIC_ONLY# 5 | nR #NUMERIC_ONLY# 6 | NR #NUMERIC_ONLY# 7 | a 8 | b 9 | c 10 | d 11 | e 12 | f 13 | g 14 | h 15 | i 16 | j 17 | k 18 | l 19 | m 20 | n 21 | o 22 | p 23 | q 24 | r 25 | s 26 | t 27 | u 28 | v 29 | w 30 | x 31 | y 32 | z 33 | ^ 34 | í 35 | á 36 | ó 37 | æ 38 | A 39 | B 40 | C 41 | D 42 | E 43 | F 44 | G 45 | H 46 | I 47 | J 48 | K 49 | L 50 | M 51 | N 52 | O 53 | P 54 | Q 55 | R 56 | S 57 | T 58 | U 59 | V 60 | W 61 | X 62 | Y 63 | Z 64 | ab.fn 65 | a.fn 66 | afs 67 | al 68 | alm 69 | alg 70 | andh 71 | ath 72 | aths 73 | atr 74 | ao 75 | au 76 | aukaf 77 | áfn 78 | áhrl.s 79 | áhrs 80 | ákv.gr 81 | ákv 82 | bh 83 | bls 84 | dr 85 | e.Kr 86 | et 87 | ef 88 | efn 89 | ennfr 90 | eink 91 | end 92 | e.st 93 | erl 94 | fél 95 | fskj 96 | fh 97 | f.hl 98 | físl 99 | fl 100 | fn 101 | fo 102 | forl 103 | frb 104 | frl 105 | frh 106 | frt 107 | fsl 108 | fsh 109 | fs 110 | fsk 111 | fst 112 | f.Kr 113 | ft 114 | fv 115 | fyrrn 116 | fyrrv 117 | germ 118 | gm 119 | gr 120 | hdl 121 | hdr 122 | hf 123 | hl 124 | hlsk 125 | hljsk 126 | hljv 127 | hljóðv 128 | hr 129 | hv 130 | hvk 131 | holl 132 | Hos 133 | höf 134 | hk 135 | hrl 136 | ísl 137 | kaf 138 | kap 139 | Khöfn 140 | kk 141 | kg 142 | kk 143 | km 144 | kl 145 | klst 146 | kr 147 | kt 148 | kgúrsk 149 | kvk 150 | leturbr 151 | lh 152 | lh.nt 153 | lh.þt 154 | lo 155 | ltr 156 | mlja 157 | mljó 158 | millj 159 | mm 160 | mms 161 | m.fl 162 | miðm 163 | mgr 164 | mst 165 | mín 166 | nf 167 | nh 168 | nhm 169 | nl 170 | nk 171 | nmgr 172 | no 173 | núv 174 | nt 175 | o.áfr 176 | o.m.fl 177 | ohf 178 | o.fl 179 | o.s.frv 180 | ófn 181 | ób 182 | óákv.gr 183 | óákv 184 | pfn 185 | PR 186 | pr 187 | Ritstj 188 | Rvík 189 | Rvk 190 | samb 191 | samhlj 192 | samn 193 | samn 194 | sbr 195 | sek 196 | sérn 197 | sf 198 | sfn 199 | sh 200 | sfn 201 | sh 202 | s.hl 203 | sk 204 | skv 205 | sl 206 | sn 207 | so 208 | ss.us 209 | s.st 210 | samþ 211 | sbr 212 | shlj 213 | sign 214 | skál 215 | st 216 | st.s 217 | stk 218 | sþ 219 | teg 220 | tbl 221 | tfn 222 | tl 223 | tvíhlj 224 | tvt 225 | till 226 | to 227 | umr 228 | uh 229 | us 230 | uppl 231 | útg 232 | vb 233 | Vf 234 | vh 235 | vkf 236 | Vl 237 | vl 238 | vlf 239 | vmf 240 | 8vo 241 | vsk 242 | vth 243 | þt 244 | þf 245 | þjs 246 | þgf 247 | þlt 248 | þolm 249 | þm 250 | þml 251 | þýð 252 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.it: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | 4 | #any single upper case letter followed by a period is not a sentence ender (excluding I occasionally, but we leave it in) 5 | #usually upper case letters are initials in a name 6 | A 7 | B 8 | C 9 | D 10 | E 11 | F 12 | G 13 | H 14 | I 15 | J 16 | K 17 | L 18 | M 19 | N 20 | O 21 | P 22 | Q 23 | R 24 | S 25 | T 26 | U 27 | V 28 | W 29 | X 30 | Y 31 | Z 32 | 33 | #List of titles. These are often followed by upper-case names, but do not indicate sentence breaks 34 | Adj 35 | Adm 36 | Adv 37 | Amn 38 | Arch 39 | Asst 40 | Avv 41 | Bart 42 | Bcc 43 | Bldg 44 | Brig 45 | Bros 46 | C.A.P 47 | C.P 48 | Capt 49 | Cc 50 | Cmdr 51 | Co 52 | Col 53 | Comdr 54 | Con 55 | Corp 56 | Cpl 57 | DR 58 | Dott 59 | Dr 60 | Drs 61 | Egr 62 | Ens 63 | Gen 64 | Geom 65 | Gov 66 | Hon 67 | Hosp 68 | Hr 69 | Id 70 | Ing 71 | Insp 72 | Lt 73 | MM 74 | MR 75 | MRS 76 | MS 77 | Maj 78 | Messrs 79 | Mlle 80 | Mme 81 | Mo 82 | Mons 83 | Mr 84 | Mrs 85 | Ms 86 | Msgr 87 | N.B 88 | Op 89 | Ord 90 | P.S 91 | P.T 92 | Pfc 93 | Ph 94 | Prof 95 | Pvt 96 | RP 97 | RSVP 98 | Rag 99 | Rep 100 | Reps 101 | Res 102 | Rev 103 | Rif 104 | Rt 105 | S.A 106 | S.B.F 107 | S.P.M 108 | S.p.A 109 | S.r.l 110 | Sen 111 | Sens 112 | Sfc 113 | Sgt 114 | Sig 115 | Sigg 116 | Soc 117 | Spett 118 | Sr 119 | St 120 | Supt 121 | Surg 122 | V.P 123 | 124 | # other 125 | a.c 126 | acc 127 | all 128 | banc 129 | c.a 130 | c.c.p 131 | c.m 132 | c.p 133 | c.s 