Skip to content

Latest commit

 

History

History
52 lines (43 loc) · 4.67 KB

README.md

File metadata and controls

52 lines (43 loc) · 4.67 KB
Raw

GenAug: Data Augmentation for Finetuning Text Generators

Code for GenAug, presented in GenAug: Data Augmentation for Finetuning Text Generators published at EMNLP 2020 DeeLIO Workshop. You can cite it as follows:

@inproceedings{feng-etal-2020-genaug,
    title = "{G}en{A}ug: Data Augmentation for Finetuning Text Generators",
    author = "Feng, Steven Y. and Gangal, Varun and Kang, Dongyeop and Mitamura, Teruko and Hovy, Eduard",
    booktitle = "Proceedings of Deep Learning Inside Out (DeeLIO): The First Workshop on Knowledge Extraction and Integration for Deep Learning Architectures",
    month = nov, year = "2020", address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.deelio-1.4",
    doi = "10.18653/v1/2020.deelio-1.4", pages = "29--42",
}

Authors: Steven Y. Feng, Varun Gangal, Dongyeop Kang, Teruko Mitamura, Eduard Hovy

Talk can be found here. Slides and other resources can be found here.

Note: inquiries should be directed to stevenyfeng@gmail.com or by opening an issue here.

drawing

Required Resources

  1. Stanford POS Tagger: https://nlp.stanford.edu/software/stanford-postagger-2018-10-16.zip
  2. Stanford CoreNLP: http://nlp.stanford.edu/software/stanford-corenlp-full-2018-10-05.zip

Pretrained/finetuned Models (on Yelp):

  1. BERT Sentiment Regressor (Finetuned on YLR reviews with star ratings): https://drive.google.com/drive/folders/1JT07ZPxmMO9my5hH3MvJf8VmAlzuuUGf?usp=sharing
  2. GPT-2 (Finetuned on 2 million Yelp reviews - for perplexity and SLOR evaluation): https://drive.google.com/drive/folders/1J3Jcw-qtdWxCYZV7LOnLjKVJZPXiFS2h?usp=sharing
  3. SMERTI-Transformer (Trained on a subset of YLR): https://drive.google.com/drive/folders/1A-jyNp5So4lmv3ZtgKmwq7be8CoFtF_B?usp=sharing

Data

  • "stopwords.txt" contains a list of stopwords used.
  • "yelp_train.txt", "yelp_val.txt", and "yelp_test.txt" contain the entire training, validation, and testing examples for YLR (both prompt and continuation portions).
  • "yelp_train_0.5.txt" and "yelp_test_0.5.txt" contain first-half versions of the training and testing examples of YLR (prompt portions only).
  • "SMERTI_chosen_REs.txt" contains the chosen 150 nouns as Replacement Entities for SMERTI inference.
  • "SMERTI_train.csv" and "SMERTI_val.csv" contain the training and validation splits for SMERTI-Transformer, respectively.

Code

  • "augmentation_methods" folder contains code for the various GenAug augmentation methods (excluding SMERTI).
  • "evaluation" folder contains code for the evaluation. Specifically, "get_unigram_distribution.py" is used for the rare_words (RWords) metric, and "evaluate_combined.py" contains code for SBLEU, UTR, TTR, and RWords. The "sentiment" sub-folder contains code for Sentiment Consistency evaluation [MORE DETAILS LATER].
  • "finetuning_and_generation" folder contains scripts and code for finetuning GPT-2 models and generating outputs from GPT-2 models. Run the two .sh scripts to finetune and generate.
  • "processing_and_setup" folder contains code to process and setup the data required for the experiments. "Continuation_Postprocessor.py" cleans the GPT-2 outputs (e.g. stripping trailing exclamation marks), "GenAug_Finetuning_Setup.ipynb" contains code to setup the final augmentation data for GPT-2 finetuning, "GenAug_SMERTI_Processing.ipynb" contains code to setup SMERTI training, validation, and inference data, and "GenAug_Yelp_Dataset_Processing.ipynb" contains code to process and setup the Yelp Reviews data for YLR and 2 million reviews subset for PPL and SLOR evaluation.

SMERTI Augmentation Method Code

Code for the SMERTI augmentation method can be found in the "GenAug SMERTI-Transformer" folder at this repo. This is the official repo for "SMERTI for Semantic Text Exchange" presented in Keep Calm and Switch On! Preserving Sentiment and Fluency in Semantic Text Exchange published at EMNLP-IJCNLP 2019.

Note: more details and example commands for all the code will be added at a later date.