QADialogSystem

We design models that generate conversational responses for factual questions using expert answer phrases from Question Answering systems. Paper: "Fluent Response Generation for Conversational Question Answering"

Setup for CoQA baseline

• Install SMRCToolkit for model code
• Run python run_bert_coqa.py to train the model
• Run python evaluate_on_qa_generation_test_using_coqa.py to get CoQA predictions on SQuAD Dev Test set

Setup for QuAC baseline

• Download and extract QuAC trained model inside the "quac_baseline" folder
• Run quac_baseline.py to extract quac model responses on squad_dev_test

Setup for SQuAD baseline

• Clone bert within "squad_baseline" folder.
• Checkout to specific commit by running git checkout 88a817c37f788702a363ff935fd173b6dc6ac0d6
• Refer to model_training_commands.txt inside "bert" folder for running instructions

STs+BERT baseline

• The outputs of STs+BERT baseline predictions on SQuAD Dev Test set can be found in mturk_evaluations/data2/bert_softmax_predictions_on_squad_dev_test_0_to_822.txt

Setup for Pointer generator models

• run git clone https://github.com/OpenNMT/OpenNMT-py to get "OpenNMT-py" folder within "QADialogSystem".
• checkout to specific commit by running git checkout 7f1fc81da864c465862f23e048802107ada714a8 from within the "OpenNMT-py" folder To get the pretrained models
• cd OpenNMT-py
• Download zip file containing saved PGN and PGN-pre model checkpoints
• unzip pgn_models.zip To re-train the models
• Extract opensub_qa_en data in "Data/Opensubtitles_qa"
• run preprocess_opensubtitles_qa.py in "Data/Opensubtitles_qa" folder to moses tokenize the opensub_qa_en data.
• Follow the training commands in all_final_model_training_and_testing_commands.txt

DialoGPT, GPT-2

Download and extract saved GPT-2, GPT-2-Pre and DGPT models in "DialoGPT" folder as follows:

• cd DialoGPT
• Download zip file containing saved GPT-2, GPT-2-Pre and DGPT model checkpoints fine-tuned on SS and SS+ data
• unzip gpt_and_dgpt_models.zip

For instructions on how to run the models refer to all_final_model_training_and_testing_commands.txt

Citation

@article{baheti2020fluent,
  title={Fluent Response Generation for Conversational Question Answering},
  author={Baheti, Ashutosh and Ritter, Alan and Small, Kevin},
  journal={arXiv preprint arXiv:2005.10464},
  year={2020}
}

TODO

• Add the instructions on how to generate STs + BERT baseline predictions on SQuAD Dev Test set