134 | c.v 135 | corr 136 | dott 137 | e.p.c 138 | ecc 139 | es 140 | fatt 141 | gg 142 | int 143 | lett 144 | ogg 145 | on 146 | p.c 147 | p.c.c 148 | p.es 149 | p.f 150 | p.r 151 | p.v 152 | post 153 | pp 154 | racc 155 | ric 156 | s.n.c 157 | seg 158 | sgg 159 | ss 160 | tel 161 | u.s 162 | v.r 163 | v.s 164 | 165 | #misc - odd period-ending items that NEVER indicate breaks (p.m. does NOT fall into this category - it sometimes ends a sentence) 166 | v 167 | vs 168 | i.e 169 | rev 170 | e.g 171 | 172 | #Numbers only. These should only induce breaks when followed by a numeric sequence 173 | # add NUMERIC_ONLY after the word for this function 174 | #This case is mostly for the english "No." which can either be a sentence of its own, or 175 | #if followed by a number, a non-breaking prefix 176 | No #NUMERIC_ONLY# 177 | Nos 178 | Art #NUMERIC_ONLY# 179 | Nr 180 | pp #NUMERIC_ONLY# 181 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.lv: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | 4 | #any single upper case letter followed by a period is not a sentence ender (excluding I occasionally, but we leave it in) 5 | #usually upper case letters are initials in a name 6 | A 7 | Ā 8 | B 9 | C 10 | Č 11 | D 12 | E 13 | Ē 14 | F 15 | G 16 | Ģ 17 | H 18 | I 19 | Ī 20 | J 21 | K 22 | Ķ 23 | L 24 | Ļ 25 | M 26 | N 27 | Ņ 28 | O 29 | P 30 | Q 31 | R 32 | S 33 | Š 34 | T 35 | U 36 | Ū 37 | V 38 | W 39 | X 40 | Y 41 | Z 42 | Ž 43 | 44 | #List of titles. These are often followed by upper-case names, but do not indicate sentence breaks 45 | dr 46 | Dr 47 | med 48 | prof 49 | Prof 50 | inž 51 | Inž 52 | ist.loc 53 | Ist.loc 54 | kor.loc 55 | Kor.loc 56 | v.i 57 | vietn 58 | Vietn 59 | 60 | #misc - odd period-ending items that NEVER indicate breaks (p.m. does NOT fall into this category - it sometimes ends a sentence) 61 | a.l 62 | t.p 63 | pārb 64 | Pārb 65 | vec 66 | Vec 67 | inv 68 | Inv 69 | sk 70 | Sk 71 | spec 72 | Spec 73 | vienk 74 | Vienk 75 | virz 76 | Virz 77 | māksl 78 | Māksl 79 | mūz 80 | Mūz 81 | akad 82 | Akad 83 | soc 84 | Soc 85 | galv 86 | Galv 87 | vad 88 | Vad 89 | sertif 90 | Sertif 91 | folkl 92 | Folkl 93 | hum 94 | Hum 95 | 96 | #Numbers only. These should only induce breaks when followed by a numeric sequence 97 | # add NUMERIC_ONLY after the word for this function 98 | #This case is mostly for the english "No." which can either be a sentence of its own, or 99 | #if followed by a number, a non-breaking prefix 100 | Nr #NUMERIC_ONLY# 101 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.nl: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | #Sources: http://nl.wikipedia.org/wiki/Lijst_van_afkortingen 4 | # http://nl.wikipedia.org/wiki/Aanspreekvorm 5 | # http://nl.wikipedia.org/wiki/Titulatuur_in_het_Nederlands_hoger_onderwijs 6 | #any single upper case letter followed by a period is not a sentence ender (excluding I occasionally, but we leave it in) 7 | #usually upper case letters are initials in a name 8 | A 9 | B 10 | C 11 | D 12 | E 13 | F 14 | G 15 | H 16 | I 17 | J 18 | K 19 | L 20 | M 21 | N 22 | O 23 | P 24 | Q 25 | R 26 | S 27 | T 28 | U 29 | V 30 | W 31 | X 32 | Y 33 | Z 34 | 35 | #List of titles. These are often followed by upper-case names, but do not indicate sentence breaks 36 | bacc 37 | bc 38 | bgen 39 | c.i 40 | dhr 41 | dr 42 | dr.h.c 43 | drs 44 | drs 45 | ds 46 | eint 47 | fa 48 | Fa 49 | fam 50 | gen 51 | genm 52 | ing 53 | ir 54 | jhr 55 | jkvr 56 | jr 57 | kand 58 | kol 59 | lgen 60 | lkol 61 | Lt 62 | maj 63 | Mej 64 | mevr 65 | Mme 66 | mr 67 | mr 68 | Mw 69 | o.b.s 70 | plv 71 | prof 72 | ritm 73 | tint 74 | Vz 75 | Z.D 76 | Z.D.H 77 | Z.E 78 | Z.Em 79 | Z.H 80 | Z.K.H 81 | Z.K.M 82 | Z.M 83 | z.v 84 | 85 | #misc - odd period-ending items that NEVER indicate breaks (p.m. does NOT fall into this category - it sometimes ends a sentence) 86 | #we seem to