├── .gitignore
├── LICENSE
├── README.md
├── arch.png
├── construct_sg.py
├── data_load.py
├── demo.gif
├── download.sh
├── encode.py
├── hparams.py
├── make_phr2sg_id.py
├── model.py
├── prepro.py
├── test.py
└── train.py
/.gitignore:
--------------------------------------------------------------------------------
1 | # Byte-compiled / optimized / DLL files
2 | __pycache__/
3 | *.py[cod]
4 | *$py.class
5 |
6 | # C extensions
7 | *.so
8 |
9 | # Distribution / packaging
10 | .Python
11 | build/
12 | develop-eggs/
13 | dist/
14 | downloads/
15 | eggs/
16 | .eggs/
17 | lib/
18 | lib64/
19 | parts/
20 | sdist/
21 | var/
22 | wheels/
23 | *.egg-info/
24 | .installed.cfg
25 | *.egg
26 | MANIFEST
27 |
28 | # PyInstaller
29 | # Usually these files are written by a python script from a template
30 | # before PyInstaller builds the exe, so as to inject date/other infos into it.
31 | *.manifest
32 | *.spec
33 |
34 | # Installer logs
35 | pip-log.txt
36 | pip-delete-this-directory.txt
37 |
38 | # Unit test / coverage reports
39 | htmlcov/
40 | .tox/
41 | .coverage
42 | .coverage.*
43 | .cache
44 | nosetests.xml
45 | coverage.xml
46 | *.cover
47 | .hypothesis/
48 | .pytest_cache/
49 |
50 | # Translations
51 | *.mo
52 | *.pot
53 |
54 | # Django stuff:
55 | *.log
56 | local_settings.py
57 | db.sqlite3
58 |
59 | # Flask stuff:
60 | instance/
61 | .webassets-cache
62 |
63 | # Scrapy stuff:
64 | .scrapy
65 |
66 | # Sphinx documentation
67 | docs/_build/
68 |
69 | # PyBuilder
70 | target/
71 |
72 | # Jupyter Notebook
73 | .ipynb_checkpoints
74 |
75 | # pyenv
76 | .python-version
77 |
78 | # celery beat schedule file
79 | celerybeat-schedule
80 |
81 | # SageMath parsed files
82 | *.sage.py
83 |
84 | # Environments
85 | .env
86 | .venv
87 | env/
88 | venv/
89 | ENV/
90 | env.bak/
91 | venv.bak/
92 |
93 | # Spyder project settings
94 | .spyderproject
95 | .spyproject
96 |
97 | # Rope project settings
98 | .ropeproject
99 |
100 | # mkdocs documentation
101 | /site
102 |
103 | # mypy
104 | .mypy_cache/
105 |
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | Apache License
2 | Version 2.0, January 2004
3 | http://www.apache.org/licenses/
4 |
5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6 |
7 | 1. Definitions.
8 |
9 | "License" shall mean the terms and conditions for use, reproduction,
10 | and distribution as defined by Sections 1 through 9 of this document.
11 |
12 | "Licensor" shall mean the copyright owner or entity authorized by
13 | the copyright owner that is granting the License.
14 |
15 | "Legal Entity" shall mean the union of the acting entity and all
16 | other entities that control, are controlled by, or are under common
17 | control with that entity. For the purposes of this definition,
18 | "control" means (i) the power, direct or indirect, to cause the
19 | direction or management of such entity, whether by contract or
20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the
21 | outstanding shares, or (iii) beneficial ownership of such entity.
22 |
23 | "You" (or "Your") shall mean an individual or Legal Entity
24 | exercising permissions granted by this License.
25 |
26 | "Source" form shall mean the preferred form for making modifications,
27 | including but not limited to software source code, documentation
28 | source, and configuration files.
29 |
30 | "Object" form shall mean any form resulting from mechanical
31 | transformation or translation of a Source form, including but
32 | not limited to compiled object code, generated documentation,
33 | and conversions to other media types.
34 |
35 | "Work" shall mean the work of authorship, whether in Source or
36 | Object form, made available under the License, as indicated by a
37 | copyright notice that is included in or attached to the work
38 | (an example is provided in the Appendix below).
39 |
40 | "Derivative Works" shall mean any work, whether in Source or Object
41 | form, that is based on (or derived from) the Work and for which the
42 | editorial revisions, annotations, elaborations, or other modifications
43 | represent, as a whole, an original work of authorship. For the purposes
44 | of this License, Derivative Works shall not include works that remain
45 | separable from, or merely link (or bind by name) to the interfaces of,
46 | the Work and Derivative Works thereof.
47 |
48 | "Contribution" shall mean any work of authorship, including
49 | the original version of the Work and any modifications or additions
50 | to that Work or Derivative Works thereof, that is intentionally
51 | submitted to Licensor for inclusion in the Work by the copyright owner
52 | or by an individual or Legal Entity authorized to submit on behalf of
53 | the copyright owner. For the purposes of this definition, "submitted"
54 | means any form of electronic, verbal, or written communication sent
55 | to the Licensor or its representatives, including but not limited to
56 | communication on electronic mailing lists, source code control systems,
57 | and issue tracking systems that are managed by, or on behalf of, the
58 | Licensor for the purpose of discussing and improving the Work, but
59 | excluding communication that is conspicuously marked or otherwise
60 | designated in writing by the copyright owner as "Not a Contribution."
61 |
62 | "Contributor" shall mean Licensor and any individual or Legal Entity
63 | on behalf of whom a Contribution has been received by Licensor and
64 | subsequently incorporated within the Work.
