├── README.md
└── prompt_source.pth


/README.md:
--------------------------------------------------------------------------------
 1 | # Transfer-Prompts-for-Text-Generation
 2 | 
 3 | This is the repository for NAACL 2022 paper "Learning to Transfer Prompts for Text Generation". The implementation is completely based on our newly-developed text generation library **[TextBox 2.0](https://github.com/RUCAIBox/TextBox)**
 4 | 
 5 | ## Prompt Source
 6 | The `prompt_source.pth` in this repository contains the source task prompts (*i.e.*, tensors of shape `[200,1024]`) trained on 14 datasets as introduced in our paper:
 7 | - CNN/Daily Mail (cmmdn)
 8 | - XSum (xsum)
 9 | - MSNews (msn)
10 | - Multi-News (mn)
11 | - NEWSROOM (nr)
12 | - SQuAD (squad)
13 | - Wiki Neutrality (wiki)
14 | - Quora (quora)
15 | - PersonaChat (pc)
16 | - DailyDialog (dd)
17 | - DSTC7-AVSD (da)
18 | - MultiWOZ (mw)
19 | - WritingPrompts (wp)
20 | 
21 | You can also download these datasets [here](https://github.com/RUCAIBox/TextBox#dataset).
22 | 
23 | ## Installation
24 | First, you should clone the TextBox repository and follow its [instructions](https://github.com/RUCAIBox/TextBox#installation).
25 | 
26 | Then, you may copy the `prompt_source.pth` into the TextBox folder (*i.e.*, \<your clone dir\>/TextBox).
27 | 
28 | ## Running PTG based on TextBox
29 | For example, you can conduct our cross-dataset experiments on cnndm dataset using this command:
30 | ```bash
31 | python run_textbox.py --model=PTG --dataset=cnndm --model_path=facebook/bart-large
32 | ```
33 | In this default case, the source tasks (datasets) is msn, mn, and nr.
34 | 
35 | You can use `--dataset=xxx` to specify the dataset name.
36 | 
37 | In addition, you can also specify the source tasks using `--source_task=list_of_task`. The default setting is equivalent to `--source_task=\[\'msn\',\'mn\',\'nr\'\]`.
38 | 
39 | We also provide several cases used in our paper:
40 | - `--source_task=cross-dataset2`: tc, da, mw.
41 | - `--source_task=cross-task1`: squad, wiki, quora, wp, cnndm.
42 | - `--source_task=cross-task2`: squad, wiki, quora, wp, pc.
43 | 
44 | 
45 | 
46 | ## Reference
47 | ```bibtex
48 | @inproceedings{li-etal-2022-learning-transfer,
49 |     title = "Learning to Transfer Prompts for Text Generation",
50 |     author = "Li, Junyi  and
51 |       Tang, Tianyi  and
52 |       Nie, Jian-Yun  and
53 |       Wen, Ji-Rong  and
54 |       Zhao, Xin",
55 |     booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
56 |     month = jul,
57 |     year = "2022",
58 |     address = "Seattle, United States",
59 |     publisher = "Association for Computational Linguistics",
60 |     url = "https://aclanthology.org/2022.naacl-main.257",
61 |     doi = "10.18653/v1/2022.naacl-main.257",
62 |     pages = "3506--3518",
63 | }
64 | ```
65 | 


--------------------------------------------------------------------------------
/prompt_source.pth:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/RUCAIBox/Transfer-Prompts-for-Text-Generation/b1ab5358e2e1bb70d243a0f811e641b215cb2808/prompt_source.pth


--------------------------------------------------------------------------------