have a lot of these in dutch i.e.: i.p.v - in plaats van (in stead of) never ends a sentence 87 | a.g.v 88 | bijv 89 | bijz 90 | bv 91 | d.w.z 92 | e.c 93 | e.g 94 | e.k 95 | ev 96 | i.p.v 97 | i.s.m 98 | i.t.t 99 | i.v.m 100 | m.a.w 101 | m.b.t 102 | m.b.v 103 | m.h.o 104 | m.i 105 | m.i.v 106 | v.w.t 107 | 108 | #Numbers only. These should only induce breaks when followed by a numeric sequence 109 | # add NUMERIC_ONLY after the word for this function 110 | #This case is mostly for the english "No." which can either be a sentence of its own, or 111 | #if followed by a number, a non-breaking prefix 112 | Nr #NUMERIC_ONLY# 113 | Nrs 114 | nrs 115 | nr #NUMERIC_ONLY# 116 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.pl: -------------------------------------------------------------------------------- 1 | adw 2 | afr 3 | akad 4 | al 5 | Al 6 | am 7 | amer 8 | arch 9 | art 10 | Art 11 | artyst 12 | astr 13 | austr 14 | bałt 15 | bdb 16 | bł 17 | bm 18 | br 19 | bryg 20 | bryt 21 | centr 22 | ces 23 | chem 24 | chiń 25 | chir 26 | c.k 27 | c.o 28 | cyg 29 | cyw 30 | cyt 31 | czes 32 | czw 33 | cd 34 | Cd 35 | czyt 36 | ćw 37 | ćwicz 38 | daw 39 | dcn 40 | dekl 41 | demokr 42 | det 43 | diec 44 | dł 45 | dn 46 | dot 47 | dol 48 | dop 49 | dost 50 | dosł 51 | h.c 52 | ds 53 | dst 54 | duszp 55 | dypl 56 | egz 57 | ekol 58 | ekon 59 | elektr 60 | em 61 | ew 62 | fab 63 | farm 64 | fot 65 | fr 66 | gat 67 | gastr 68 | geogr 69 | geol 70 | gimn 71 | głęb 72 | gm 73 | godz 74 | górn 75 | gosp 76 | gr 77 | gram 78 | hist 79 | hiszp 80 | hr 81 | Hr 82 | hot 83 | id 84 | in 85 | im 86 | iron 87 | jn 88 | kard 89 | kat 90 | katol 91 | k.k 92 | kk 93 | kol 94 | kl 95 | k.p.a 96 | kpc 97 | k.p.c 98 | kpt 99 | kr 100 | k.r 101 | krak 102 | k.r.o 103 | kryt 104 | kult 105 | laic 106 | łac 107 | niem 108 | woj 109 | nb 110 | np 111 | Nb 112 | Np 113 | pol 114 | pow 115 | m.in 116 | pt 117 | ps 118 | Pt 119 | Ps 120 | cdn 121 | jw 122 | ryc 123 | rys 124 | Ryc 125 | Rys 126 | tj 127 | tzw 128 | Tzw 129 | tzn 130 | zob 131 | ang 132 | ub 133 | ul 134 | pw 135 | pn 136 | pl 137 | al 138 | k 139 | n 140 | nr #NUMERIC_ONLY# 141 | Nr #NUMERIC_ONLY# 142 | ww 143 | wł 144 | ur 145 | zm 146 | żyd 147 | żarg 148 | żyw 149 | wył 150 | bp 151 | bp 152 | wyst 153 | tow 154 | Tow 155 | o 156 | sp 157 | Sp 158 | st 159 | spółdz 160 | Spółdz 161 | społ 162 | spółgł 163 | stoł 164 | stow 165 | Stoł 166 | Stow 167 | zn 168 | zew 169 | zewn 170 | zdr 171 | zazw 172 | zast 173 | zaw 174 | zał 175 | zal 176 | zam 177 | zak 178 | zakł 179 | zagr 180 | zach 181 | adw 182 | Adw 183 | lek 184 | Lek 185 | med 186 | mec 187 | Mec 188 | doc 189 | Doc 190 | dyw 191 | dyr 192 | Dyw 193 | Dyr 194 | inż 195 | Inż 196 | mgr 197 | Mgr 198 | dh 199 | dr 200 | Dh 201 | Dr 202 | p 203 | P 204 | red 205 | Red 206 | prof 207 | prok 208 | Prof 209 | Prok 210 | hab 211 | płk 212 | Płk 213 | nadkom 214 | Nadkom 215 | podkom 216 | Podkom 217 | ks 218 | Ks 219 | gen 220 | Gen 221 | por 222 | Por 223 | reż 224 | Reż 225 | przyp 226 | Przyp 227 | śp 228 | św 229 | śW 230 | Śp 231 | Św 232 | ŚW 233 | szer 234 | Szer 235 | pkt #NUMERIC_ONLY# 236 | str #NUMERIC_ONLY# 237 | tab #NUMERIC_ONLY# 238 | Tab #NUMERIC_ONLY# 239 | tel 240 | ust #NUMERIC_ONLY# 241 | par #NUMERIC_ONLY# 242 | poz 243 | pok 244 | oo 245 | oO 246 | Oo 247 | OO 248 | r #NUMERIC_ONLY# 249 | l #NUMERIC_ONLY# 250 | s #NUMERIC_ONLY# 251 | najśw 252 | Najśw 253 | A 254 | B 255 | C 256 | D 257 | E 258 | F 259 | G 260 | H 261 | I 262 | J 263 | K 264 | L 265 | M 266 | N 267 | O 268 | P 269 | Q 270 | R 271 | S 272 | T 273 | U 274 | V 275 | W 276 | X 277 | Y 278 | Z 279 | Ś 280 | Ć 281 | Ż 282 | Ź 283 | Dz 284 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.ro: -------------------------------------------------------------------------------- 1 | A 2 | B 3 | C 4 | D 5 | E 6 | F 7 | G 8 | H 9 | I 10 | J 11 | K 12 | L 13 | M 14 | N 15 | O 16 | P 17 | Q 18 | R 19 | S 20 | T 21 | U 22 | V 23 | W 24 | X 25 | Y 26 | Z 27 | dpdv 28 | etc 29 | șamd 30 | M.Ap.N 31 | dl 32 | Dl 33 | d-na 34 | D-na 35 | dvs 36 | Dvs 37 | pt 38 | Pt 39 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.ru: -------------------------------------------------------------------------------- 1 | # added Cyrillic uppercase letters [А-Я] 2 | # removed 000D carriage return (this is not removed by chomp in tokenizer.perl, and prevents recognition of the prefixes) 3 | # edited by Kate Young (nspaceanalysis@earthlink.net) 21 May 2013 4 | А 5 | Б 6 | В 7 | Г 8 | Д 9 | Е 10 | Ж 11 | З 12 | И 13 | Й 14 | К 15 | Л 16 | М 17 | Н 18 | О 19 | П 20 | Р 21 | С 22 | Т 23 | У 24 | Ф 25 | Х 26 | Ц 27 | Ч 28 | Ш 29 | Щ 30 | Ъ 31 | Ы 32 | Ь 33 | Э 34 | Ю 35 | Я 36 | A 37 | B 38 | C 39 | D 40 | E 41 | F 42 | G 43 | H 44 | I 45 | J 46 | K 47 | L 48 | M 49 | N 50 | O 51 | P 52 | Q 53 | R 54 | S 55 | T 56 | U 57 | V 58 | W 59 | X 60 | Y 61 | Z 62 | 0гг 63 | 1гг 64 | 2гг 65 | 3гг 66 | 4гг 67 | 5гг 68 | 6гг 69 | 7гг 70 | 8гг 71 | 9гг 72 | 0г 73 | 1г 74 | 2г 75 | 3г 76 | 4г 77 | 5г 78 | 6г 79 | 7г 80 | 8г 81 | 9г 82 | Xвв 83 | Vвв 84 | Iвв 85 | Lвв 86 | Mвв 87 | Cвв 88 | Xв 89 | Vв 90 | Iв 91 | Lв 92 | Mв 93 | Cв 94 | 0м 95 | 1м 96 | 2м 97 | 3м 98 | 4м 99 | 5м 100 | 6м 101 | 7м 102 | 8м 103 | 9м 104 | 0мм 105 | 1мм 106 | 2мм 107 | 3мм 108 | 4мм 109 | 5мм 110 | 6мм 111 | 7мм 112 | 8мм 113 | 9мм 114 | 0см 115 | 1см 116 | 2см 117 | 3см 118 | 4см 119 | 5см 120 | 6см 121 | 7см 122 | 8см 123 | 9см 124 | 0дм 125 | 1дм 126 | 2дм 127 | 3дм 128 | 4дм 129 | 5дм 130 | 6дм 131 | 7дм 132 | 8дм 133 | 9дм 134 | 0л 135 | 1л 136 | 2л 137 | 3л 138 | 4л 139 | 5л 140 | 6л 141 | 7л 142 | 8л 143 | 9л 144 | 0км 145 | 1км 146 | 2км 147 | 3км 148 | 4км 149 | 5км 150 | 6км 151 | 7км 152 | 8км 153 | 9км 154 | 0га 155 | 1га 156 | 2га 157 | 3га 158 | 4га 159 | 5га 160 | 6га 161 | 7га 162 | 8га 163 | 9га 164 | 0кг 165 | 1кг 166 | 2кг 167 | 3кг 168 | 4кг 169 | 5кг 170 | 6кг 171 | 7кг 172 | 8кг 173 | 9кг 174 | 0т 175 | 1т 176 | 2т 177 | 3т 178 | 4т 179 | 5т 180 | 6т 181 | 7т 182 | 8т 183 | 9т 184 | 0г 185 | 1г 186 | 2г 187 | 3г 188 | 4г 189 | 5г 190 | 6г 191 | 7г 192 | 8г 193 | 9г 194 | 0мг 195 | 1мг 196 | 2мг 197 | 3мг 198 | 4мг 199 | 5мг 200 | 6мг 201 | 7мг 202 | 8мг 203 | 9мг 204 | бульв 205 | в 206 | вв 207 | г 208 | га 209 | гг 210 | гл 211 | гос 212 | д 213 | дм 214 | доп 215 | др 216 | е 217 | ед 218 | ед 219 | зам 220 | и 221 | инд 222 | исп 223 | Исп 224 | к 225 | кап 226 | кг 227 | кв 228 | кл 229 | км 230 | кол 231 | комн 232 | коп 233 | куб 234 | л 235 | лиц 236 | лл 237 | м 238 | макс 239 | мг 240 | мин 241 | мл 242 | млн 243 | млрд 244 | мм 245 | н 246 | наб 247 | нач 248 | неуд 249 | ном 250 | о 251 | обл 252 | обр 253 | общ 254 | ок 255 | ост 256 | отл 257 | п 258 | пер 259 | перераб 260 | пл 261 | пос 262 | пр 263 | просп 264 | проф 265 | р 266 | ред 267 | руб 268 | с 269 | сб 270 | св 271 | см 272 | соч 273 | ср 274 | ст 275 | стр 276 | т 277 | тел 278 | Тел 279 | тех 280 | тт 281 | туп 282 | тыс 283 | уд 284 | ул 285 | уч 286 | физ 287 | х 288 | хор 289 | ч 290 | чел 291 | шт 292 | экз 293 | э 294 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.sk: -------------------------------------------------------------------------------- 1 | Bc 2 | Mgr 3 | RNDr 4 | PharmDr 5 | PhDr 6 | JUDr 7 | PaedDr 8 | ThDr 9 | Ing 10 | MUDr 11 | MDDr 12 | MVDr 13 | Dr 14 | ThLic 15 | PhD 16 | ArtD 17 | ThDr 18 | Dr 19 | DrSc 20 | CSs 21 | prof 22 | obr 23 | Obr 24 | Č 25 | č 26 | absol 27 | adj 28 | admin 29 | adr 30 | Adr 31 | adv 32 | advok 33 | afr 34 | ak 35 | akad 36 | akc 37 | akuz 38 | et 39 | al 40 | alch 41 | amer 42 | anat 43 | angl 44 | Angl 45 | anglosas 46 | anorg 47 | ap 48 | apod 49 | arch 50 | archeol 51 | archit 52 | arg 53 | art 54 | astr 55 | astrol 56 | astron 57 | atp 58 | atď 59 | austr 60 | Austr 61 | aut 62 | belg 63 | Belg 64 | bibl 65 | Bibl 66 | biol 67 | bot 68 | bud 69 | bás 70 | býv 71 | cest 72 | chem 73 | cirk 74 | csl 75 | čs 76 | Čs 77 | dat 78 | dep 79 | det 80 | dial 81 | diaľ 82 | dipl 83 | distrib 84 | dokl 85 | dosl 86 | dopr 87 | dram 88 | duš 89 | dv 90 | dvojčl 91 | dór 92 | ekol 93 | ekon 94 | el 95 | elektr 96 | elektrotech 97 | energet 98 | epic 99 | est 100 | etc 101 | etonym 102 | eufem 103 | európ 104 | Európ 105 | ev 106 | evid 107 | expr 108 | fa 109 | fam 110 | farm 111 | fem 112 | feud 113 | fil 114 | filat 115 | filoz 116 | fi 117 | fon 118 | form 119 | fot 120 | fr 121 | Fr 122 | franc 123 | Franc 124 | fraz 125 | fut 126 | fyz 127 | fyziol 128 | garb 129 | gen 130 | genet 131 | genpor 132 | geod 133 | geogr 134 | geol 135 | geom 136 | germ 137 | gr 138 | Gr 139 | gréc 140 | Gréc 141 | gréckokat 142 | hebr 143 | herald 144 | hist 145 | hlav 146 | hosp 147 | hromad 148 | hud 149 | hypok 150 | ident 151 | i.e 152 | ident 153 | imp 154 | impf 155 | indoeur 156 | inf 157 | inform 158 | instr 159 | int 160 | interj 161 | inšt 162 | inštr 163 | iron 164 | jap 165 | Jap 166 | jaz 167 | jedn 168 | juhoamer 169 | juhových 170 | juhozáp 171 | juž 172 | kanad 173 | Kanad 174 | kanc 175 | kapit 176 | kpt 177 | kart 178 | katastr 179 | knih 180 | kniž 181 | komp 182 | konj 183 | konkr 184 | kozmet 185 | krajč 186 | kresť 187 | kt 188 | kuch 189 | lat 190 | latinskoamer 191 | lek 192 | lex 193 | lingv 194 | lit 195 | litur 196 | log 197 | lok 198 | max 199 | Max 200 | maď 201 | Maď 202 | medzinár 203 | mest 204 | metr 205 | mil 206 | Mil 207 | min 208 | Min 209 | miner 210 | ml 211 | mld 212 | mn 213 | mod 214 | mytol 215 | napr 216 | nar 217 | Nar 218 | nasl 219 | nedok 220 | neg 221 | negat 222 | neklas 223 | nem 224 | Nem 225 | neodb 226 | neos 227 | neskl 228 | nesklon 229 | nespis 230 | nespráv 231 | neved 232 | než 233 | niekt 234 | niž 235 | nom 236 | náb 237 | nákl 238 | námor 239 | nár 240 | obch 241 | obj 242 | obv 243 | obyč 244 | obč 245 | občian 246 | odb 247 | odd 248 | ods 249 | ojed 250 | okr 251 | Okr 252 | opt 253 | opyt 254 | org 255 | os 256 | osob 257 | ot 258 | ovoc 259 | par 260 | part 261 | pejor 262 | pers 263 | pf 264 | Pf 265 | P.f 266 | p.f 267 | pl 268 | Plk 269 | pod 270 | podst 271 | pokl 272 | polit 273 | politol 274 | polygr 275 | pomn 276 | popl 277 | por 278 | porad 279 | porov 280 | posch 281 | potrav 282 | použ 283 | poz 284 | pozit 285 | poľ 286 | poľno 287 | poľnohosp 288 | poľov 289 | pošt 290 | pož 291 | prac 292 | predl 293 | pren 294 | prep 295 | preuk 296 | priezv 297 | Priezv 298 | privl 299 | prof 300 | práv 301 | príd 302 | príj 303 | prík 304 | príp 305 | prír 306 | prísl 307 | príslov 308 | príč 309 | psych 310 | publ 311 | pís 312 | písm 313 | pôv 314 | refl 315 | reg 316 | rep 317 | resp 318 | rozk 319 | rozlič 320 | rozpráv 321 | roč 322 | Roč 323 | ryb 324 | rádiotech 325 | rím 326 | samohl 327 | semest 328 | sev 329 | severoamer 330 | severových 331 | severozáp 332 | sg 333 | skr 334 | skup 335 | sl 336 | Sloven 337 | soc 338 | soch 339 | sociol 340 | sp 341 | spol 342 | Spol 343 | spoloč 344 | spoluhl 345 | správ 346 | spôs 347 | st 348 | star 349 | starogréc 350 | starorím 351 | s.r.o 352 | stol 353 | stor 354 | str 355 | stredoamer 356 | stredoškol 357 | subj 358 | subst 359 | superl 360 | sv 361 | sz 362 | súkr 363 | súp 364 | súvzť 365 | tal 366 | Tal 367 | tech 368 | tel 369 | Tel 370 | telef 371 | teles 372 | telev 373 | teol 374 | trans 375 | turist 376 | tuzem 377 | typogr 378 | tzn 379 | tzv 380 | ukaz 381 | ul 382 | Ul 383 | umel 384 | univ 385 | ust 386 | ved 387 | vedľ 388 | verb 389 | veter 390 | vin 391 | viď 392 | vl 393 | vod 394 | vodohosp 395 | pnl 396 | vulg 397 | vyj 398 | vys 399 | vysokoškol 400 | vzťaž 401 | vôb 402 | vých 403 | výd 404 | výrob 405 | výsk 406 | výsl 