65 |
66 | 2. Grant of Copyright License. Subject to the terms and conditions of
67 | this License, each Contributor hereby grants to You a perpetual,
68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69 | copyright license to reproduce, prepare Derivative Works of,
70 | publicly display, publicly perform, sublicense, and distribute the
71 | Work and such Derivative Works in Source or Object form.
72 |
73 | 3. Grant of Patent License. Subject to the terms and conditions of
74 | this License, each Contributor hereby grants to You a perpetual,
75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76 | (except as stated in this section) patent license to make, have made,
77 | use, offer to sell, sell, import, and otherwise transfer the Work,
78 | where such license applies only to those patent claims licensable
79 | by such Contributor that are necessarily infringed by their
80 | Contribution(s) alone or by combination of their Contribution(s)
81 | with the Work to which such Contribution(s) was submitted. If You
82 | institute patent litigation against any entity (including a
83 | cross-claim or counterclaim in a lawsuit) alleging that the Work
84 | or a Contribution incorporated within the Work constitutes direct
85 | or contributory patent infringement, then any patent licenses
86 | granted to You under this License for that Work shall terminate
87 | as of the date such litigation is filed.
88 |
89 | 4. Redistribution. You may reproduce and distribute copies of the
90 | Work or Derivative Works thereof in any medium, with or without
91 | modifications, and in Source or Object form, provided that You
92 | meet the following conditions:
93 |
94 | (a) You must give any other recipients of the Work or
95 | Derivative Works a copy of this License; and
96 |
97 | (b) You must cause any modified files to carry prominent notices
98 | stating that You changed the files; and
99 |
100 | (c) You must retain, in the Source form of any Derivative Works
101 | that You distribute, all copyright, patent, trademark, and
102 | attribution notices from the Source form of the Work,
103 | excluding those notices that do not pertain to any part of
104 | the Derivative Works; and
105 |
106 | (d) If the Work includes a "NOTICE" text file as part of its
107 | distribution, then any Derivative Works that You distribute must
108 | include a readable copy of the attribution notices contained
109 | within such NOTICE file, excluding those notices that do not
110 | pertain to any part of the Derivative Works, in at least one
111 | of the following places: within a NOTICE text file distributed
112 | as part of the Derivative Works; within the Source form or
113 | documentation, if provided along with the Derivative Works; or,
114 | within a display generated by the Derivative Works, if and
115 | wherever such third-party notices normally appear. The contents
116 | of the NOTICE file are for informational purposes only and
117 | do not modify the License. You may add Your own attribution
118 | notices within Derivative Works that You distribute, alongside
119 | or as an addendum to the NOTICE text from the Work, provided
120 | that such additional attribution notices cannot be construed
121 | as modifying the License.
122 |
123 | You may add Your own copyright statement to Your modifications and
124 | may provide additional or different license terms and conditions
125 | for use, reproduction, or distribution of Your modifications, or
126 | for any such Derivative Works as a whole, provided Your use,
127 | reproduction, and distribution of the Work otherwise complies with
128 | the conditions stated in this License.
129 |
130 | 5. Submission of Contributions. Unless You explicitly state otherwise,
131 | any Contribution intentionally submitted for inclusion in the Work
132 | by You to the Licensor shall be under the terms and conditions of
133 | this License, without any additional terms or conditions.
134 | Notwithstanding the above, nothing herein shall supersede or modify
135 | the terms of any separate license agreement you may have executed
136 | with Licensor regarding such Contributions.
137 |
138 | 6. Trademarks. This License does not grant permission to use the trade
139 | names, trademarks, service marks, or product names of the Licensor,
140 | except as required for reasonable and customary use in describing the
141 | origin of the Work and reproducing the content of the NOTICE file.
142 |
143 | 7. Disclaimer of Warranty. Unless required by applicable law or
144 | agreed to in writing, Licensor provides the Work (and each
145 | Contributor provides its Contributions) on an "AS IS" BASIS,
146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147 | implied, including, without limitation, any warranties or conditions
148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149 | PARTICULAR PURPOSE. You are solely responsible for determining the
150 | appropriateness of using or redistributing the Work and assume any
151 | risks associated with Your exercise of permissions under this License.
152 |
153 | 8. Limitation of Liability. In no event and under no legal theory,
154 | whether in tort (including negligence), contract, or otherwise,
155 | unless required by applicable law (such as deliberate and grossly
156 | negligent acts) or agreed to in writing, shall any Contributor be
157 | liable to You for damages, including any direct, indirect, special,
158 | incidental, or consequential damages of any character arising as a
159 | result of this License or out of the use or inability to use the
160 | Work (including but not limited to damages for loss of goodwill,
161 | work stoppage, computer failure or malfunction, or any and all
162 | other commercial damages or losses), even if such Contributor
163 | has been advised of the possibility of such damages.
164 |
165 | 9. Accepting Warranty or Additional Liability. While redistributing
166 | the Work or Derivative Works thereof, You may choose to offer,
167 | and charge a fee for, acceptance of support, warranty, indemnity,
168 | or other liability obligations and/or rights consistent with this
169 | License. However, in accepting such obligations, You may act only
170 | on Your own behalf and on Your sole responsibility, not on behalf
171 | of any other Contributor, and only if You agree to indemnify,
172 | defend, and hold each Contributor harmless for any liability
173 | incurred by, or claims asserted against, such Contributor by reason
174 | of your accepting any such warranty or additional liability.
175 |
176 | END OF TERMS AND CONDITIONS
177 |
178 | APPENDIX: How to apply the Apache License to your work.
