MultihopQA

This project is part of course "Natural Language Processing Application" from IIIT

Multi-hop QA where answering to question requires reasoning and aggregation across several paragraphs. Most of the existing QA problems are confined to finding the answers from a single paragraph (single-hop). For example, in SQuAD questions are designed to be answered given a single paragraph as the context. MultihopQA problem not only seeks to provide correct answers but also requires explanation through selection of supporting facts.

Requirements

install required packages using anaconda

conda env create -n environment.yml

Data Download and Preprocessing

Run the script to download the data, including HotpotQA data and GloVe embeddings, as well as spacy packages.

./download.sh

Model Descriptions

Baseline Model (HotPotQA)

As described in HotPotQA paper, the baseline model was implemented by augmenting the model presented simple multiple-paragraph RC. They have added an extra RNN layer combined with a binary classifier for classifying supporting sentences from the paragraph.

preprocessing

python main.py --mode prepro --data_file hotpot_train_v1.1.json --para_limit 2250 --data_split train
python main.py --mode prepro --data_file hotpot_dev_distractor_v1.json --para_limit 2250 --data_split dev
python main.py --mode prepro --data_file hotpot_dev_fullwiki_v1.json --data_split dev --fullwiki --para_limit 2250

training

python main.py --mode train --para_limit 2250 --batch_size 24 --init_lr 0.1 --keep_prob 1.0 --sp_lambda 1.0

results

setting	split	answer EM	answer F1	sup fact EM	sup fact F1	joint EM	joint F1
distractor	dev	41.8	55.8	17.5	61.3	8.7	36.3
Full Wiki	dev	16.0	24.0	1.78	21.9	0.87	7.60

Model weights

The weights of models implemented in this repository can be found here