Question Answering through Transfer Learning

• This is the original implementation of "Question Answering through Transfer Learning from Large Fine-grained Supervision Data". [paper] [poster]
• Most parts were adapted & modified from "Bi-directional Attention Flow". [paper] [code]
• Evaluation scripts for SemEval were adapted & modified from SemEval-2016 official scorer.
• Please contact Sewon Min (email) for questions and suggestions.

Codes include

• pretraining BiDAF (span-level QA) and BiDAF-T (sentence-level QA) on SQuAD
• training on WikiQA
• training on SemEval-2016 (Task 3A)

0. Requirements

General

• Python3 (verified on 3.5.2.)
• Python2 (verified on 2.7.12., only for Semeval-2016 Scorer)
• unzip, wget (for running download.sh only)

Python Packages

• tensorflow (deep learning library, only works on r0.11)
• nltk (NLP tools, verified on 3.2.1)
• tqdm (progress bar, verified on 4.7.4)
• jinja2 (for visaulization; if you only train and test, not needed)

1. Quick Tutorial

First, download data (SQuAD, WikiQA, SemEval-2016, GLoVe, NLTK). This will download files to $HOME/data. Also, preprocess data and save them in data.

chmod +x download.sh; ./download.sh
chmod +x prepro.sh; ./prepro.sh

Then, pretrain the model on SQuAD.

chmod +x pretrain.sh
./pretrain.sh span 		# to pretrain BiDAF on SQuAD
./pretrain.sh class		# to pretrain BiDAF-T on SQuAD-T

You can use trained model from original BiDAF code. Just place saved directory to out/squad/basic/00.

Finetune the model on WikiQA / Semeval.

chmod +x train.sh; ./train.sh DATA finetune RUN_ID PRETR_FROM STEP

• DATA: [wikiqa | semeval]
• RUN_ID: run id for finetuning. use unique run id for the same data.
• PRETR_FROM: [basic | basic-class]. use basic for span-level pretrained model, and basic-class for class-level pretrained model.
• STEP: global step of pretrained data. For a quick start, use 18000 for span-level pretrained model and 34000 for class-level pretrained model. However, monitoring tensorboard and pick the best global step is recommended, because results would depend much on the quality of pretrained model.

Finally, evaluate your model.

chmod +x evaluate.sh; ./evaluate.sh DATA RUN_ID START_STEP END_STEP

• DATA: [wikiqa | semeval]
• RUN_ID: run_id you used for finetuning
• START_STEP: STEP+200 when you used for finetuning
• END_STEP: STEP+5000

This is just for a quick tutorial. Please take a look at run.md for details about running the code.