179 |
180 | To apply the Apache License to your work, attach the following
181 | boilerplate notice, with the fields enclosed by brackets "[]"
182 | replaced with your own identifying information. (Don't include
183 | the brackets!) The text should be enclosed in the appropriate
184 | comment syntax for the file format. We also recommend that a
185 | file or class name and description of purpose be included on the
186 | same "printed page" as the copyright notice for easier
187 | identification within third-party archives.
188 |
189 | Copyright [yyyy] [name of copyright owner]
190 |
191 | Licensed under the Apache License, Version 2.0 (the "License");
192 | you may not use this file except in compliance with the License.
193 | You may obtain a copy of the License at
194 |
195 | http://www.apache.org/licenses/LICENSE-2.0
196 |
197 | Unless required by applicable law or agreed to in writing, software
198 | distributed under the License is distributed on an "AS IS" BASIS,
199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200 | See the License for the specific language governing permissions and
201 | limitations under the License.
202 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # Smart Message Reply
2 |
3 | Have you ever seen or used [Google Smart Reply](https://firebase.google.com/docs/ml-kit/generate-smart-replies)? It's a service that provides automatic reply suggestions for user messages. See below.
4 |
5 | 
6 |
7 | This is a useful application of the retrieval based chatbot. Think about it. How many times do we text a message like thx, hey, or see you later?
8 | In this project, we build a simple message reply suggestion system.
9 |
10 | Kyubyong Park
11 | Code-review by [Yj Choe](https://github.com/yjchoe)
12 |
13 | ## Synonym group
14 | * We need to set the list of suggestions to show. Naturally, frequency is considered first. But what about those phrases that are similar in meaning? For example, should thank you so much and thxbe treated independently? We don't think so. We want to group them and save our slots. How? We make use of a parallel corpus. Both thank you so much and thx are likely to be translated into the same text. Based on this assumption, we construct English synonym groups that share the same translation.
15 |
16 | ## Model
17 | We fine-tune [huggingface's](https://github.com/huggingface/pytorch-pretrained-BERT) the [Bert](https://arxiv.org/abs/1810.04805) pretrained model for sequence classification. In it, a special starting token [CLS] stores the entire information of a sentence. Extra layers are attached to project the condensed information to classification units (here 100).
18 |
19 | 
20 |
21 | ## Data
22 | * We use [OpenSubtitles 2018](http://opus.nlpl.eu/OpenSubtitles-v2018.php) Spanish-English parallel corpus to construct synonym groups. OpenSubtitles is a large collection of translated movie subtitles. The en-es data consists of more than 61M aligned lines.
23 | * Ideally, a (very) large dialog corpus is needed for training, which we failed to find. We use the Cornell Movie Dialogue Corpus, instead. It's composed of 83,097 dialogues or 304,713 lines.
24 |
25 | ## Requirements
26 | * python>=3.6
27 | * tqdm>=4.30.0
28 | * pytorch>=1.0
29 | * pytorch_pretrained_bert>=0.6.1
30 | * nltk>=3.4
31 |
32 | ## Training
33 | * STEP 0. Download OpenSubtitles 2018 Spanish-English Parallel data.
34 | ```
35 | bash download.sh
36 | ```
37 |
38 | * STEP 1. Construct synonym groups from the corpus.
39 | ```
40 | python construct_sg.py
41 | ```
42 | * STEP 2. Make phr2sg_id and sg_id2phr dictionaries.
43 | ```
44 | python make_phr2sg_id.py
45 | ```
46 | * STEP 3. Convert a monolingual English text to ids.
47 | ```
48 | python encode.py
49 | ```
50 | * STEP 4. Create training data and save them as pickle.
51 | ```
52 | python prepro.py
53 | ```
54 | * STEP 5. Train.
55 | ```
56 | python train.py
57 | ```
58 |
59 | ## Test (Demo)
60 |
61 | 
62 |
63 | * Download and extract the [pre-trained model](https://www.dropbox.com/s/fqomn5flbwlvndc/log.tar.gz?dl=0) and run the following command.
64 | ```
65 | python test.py --ckpt log/9500_ACC0.1.pt
66 | ```
67 |
68 | ## Notes
69 | * Training loss slowly but steadily decreases.
70 | * Accuracy@5 on the evaluation data is from 10 to 20 percent.
71 | * For real application, a much much larger corpus is needed.
72 | * Not sure how much movie scripts are similar to message dialogues.
73 | * A better strategy for constructing synonym groups is necessary.
74 | * A retrieval-based chatbot is a realistic application as it is safter and easier than generation-based one.
75 |
76 |
--------------------------------------------------------------------------------
/arch.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Kyubyong/msg_reply/046f6308785d8e65d7ae429964df40a001a9675d/arch.png
--------------------------------------------------------------------------------
/construct_sg.py:
--------------------------------------------------------------------------------
1 | '''
2 | Construct synonym groups looking like this:
3 |
4 | [
5 | "0": {
6 | "_translation": "Te pondré más.",
7 | "phrases": [
8 | [
9 | "I'll give you more.",
10 | 1
11 | ],
12 | [
13 | "You'll have to have some more.",
14 | 1
15 | ],
16 | ...