407 | výtv 408 | výtvar 409 | význ 410 | včel 411 | vš 412 | všeob 413 | zahr 414 | zar 415 | zariad 416 | zast 417 | zastar 418 | zastaráv 419 | zb 420 | zdravot 421 | združ 422 | zjemn 423 | zlat 424 | zn 425 | Zn 426 | zool 427 | zr 428 | zried 429 | zv 430 | záhr 431 | zák 432 | zákl 433 | zám 434 | záp 435 | západoeur 436 | zázn 437 | územ 438 | účt 439 | čast 440 | čes 441 | Čes 442 | čl 443 | čísl 444 | živ 445 | pr 446 | fak 447 | Kr 448 | p.n.l 449 | A 450 | B 451 | C 452 | D 453 | E 454 | F 455 | G 456 | H 457 | I 458 | J 459 | K 460 | L 461 | M 462 | N 463 | O 464 | P 465 | Q 466 | R 467 | S 468 | T 469 | U 470 | V 471 | W 472 | X 473 | Y 474 | Z 475 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.sl: -------------------------------------------------------------------------------- 1 | dr 2 | Dr 3 | itd 4 | itn 5 | št #NUMERIC_ONLY# 6 | Št #NUMERIC_ONLY# 7 | d 8 | jan 9 | Jan 10 | feb 11 | Feb 12 | mar 13 | Mar 14 | apr 15 | Apr 16 | jun 17 | Jun 18 | jul 19 | Jul 20 | avg 21 | Avg 22 | sept 23 | Sept 24 | sep 25 | Sep 26 | okt 27 | Okt 28 | nov 29 | Nov 30 | dec 31 | Dec 32 | tj 33 | Tj 34 | npr 35 | Npr 36 | sl 37 | Sl 38 | op 39 | Op 40 | gl 41 | Gl 42 | oz 43 | Oz 44 | prev 45 | dipl 46 | ing 47 | prim 48 | Prim 49 | cf 50 | Cf 51 | gl 52 | Gl 53 | A 54 | B 55 | C 56 | D 57 | E 58 | F 59 | G 60 | H 61 | I 62 | J 63 | K 64 | L 65 | M 66 | N 67 | O 68 | P 69 | Q 70 | R 71 | S 72 | T 73 | U 74 | V 75 | W 76 | X 77 | Y 78 | Z 79 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.sv: -------------------------------------------------------------------------------- 1 | #single upper case letter are usually initials 2 | A 3 | B 4 | C 5 | D 6 | E 7 | F 8 | G 9 | H 10 | I 11 | J 12 | K 13 | L 14 | M 15 | N 16 | O 17 | P 18 | Q 19 | R 20 | S 21 | T 22 | U 23 | V 24 | W 25 | X 26 | Y 27 | Z 28 | #misc abbreviations 29 | AB 30 | G 31 | VG 32 | dvs 33 | etc 34 | from 35 | iaf 36 | jfr 37 | kl 38 | kr 39 | mao 40 | mfl 41 | mm 42 | osv 43 | pga 44 | tex 45 | tom 46 | vs 47 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.ta: -------------------------------------------------------------------------------- 1 | #Anything in this file, followed by a period (and an upper-case word), does NOT indicate an end-of-sentence marker. 2 | #Special cases are included for prefixes that ONLY appear before 0-9 numbers. 3 | 4 | #any single upper case letter followed by a period is not a sentence ender (excluding I occasionally, but we leave it in) 5 | #usually upper case letters are initials in a name 6 | அ 7 | ஆ 8 | இ 9 | ஈ 10 | உ 11 | ஊ 12 | எ 13 | ஏ 14 | ஐ 15 | ஒ 16 | ஓ 17 | ஔ 18 | ஃ 19 | க 20 | கா 21 | கி 22 | கீ 23 | கு 24 | கூ 25 | கெ 26 | கே 27 | கை 28 | கொ 29 | கோ 30 | கௌ 31 | க் 32 | ச 33 | சா 34 | சி 35 | சீ 36 | சு 37 | சூ 38 | செ 39 | சே 40 | சை 41 | சொ 42 | சோ 43 | சௌ 44 | ச் 45 | ட 46 | டா 47 | டி 48 | டீ 49 | டு 50 | டூ 51 | டெ 52 | டே 53 | டை 54 | டொ 55 | டோ 56 | டௌ 57 | ட் 58 | த 59 | தா 60 | தி 61 | தீ 62 | து 63 | தூ 64 | தெ 65 | தே 66 | தை 67 | தொ 68 | தோ 69 | தௌ 70 | த் 71 | ப 72 | பா 73 | பி 74 | பீ 75 | பு 76 | பூ 77 | பெ 78 | பே 79 | பை 80 | பொ 81 | போ 82 | பௌ 83 | ப் 84 | ற 85 | றா 86 | றி 87 | றீ 88 | று 89 | றூ 90 | றெ 91 | றே 92 | றை 93 | றொ 94 | றோ 95 | றௌ 96 | ற் 97 | ய 98 | யா 99 | யி 100 | யீ 101 | யு 102 | யூ 103 | யெ 104 | யே 105 | யை 106 | யொ 107 | யோ 108 | யௌ 109 | ய் 110 | ர 111 | ரா 112 | ரி 113 | ரீ 114 | ரு 115 | ரூ 116 | ரெ 117 | ரே 118 | ரை 119 | ரொ 120 | ரோ 121 | ரௌ 122 | ர் 123 | ல 124 | லா 125 | லி 126 | லீ 127 | லு 128 | லூ 129 | லெ 130 | லே 131 | லை 132 | லொ 133 | லோ 134 | லௌ 135 | ல் 136 | வ 137 | வா 138 | வி 139 | வீ 140 | வு 141 | வூ 142 | வெ 143 | வே 144 | வை 145 | வொ 146 | வோ 147 | வௌ 148 | வ் 149 | ள 150 | ளா 151 | ளி 152 | ளீ 153 | ளு 154 | ளூ 155 | ளெ 156 | ளே 157 | ளை 158 | ளொ 159 | ளோ 160 | ளௌ 161 | ள் 162 | ழ 163 | ழா 164 | ழி 165 | ழீ 166 | ழு 