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"
- 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
- Download and extract QuAC trained model inside the
"quac_baseline"
folder - Run
quac_baseline.py
to extract quac model responses onsquad_dev_test
- 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
- 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
- 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 theopensub_qa_en
data. - Follow the training commands in
all_final_model_training_and_testing_commands.txt
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
@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}
}
- Add the instructions on how to generate STs + BERT baseline predictions on SQuAD Dev Test set