17 | ]
18 | '''
19 |
20 | import json
21 | from collections import Counter
22 | from operator import itemgetter
23 | from hparams import hp
24 | import os
25 | from tqdm import tqdm
26 |
27 | def normalize(text):
28 | text = text.strip(" -\n")
29 | return text
30 |
31 | if __name__ == "__main__":
32 | # Group phrases
33 | es2ens = dict()
34 | en_lines = open(hp.opus_en, 'r').read().splitlines()
35 | es_lines = open(hp.opus_es, 'r').read().splitlines()
36 | for en, es in tqdm(zip(en_lines, es_lines), total=len(en_lines)):
37 | en = normalize(en)
38 | es = normalize(es)
39 | if len(es) <= 1: continue
40 | if es not in es2ens: es2ens[es] = []
41 | es2ens[es].append(en)
42 | print(f"Grouped all synonymous phrases: {len(es2ens)}")
43 |
44 | # Sort
45 | data = dict()
46 | i = 0
47 | for es, ens in es2ens.items():
48 | en2cnt = Counter(ens)
49 | phrases = sorted(en2cnt.items(), key=itemgetter(1), reverse=True)
50 | if len(phrases) > 1:
51 | val = dict()
52 | val["_translation"] = es
53 | val["phrases"] = phrases
54 | data[i] = val
55 | i += 1
56 | print(f"Sorted all synonymous groups by frequency: {len(data)}")
57 |
58 | # Write
59 | os.makedirs(os.path.dirname(hp.sg), exist_ok=True)
60 | with open(hp.sg, 'w') as fout:
61 | json.dump(data, fout, ensure_ascii=False, indent=4, separators=(',', ': '), sort_keys=True)
62 |
--------------------------------------------------------------------------------
/data_load.py:
--------------------------------------------------------------------------------
1 | from hparams import hp
2 | import random
3 | import pickle
4 | from itertools import chain
5 | import torch
6 | from glob import glob
7 |
8 | print("Loading training files")
9 |
10 | train_data = pickle.load(open(hp.pkl_train, 'rb'))
11 | dev_data = pickle.load(open(hp.pkl_dev, 'rb'))
12 |
13 | def pad(batch, maxlen):
14 | '''Pads to the longest sample'''
15 | return [sample + [0]*(maxlen-len(sample)) for sample in batch]
16 |
17 |
18 | def get_batch(max_span, batch_size, n_classes, train=True):
19 | '''f
20 | Returns
21 | x: (N, T)
22 | y: (N,)
23 | '''
24 | contexts_li = train_data if train else dev_data
25 |
26 | x, y, maxlen = [], [], 0
27 | for _ in range(batch_size):
28 | label = random.randint(0, n_classes-1) # randint: [a, b]
29 | try:
30 | contexts = contexts_li[label] # list of lists of lists
31 | except IndexError:
32 | continue
33 | if len(contexts) == 0: continue
34 | ctx = random.choice(contexts) # list of lists
35 | history_span = random.randint(1, len(ctx) + 1)
36 | history = ctx[-history_span:] # lists
37 |
38 | history = list(chain.from_iterable(history) ) # list
39 | history = history[-max_span+2:] # [3, 4, 5, ...]
40 | history = [101] + history + [102] # 101: [CLS], 102: [SEP]
41 | x.append(history)
42 | y.append(label)
43 | maxlen = max(maxlen, len(history))
44 |
45 | # print(f"len(x)={len(x)}, len(y)={len(y)}, maxlen={maxlen}")
46 | x = pad(x, maxlen)
47 | x = torch.LongTensor(x)
48 | y = torch.LongTensor(y)
49 | return x, y
50 |
51 |
52 |
53 |
--------------------------------------------------------------------------------
/demo.gif:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Kyubyong/msg_reply/046f6308785d8e65d7ae429964df40a001a9675d/demo.gif
--------------------------------------------------------------------------------
/download.sh:
--------------------------------------------------------------------------------
1 | #!/bin/bash
2 |
3 | echo "Download and extract OpenSubtitles 2018 en-es parallel data"
4 | echo "to opensubtitles2018"
5 | wget http://opus.nlpl.eu/download.php?f=OpenSubtitles/v2018/moses/en-es.txt.zip -O temp.zip;
6 | unzip temp.zip -d opensubtitles2018/;
7 | rm temp.zip
8 |
9 | echo "Download Cornell Movie Dialogue Corpus"
10 | wget http://www.cs.cornell.edu/~cristian/data/cornell_movie_dialogs_corpus.zip -O temp.zip;
11 | unzip temp.zip;
12 | rm temp.zip
13 |
14 |
--------------------------------------------------------------------------------
/encode.py:
--------------------------------------------------------------------------------
1 | '''
2 | Generate data/cornell.txt that encodes cornell corpus.