167 | ழூ 168 | ழெ 169 | ழே 170 | ழை 171 | ழொ 172 | ழோ 173 | ழௌ 174 | ழ் 175 | ங 176 | ஙா 177 | ஙி 178 | ஙீ 179 | ஙு 180 | ஙூ 181 | ஙெ 182 | ஙே 183 | ஙை 184 | ஙொ 185 | ஙோ 186 | ஙௌ 187 | ங் 188 | ஞ 189 | ஞா 190 | ஞி 191 | ஞீ 192 | ஞு 193 | ஞூ 194 | ஞெ 195 | ஞே 196 | ஞை 197 | ஞொ 198 | ஞோ 199 | ஞௌ 200 | ஞ் 201 | ண 202 | ணா 203 | ணி 204 | ணீ 205 | ணு 206 | ணூ 207 | ணெ 208 | ணே 209 | ணை 210 | ணொ 211 | ணோ 212 | ணௌ 213 | ண் 214 | ந 215 | நா 216 | நி 217 | நீ 218 | நு 219 | நூ 220 | நெ 221 | நே 222 | நை 223 | நொ 224 | நோ 225 | நௌ 226 | ந் 227 | ம 228 | மா 229 | மி 230 | மீ 231 | மு 232 | மூ 233 | மெ 234 | மே 235 | மை 236 | மொ 237 | மோ 238 | மௌ 239 | ம் 240 | ன 241 | னா 242 | னி 243 | னீ 244 | னு 245 | னூ 246 | னெ 247 | னே 248 | னை 249 | னொ 250 | னோ 251 | னௌ 252 | ன் 253 | 254 | 255 | #List of titles. These are often followed by upper-case names, but do not indicate sentence breaks 256 | திரு 257 | திருமதி 258 | வண 259 | கௌரவ 260 | 261 | 262 | #misc - odd period-ending items that NEVER indicate breaks (p.m. does NOT fall into this category - it sometimes ends a sentence) 263 | உ.ம் 264 | #கா.ம் 265 | #எ.ம் 266 | 267 | 268 | #Numbers only. These should only induce breaks when followed by a numeric sequence 269 | # add NUMERIC_ONLY after the word for this function 270 | #This case is mostly for the english "No." which can either be a sentence of its own, or 271 | #if followed by a number, a non-breaking prefix 272 | No #NUMERIC_ONLY# 273 | Nos 274 | Art #NUMERIC_ONLY# 275 | Nr 276 | pp #NUMERIC_ONLY# 277 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.yue: -------------------------------------------------------------------------------- 1 | # 2 | # Cantonese (Chinese) 3 | # 4 | # Anything in this file, followed by a period, 5 | # does NOT indicate an end-of-sentence marker. 6 | # 7 | # English/Euro-language given-name initials (appearing in 8 | # news, periodicals, etc.) 9 | A 10 | Ā 11 | B 12 | C 13 | Č 14 | D 15 | E 16 | Ē 17 | F 18 | G 19 | Ģ 20 | H 21 | I 22 | Ī 23 | J 24 | K 25 | Ķ 26 | L 27 | Ļ 28 | M 29 | N 30 | Ņ 31 | O 32 | P 33 | Q 34 | R 35 | S 36 | Š 37 | T 38 | U 39 | Ū 40 | V 41 | W 42 | X 43 | Y 44 | Z 45 | Ž 46 | 47 | # Numbers only. These should only induce breaks when followed by 48 | # a numeric sequence. 49 | # Add NUMERIC_ONLY after the word for this function. This case is 50 | # mostly for the english "No." which can either be a sentence of its 51 | # own, or if followed by a number, a non-breaking prefix. 52 | No #NUMERIC_ONLY# 53 | Nr #NUMERIC_ONLY# 54 | -------------------------------------------------------------------------------- /tools/nonbreaking_prefixes/nonbreaking_prefix.zh: -------------------------------------------------------------------------------- 1 | # 2 | # Mandarin (Chinese) 3 | # 4 | # Anything in this file, followed by a period, 5 | # does NOT indicate an end-of-sentence marker. 6 | # 7 | # English/Euro-language given-name initials (appearing in 8 | # news, periodicals, etc.) 9 | A 10 | Ā 11 | B 12 | C 13 | Č 14 | D 15 | E 16 | Ē 17 | F 18 | G 19 | Ģ 20 | H 21 | I 22 | Ī 23 | J 24 | K 25 | Ķ 26 | L 27 | Ļ 28 | M 29 | N 30 | Ņ 31 | O 32 | P 33 | Q 34 | R 35 | S 36 | Š 37 | T 38 | U 39 | Ū 40 | V 41 | W 42 | X 43 | Y 44 | Z 45 | Ž 46 | 47 | # Numbers only. These should only induce breaks when followed by 48 | # a numeric sequence. 49 | # Add NUMERIC_ONLY after the word for this function. This case is 50 | # mostly for the english "No." which can either be a sentence of its 51 | # own, or if followed by a number, a non-breaking prefix. 52 | No #NUMERIC_ONLY# 53 | Nr #NUMERIC_ONLY# 54 | -------------------------------------------------------------------------------- /tools/release_model.