3 | It looks like this:
4 | [sg_id] [text] [encoding]
5 | 0 You makin' any headway? 2017 5003 4939 1005 2151 2132 4576 1029 1064
6 | 0 She kissed me. 2016 4782 2033 1012 1064
7 | 200020 Where? 2073 1029 1064
8 |
9 | '''
10 |
11 | import re, os
12 | import pickle
13 | from hparams import hp
14 | from pytorch_pretrained_bert import BertTokenizer
15 | from nltk.tokenize import sent_tokenize
16 | from tqdm import tqdm
17 | import codecs
18 |
19 |
20 | def refine(text):
21 | text = text.lower()
22 | text = re.sub("[^ A-Za-z\|]", "", text)
23 | return text
24 |
25 | def get_utterances(line):
26 | text = re.search("\[(.+?)\]", line).group(1)
27 | text = re.sub("[',]", "", text)
28 | utts = text.split()
29 | if len(utts) < 2:
30 | print(line)
31 | return utts
32 |
33 |
34 | if __name__ == "__main__":
35 | tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
36 |
37 | # Load dictionaries
38 | phr2sg_id = pickle.load(open(hp.phr2sg_id, 'rb'))
39 | sg_id2phr = pickle.load(open(hp.sg_id2phr, 'rb'))
40 |
41 | # Load Cornell movie corpus
42 | convs = os.path.join(hp.corpus, "movie_conversations.txt")
43 | lines = os.path.join(hp.corpus, "movie_lines.txt")
44 |
45 | indices_li = [get_utterances(line) for line in codecs.open(convs, 'r', "utf-8").read().splitlines()] # list of lists
46 | idx2utt = dict()
47 | for line in codecs.open(lines, 'r', "utf-8", errors="ignore").read().splitlines():
48 | cols = line.split("+++$+++")
49 | idx, utt = cols[0].strip(), cols[-1].strip()
50 | idx2utt[idx] = utt
51 |
52 | os.makedirs(os.path.dirname(hp.text), exist_ok=True)
53 | with open(hp.text, 'w') as fout:
54 | for i, indices in tqdm(enumerate(indices_li), total=len(indices_li)):
55 | if len(indices) < 2:
56 | print(indices)
57 | utts = [idx2utt[idx] for idx in indices]
58 |
59 | is_valid = True
60 | for utt in utts:
61 | if len(utt.strip()) < 1:
62 | is_valid = False
63 | break
64 | if not is_valid: continue
65 |
66 | for utt in utts:
67 | utt = utt.replace("\t", " ").replace(" ", " ")
68 | utt0 = sent_tokenize(utt)[0]
69 | utt0 = refine(utt0)
70 | sg_id = phr2sg_id.get(utt0, 0)
71 |
72 | tokens = tokenizer.tokenize(utt)[:512-1] # 512: max length of bert
73 | if len(tokens) == 0: continue
74 | tokens += ["|"] # utterance delimiter
75 | ids = tokenizer.convert_tokens_to_ids(tokens)
76 | ids = " ".join(str(idx) for idx in ids)
77 |
78 | # save
79 | fout.write(f"{sg_id}\t{utt}\t{ids}\n")
80 | fout.write("\n")
81 |
82 |
--------------------------------------------------------------------------------
/hparams.py:
--------------------------------------------------------------------------------
1 | class Hparams:
2 | # construct_sg
3 | opus_en = "opensubtitles2018/OpenSubtitles.en-es.en"
4 | opus_es = "opensubtitles2018/OpenSubtitles.en-es.es"
5 | sg = "data/sg.en.es.json"
6 |
7 | # make_phr2sg_id
8 | min_cnt = 5 # a phrase whose count is 5 or more is included
9 | n_phrs = 10000 # number of phrases
10 | phr2sg_id = "data/phr2sg_id.pkl"
11 | sg_id2phr = "data/sg_id2phr.pkl"
12 |
13 | # encode
14 | corpus = "cornell movie-dialogs corpus"
15 | text = "data/cornell.txt"
16 |
17 | # prepro
18 | pkl_train = 'data/train.pkl'
19 | pkl_dev = 'data/dev.pkl'
20 | n_classes = 100
21 | phr2idx = "data/phr2idx.pkl"
22 | idx2phr = "data/idx2phr.pkl"
23 |
24 | # train
25 | batch_size = 32*8 # 8 GPUs
26 | lr = 2e-5
27 | logdir = 'log'
28 | vocab_size = 28996
29 | max_span = 128 # maximum token length for context
30 | n_train_steps = 10000
31 |
32 | # also test
33 | n_candidates = 5
34 |
35 | hp = Hparams()
--------------------------------------------------------------------------------
/make_phr2sg_id.py:
--------------------------------------------------------------------------------
1 | '''
2 | Make two dictionaries: phr2sg_id and sg_id2phr
3 |
4 | phr2sg_id["nice work']==6152
5 | phr2sg_id["nicely done']==6152
6 | phr2sg_id["nice going']==6152
7 | sg_id2phr[6152]=="Well done."
8 |
9 | '''
10 |
11 |
12 | import json, os
13 | import operator
14 | import pickle
15 | from hparams import hp
16 | import re
17 | from tqdm import tqdm
18 |
19 | def refine(text):
20 | text = text.lower()
21 | text = re.sub("[^ A-Za-z]", "", text)
22 | return text
23 |
24 | if __name__ == "__main__":
25 | print("Determine the most frequent Synonym Groups")
26 | data = json.load(open(hp.sg))
27 | sg_id2cnt = dict()
28 | for sg_id, sg in tqdm(data.items()):
29 | sg_id = int(sg_id)
30 | phrs = sg["phrases"] # [['i am mormon', 1], ["i'm a mormon", 1]]
31 | sg_cnt = 0 # total cnt
32 | for phr, cnt in phrs:
33 | if cnt >= hp.min_cnt:
34 | sg_cnt += cnt
35 |
36 | sg_id2cnt[sg_id] = sg_cnt
37 |
38 | sg_id_cnt = sorted(sg_id2cnt.items(), key=operator.itemgetter(1), reverse=True)
39 | sg_ids = [sg_id for sg_id, _ in sg_id_cnt][:hp.n_phrs]
40 |
41 | print("Determine the group of phrases")
42 | sg_id2phr = dict()
43 | phr2sg_id, phr2cnt = dict(), dict()
44 | for sg_id in tqdm(sg_ids):
45 | sg = data[str(sg_id)]
46 | phrs = sg["phrases"] # [['i am mormon', 1], ["i'm a mormon", 1]]
47 |
48 | sg_id2phr[sg_id] = phrs[0][0]
49 | for phr, cnt in phrs:
50 | if cnt >= hp.min_cnt:
51 | phr = refine(phr)
52 | if phr in phr2cnt and cnt > phr2cnt[phr]: # overwrite
53 | phr2cnt[phr] = cnt
54 | phr2sg_id[phr] = sg_id
55 | else:
56 | phr2cnt[phr] = cnt
57 | phr2sg_id[phr] = sg_id
58 |
59 | print("save")
60 | os.makedirs(os.path.dirname(hp.phr2sg_id), exist_ok=True)
61 | os.makedirs(os.path.dirname(hp.sg_id2phr), exist_ok=True)
62 | pickle.dump(phr2sg_id, open(hp.phr2sg_id, 'wb'))
63 | pickle.dump(sg_id2phr, open(hp.sg_id2phr, 'wb'))
--------------------------------------------------------------------------------
/model.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import torch.nn as nn
3 | from pytorch_pretrained_bert import BertForSequenceClassification
4 | from hparams import hp
5 |
6 | class Net(nn.Module):
7 | def __init__(self, n_classes):
8 | super().__init__()
9 | self.bert = BertForSequenceClassification.from_pretrained('bert-base-uncased',
10 | num_labels=n_classes)
11 | self.softmax = nn.Softmax(-1)
12 |
13 | def forward(self, x):
14 | '''
15 | x: (N, T). int64
16 |
17 | Returns
18 | logits: (N, n_classes)
19 | y_hat: (N, n_candidates)
20 | y_hat_prob: (N, n_candidates)
21 |
22 | '''
23 | if self.training:
24 | self.bert.train()
25 | logits = self.bert(x)
26 | else:
27 | self.bert.eval()
28 | with torch.no_grad():
29 | logits = self.bert(x)
30 |
31 | activated = self.softmax(logits)
32 | y_hat_prob, y_hat = activated.sort(-1, descending=True)
33 | y_hat_prob = y_hat_prob[:, :hp.n_candidates]
34 | y_hat = y_hat[:, :hp.n_candidates]
35 |
36 | return logits, y_hat, y_hat_prob
37 |
38 |
--------------------------------------------------------------------------------
/prepro.py:
--------------------------------------------------------------------------------
1 | '''
2 | Make phr2idx, idx2phr, and {train|dev}.pkl
3 |
4 | idx2phr:
5 | {0: 'Yes!',
6 | 1: 'Good answer.',
7 | 2: 'What?',
8 | 3: 'Good.',
9 | 4: 'Of course I do.',
10 | 5: "I don't know.",
11 | 6: 'What for?',
12 | 7: 'Oh!',
13 | 8: 'Thank you.',
14 | 9: 'Hello?',
15 | 10: 'Right.',
16 | 11: 'I know.',
17 | 12: "What's wrong?",
18 | 13: 'Really?',
19 | 14: "Oh, I'm sorry.",
20 | 15: 'Oh, yes!',
21 | 16: 'Well...',
22 | 17: 'Yes, sir?',
23 | 18: 'Nothing.',
24 | 19: 'Hi!',
25 | 20: 'Huh!',
26 | 21: 'Why not?',
27 | 22: '10.',
28 | 23: 'Who?',
29 | 24: 'Stop it.',
30 | 25: 'Shit!',
31 | 26: 'What do you mean?',
32 | 27: 'Aha.',
33 | 28: 'Yes.',
34 | 29: 'Come on!',
35 | 30: 'Shut up!',
36 | 31: 'What the hell are you talking about?',
37 | 32: 'So.',
38 | 33: 'Excuse me...',
39 | 34: 'Which one?',
40 | 35: 'What are you doing?',
41 | 36: 'Where?',
42 | 37: 'Oh, I see.',
43 | 38: 'I beg you!',
44 | 39: 'Me!',
45 | 40: 'What happened?',
46 | 41: 'Great!',
47 | 42: 'Oh, no.',
48 | 43: 'Jesus!',
49 | 44: 'Maybe.',
50 | 45: 'This is it.',
51 | 46: 'Excuse me!',
52 | 47: 'No.',
53 | 48: 'I do.',
54 | 49: 'Wait?',
55 | 50: 'How?',
56 | 51: 'No, thank you.',
57 | 52: 'Forget it.',
58 | 53: 'Just like me.',
59 | 54: "I don't think so.",
60 | 55: 'I...',
61 | 56: 'We will.',
62 | 57: 'Nonsense.',
63 | 58: 'No, no',
64 | 59: 'Oh, my God.',
65 | 60: 'What is this?',
66 | 61: 'Look!',
67 | 62: "Can't I?",
68 | 63: 'No, sir.',
69 | 64: 'Here...',
70 | 65: "I'm fine.",
71 | 66: 'All right?',
72 | 67: "I don't understand!",
73 | 68: 'What do you want?',
74 | 69: 'Wait a minute!',
75 | 70: 'You!',
76 | 71: 'How wonderful!',
77 | 72: 'OK!',
78 | 73: 'When was it?',
79 | 74: 'All in order.',
80 | 75: 'Did I?',
81 | 76: 'I got it.',
82 | 77: 'Nope.',
83 | 78: 'Mmm?',
84 | 79: 'Sir',
85 | 80: 'Not a chance.',
86 | 81: 'Who are you?',
87 | 82: 'Good night...',
88 | 83: 'Die!',
89 | 84: 'What do you think?',
90 | 85: 'Not exactly.',
91 | 86: 'Where are you going?',
92 | 87: 'Are you all right?',
93 | 88: "I'm...",
94 | 89: 'Like what?',
95 | 90: 'I can imagine.',
96 | 91: "Don't be afraid.",
97 | 92: 'Huh?',
98 | 93: 'Of course.',
99 | 94: 'Bye!',
100 | 95: 'Yeah.',
101 | 96: 'Of course not!',
102 | 97: 'I got it.',
103 | 98: "No, it's not true.",
104 | 99: 'What does that mean?'}
105 |
106 | '''
107 |
108 |
109 | from hparams import hp
110 | import pickle, os
111 | from tqdm import tqdm
112 | from collections import Counter
113 |