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | import argparse 3 | import torch 4 | 5 | if __name__ == "__main__": 6 | parser = argparse.ArgumentParser( 7 | description="Removes the optim data of PyTorch models") 8 | parser.add_argument("--model", "-m", 9 | help="The model filename (*.pt)", required=True) 10 | parser.add_argument("--output", "-o", 11 | help="The output filename (*.pt)", required=True) 12 | opt = parser.parse_args() 13 | 14 | model = torch.load(opt.model) 15 | model['optim'] = None 16 | torch.save(model, opt.output) 17 | -------------------------------------------------------------------------------- /tools/test_rouge.py: -------------------------------------------------------------------------------- 1 | # -*- encoding: utf-8 -*- 2 | import argparse 3 | import os 4 | import time 5 | import pyrouge 6 | import shutil 7 | import sys 8 | 9 | 10 | def test_rouge(cand, ref): 11 | """Calculate ROUGE scores of sequences passed as an iterator 12 | e.g. a list of str, an open file, StringIO or even sys.stdin 13 | """ 14 | current_time = time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime()) 15 | tmp_dir = ".rouge-tmp-{}".format(current_time) 16 | try: 17 | if not os.path.isdir(tmp_dir): 18 | os.mkdir(tmp_dir) 19 | os.mkdir(tmp_dir + "/candidate") 20 | os.mkdir(tmp_dir + "/reference") 21 | candidates = [line.strip() for line in cand] 22 | references = [line.strip() for line in ref] 23 | assert len(candidates) == len(references) 24 | cnt = len(candidates) 25 | for i in range(cnt): 26 | if len(references[i]) < 1: 27 | continue 28 | with open(tmp_dir + "/candidate/cand.{}.txt".format(i), "w", 29 | encoding="utf-8") as f: 30 | f.write(candidates[i]) 31 | with open(tmp_dir + "/reference/ref.{}.txt".format(i), "w", 32 | encoding="utf-8") as f: 33 | f.write(references[i]) 34 | r = pyrouge.Rouge155() 35 | r.model_dir = tmp_dir + "/reference/" 36 | r.system_dir = tmp_dir + "/candidate/" 37 | r.model_filename_pattern = 'ref.#ID#.txt' 38 | r.system_filename_pattern = 'cand.(\d+).txt' 39 | rouge_results = r.convert_and_evaluate() 40 | results_dict = r.output_to_dict(rouge_results) 41 | return results_dict 42 | finally: 43 | pass 44 | if os.path.isdir(tmp_dir): 45 | shutil.rmtree(tmp_dir) 46 | 47 | 48 | def rouge_results_to_str(results_dict): 49 | return ">> ROUGE(1/2/3/L/SU4): {:.2f}/{:.2f}/{:.2f}/{:.2f}/{:.2f}".format( 50 | results_dict["rouge_1_f_score"] * 100, 51 | results_dict["rouge_2_f_score"] * 100, 52 | results_dict["rouge_3_f_score"] * 100, 53 | results_dict["rouge_l_f_score"] * 100, 54 | results_dict["rouge_su*_f_score"] * 100) 55 | 56 | 57 | if __name__ == "__main__": 58 | parser = argparse.ArgumentParser() 59 | parser.add_argument('-c', type=str, default="candidate.txt", 60 | help='candidate file') 61 | parser.add_argument('-r', type=str, default="reference.txt", 62 | help='reference file') 63 | args = parser.parse_args() 64 | if args.c.upper() == "STDIN": 65 | args.c = sys.stdin 66 | results_dict = test_rouge(args.c, args.r) 67 | print(rouge_results_to_str(results_dict)) 68 | -------------------------------------------------------------------------------- /translate.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | from __future__ import division, unicode_literals 3 | import argparse 4 | 5 | from onmt.translate.Translator import make_translator 6 | 7 | import onmt.io 8 | import onmt.translate 9 | import onmt 10 | import onmt.ModelConstructor 11 | import onmt.modules 12 | import onmt.opts 13 | 14 | 15 | def main(opt): 16 | translator = make_translator(opt, report_score=True) 17 | translator.translate(opt.src_dir, opt.src, opt.tgt, 18 | opt.batch_size, opt.attn_debug) 19 | 20 | 21 | if __name__ == "__main__": 22 | parser = argparse.ArgumentParser( 23 | description='translate.py', 24 | formatter_class=argparse.ArgumentDefaultsHelpFormatter) 25 | onmt.opts.add_md_help_argument(parser) 26 | onmt.opts.translate_opts(parser) 27 | 28 | opt = parser.parse_args() 29 | main(opt) 30 | --------------------------------------------------------------------------------