114 | def get_most_frequent_sgs(fin, n_classes):
115 | sg_ids = []
116 | for line in open(fin, 'r'):
117 | if len(line) > 1:
118 | sg_id = line.split("\t")[0]
119 | sg_id = int(sg_id)
120 | if sg_id != 0: # 0: non-sg
121 | sg_ids.append(sg_id)
122 | sg_id2cnt = Counter(sg_ids)
123 | sg_ids = [sg_id for sg_id, cnt in sg_id2cnt.most_common(n_classes)]
124 | idx2sg_id = {idx: sg_id for idx, sg_id in enumerate(sg_ids)}
125 | sg_id2idx = {sg_id: idx for idx, sg_id in enumerate(sg_ids)}
126 | return idx2sg_id, sg_id2idx
127 |
128 | def prepro(fin, pkl_train, pkl_dev, n_classes, sg_id2idx):
129 | contexts_li = [[] for _ in range(n_classes)]
130 |
131 | entries = open(fin, 'r').read().split("\n\n")
132 | for entry in tqdm(entries):
133 | lines = entry.splitlines()
134 | for i, line in enumerate(lines):
135 | if i==0: continue
136 | cols = line.strip().split("\t")
137 | sg_id, sent, ids = cols
138 | sg_id = int(sg_id)
139 | if sg_id in sg_id2idx:
140 | idx = sg_id2idx[sg_id]
141 | ctx = [] # e.g. [ [3, 4, 5], [23, 9, 4, 5] ]
142 | for l in lines[:i]:
143 | ctx.append([int(id) for id in l.strip().split("\t")[-1].split()])
144 | contexts = contexts_li[idx]
145 | contexts.append(ctx)
146 | train, dev = [], []
147 | for contexts in contexts_li:
148 | if len(contexts) > 1:
149 | train.append(contexts[1:])
150 | dev.append(contexts[:1])
151 | else:
152 | train.append(contexts)
153 | dev.append([])
154 |
155 |
156 | pickle.dump(train, open(pkl_train, 'wb'))
157 | pickle.dump(dev, open(pkl_dev, 'wb'))
158 | print("done")
159 |
160 | if __name__ == "__main__":
161 | os.makedirs(os.path.dirname(hp.pkl_train), exist_ok=True)
162 | os.makedirs(os.path.dirname(hp.pkl_dev), exist_ok=True)
163 |
164 | idx2sg_id, sg_id2idx = get_most_frequent_sgs(hp.text, hp.n_classes)
165 |
166 | phr2sg_id = pickle.load(open(hp.phr2sg_id, 'rb'))
167 | sg_id2phr = pickle.load(open(hp.sg_id2phr, 'rb'))
168 |
169 | phr2idx = dict()
170 | for phr, sg_id in phr2sg_id.items():
171 | if sg_id in sg_id2idx:
172 | phr2idx[phr] = sg_id2idx[sg_id]
173 |
174 | idx2phr = dict()
175 | for idx, sg_id in idx2sg_id.items():
176 | if sg_id in sg_id2phr:
177 | idx2phr[idx] = sg_id2phr[sg_id]
178 |
179 | pickle.dump(phr2idx, open(hp.phr2idx, 'wb'))
180 | pickle.dump(idx2phr, open(hp.idx2phr, 'wb'))
181 |
182 | prepro(hp.text, hp.pkl_train, hp.pkl_dev, hp.n_classes, sg_id2idx)
183 | print("DONE")
--------------------------------------------------------------------------------
/test.py:
--------------------------------------------------------------------------------
1 | from hparams import hp
2 | import torch
3 | from model import Net
4 | from pytorch_pretrained_bert import BertTokenizer
5 | from collections import OrderedDict
6 | from colorama import Fore, Style
7 | import pickle, re
8 |
9 | import argparse
10 |
11 | def prepare_inputs(context, tokenizer):
12 | '''context
13 | context: I love you. [SEP] Sorry, I hate you.
14 | '''
15 | tokens = tokenizer.tokenize(context)
16 | tokens = tokenizer.convert_tokens_to_ids(tokens)[-hp.max_span+2:]
17 | tokens = [101] + tokens + [102]
18 | # print(f"{Fore.LIGHTBLACK_EX}context:{tokenizer.convert_ids_to_tokens(tokens)}{Style.RESET_ALL}")
19 | tokens = torch.LongTensor(tokens)
20 | tokens = tokens.unsqueeze(0) # (1, T)
21 | tokens = tokens.to("cuda")
22 | return tokens
23 |
24 | def suggest(context, tokenizer, model, idx2phr):
25 | x = prepare_inputs(context, tokenizer)
26 | model.eval()
27 | with torch.no_grad():
28 | _, y_hat, y_hat_prob = model(x)
29 | y_hat = y_hat.cpu().numpy().flatten() # (3)
30 | y_hat_prob = y_hat_prob.cpu().numpy().flatten() # (3)
31 | y_hat_prob = [round(each, 2) for each in y_hat_prob]
32 | preds = [idx2phr.get(h, "None") for h in y_hat]
33 | preds = " | ".join(preds)
34 | print(f"{Fore.RED}{preds}{Style.RESET_ALL}")
35 | print(f"{Fore.GREEN}{y_hat_prob}{Style.RESET_ALL}")
36 |
37 |
38 |
39 | if __name__ == "__main__":
40 | parser = argparse.ArgumentParser()
41 | parser.add_argument("--ckpt", type=str, required=True,
42 | help="checkpoint file path")
43 | args = parser.parse_args()
44 |
45 |
46 | tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
47 |
48 | print("Wait... loading model")
49 | ckpt = args.ckpt
50 |
51 | model = Net(hp.n_classes)
52 | model = model.cuda()
53 | ckpt = torch.load(ckpt)
54 | # model.load_state_dict(ckpt)
55 |
56 | # ckpt = OrderedDict([(k.replace("module.", "").replace("LayerNorm.weight", "LayerNorm.gamma").replace("LayerNorm.bias", "LayerNorm.beta"), v) for k, v in ckpt.items()])
57 | ckpt = OrderedDict([(k.replace("module.", ""), v) for k, v in ckpt.items()])
58 | model.load_state_dict(ckpt)
59 | print("Model loaded.")
60 |
61 | print("# loading dictionaries ..")
62 | idx2phr = pickle.load(open(hp.idx2phr, 'rb'))
63 |
64 | context = ""
65 | print("Let's start a conversation. If you want to start a new one, please press Enter.")
66 | while True:
67 | line = input("A:")
68 | if line == "":
69 | context = ""
70 | print("NEW CONVERSATION---")
71 | continue
72 | else:
73 | context += line + " | "
74 |
75 | suggest(context, tokenizer, model, idx2phr)
76 |
77 | line = input("B:")
78 | if line == "":
79 | context = ""
80 | print("NEW CONVERSATION---")
81 | continue
82 | else:
83 | context += line + " | "
84 |
85 | suggest(context, tokenizer, model, idx2phr)
86 |
87 |
--------------------------------------------------------------------------------
/train.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import torch.nn as nn
3 | import torch.optim as optim
4 | from data_load import get_batch
5 | from hparams import hp
6 | from model import Net
7 | from tqdm import tqdm
8 | import os
9 | import random
10 | from pytorch_pretrained_bert import BertTokenizer
11 | import pickle
12 |
13 | def train_and_eval(model, optimizer, criterion, ids2tokens, idx2phr):
14 | model.train()
15 | for step in tqdm(range(hp.n_train_steps+1)):
16 | x, y = get_batch(hp.max_span, hp.batch_size, hp.n_classes, True)
17 | x = x.cuda()
18 | y = y.cuda()
19 |
20 | optimizer.zero_grad()
21 |
22 | logits, y_hat, _ = model(x) # logits: (N, classes), y_hat: (N,)
23 |
24 | loss = criterion(logits, y)
25 | loss.backward()
26 |
27 | optimizer.step()
28 |
29 | # evaluation
30 | if step and step%500==0: # monitoring
31 | eval(model, f'{hp.logdir}/{step}', ids2tokens, idx2phr)
32 | print(f"step: {step}, loss: {loss.item()}")
33 | model.train()
34 |
35 | def eval(model, f, ids2tokens, idx2phr):
36 | model.eval()
37 |
38 | Y, Y_hat = [], []
39 | with torch.no_grad():
40 | x, y = get_batch(hp.max_span, hp.batch_size, hp.n_classes, False)
41 | x = x.cuda()
42 |
43 | _, y_hat, _ = model(x) # y_hat: (N, n_candidates)
44 |
45 | x = x.cpu().numpy().tolist()
46 | y = y.cpu().numpy().tolist()
47 | y_hat = y_hat.cpu().numpy().tolist()
48 |
49 | Y.extend(y)
50 | Y_hat.extend(y_hat)
51 |
52 | # monitoring
53 | pointer = random.randint(0, len(x)-1)
54 | xx, yy, yy_hat = x[pointer], y[pointer], y_hat[pointer] # one sample
55 |
56 | tokens = ids2tokens(xx) # this is a function.
57 | ctx = " ".join(tokens).replace(" ##", "").split("[PAD]")[0] # bert detokenization
58 | gt = idx2phr[yy] # this is a dict.
59 | ht = " | ".join(idx2phr[each] for each in yy_hat)
60 |
61 | print(f"context: {ctx}")
62 | print(f"ground truth: {gt}")
63 | print(f"predictions: {ht}")
64 |
65 | # calc acc.
66 | n_samples = len(Y)
67 | n_correct = 0
68 | for y, y_hat in zip(Y, Y_hat):
69 | if y in y_hat:
70 | n_correct += 1
71 | acc = n_correct / n_samples
72 | print(f"acc@{hp.n_candidates}: %.2f"%acc)
73 |
74 | acc = str(round(acc, 2))
75 |
76 | torch.save(model.state_dict(), f"{f}_ACC{acc}.pt")
77 |
78 |
79 | if __name__=="__main__":
80 | os.makedirs(hp.logdir, exist_ok=True)
81 |
82 | print("==== Load tokenizer")
83 | tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
84 | ids2tokens = tokenizer.convert_ids_to_tokens
85 |
86 | print("==== Load dictionaries")
87 | idx2phr = pickle.load(open(hp.idx2phr, 'rb'))
88 |
89 | device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
90 |
91 | print("==== Building model")
92 | model = Net(hp.n_classes)
93 | model = model.to(device)
94 | model = nn.DataParallel(model)
95 |
96 | optimizer = optim.Adam(model.parameters(), lr=hp.lr)
97 | criterion = nn.CrossEntropyLoss()
98 |
99 | train_and_eval(model, optimizer, criterion, ids2tokens, idx2phr)
100 |
101 |
102 |
--------------------------------------------------------------------------------