manoj98 / improving-question-answering-with-external-context-fusion

Question answering system based on QA-GNN architecture with utilization of external context. QA-GNN involves using language models(LMs) to encode QA context and retrieve a knowledge subgraph (KG) from the QA entities. The joint representation of QA context and KG sub graph is passed through graph neural network(GNN) module for reasoning.

Question-Answering-using-QA-GNN-with-external-context-fusion

The project was completed as part of the CSE6240 course at Georgia Tech. It was completed while working with Palash Choudhary, Sanchita Porwal, and Shreyas Verma.

The final project report can be found here: Link

The final project presentation can be found here: Link

The result of the project is a question-answering system based on a QA-GNN architecture with the utilization of external context. QA-GNN involves using language models (LMs) to encode QA context and retrieve a knowledge subgraph (KG) from the QA entities. The joint representation of the QA context and the KG subgraph is passed through the graph neural network (GNN) module for reasoning. The final model incorporates language understanding through LMs, structured reasoning through KGs, and external context on LM entities through knowledge facts.

APPROACH 1 : LM-ONLY QA MODEL

Fig 1: LM Architecture

This approach only utilizes LM models for question-answering problems.

• Run LMQA_final.ipynb

APPROACH 2 : QAGNN [LM+KG QA MODEL]

Fig 2: QA-GNN Architecture

This approach implements the QA-GNN model for question-answering problems.

Step 1 : COMPUTE USED

2 RTX-6000 GPUs [Cuda 11.7.0]

Step 2 : CREATE ENVIRONMENT**

conda create -n qagnn python=3.7 source activate qagnn pip install torch -f https://download.pytorch.org/whl/torch_stable.html pip install transformers==3.4.0 pip install nltk spacy==2.1.6 python -m spacy download en pip install torch-sparse torch-geometric torch-scatter -f https://pytorch-geometric.com/whl/torch-{torch.__version__}.html

Step 4 : DOWNLOAD RAW DATA

./ download_raw_data.sh

Step 3 : CREATE PREPROCESSED DATA

python preprocess.py

Step 4 : RUN BASH FILE

./ run_qagnn__obqa.sh

• Check saved_model folder for epoch-wise predictions on test and metric csv
• Check logs/ for training status.

Results

Below are few examples highlighting the impact of incorporating external context (knowledge facts) in question-answering task. The final model is a QA-GNN with additional external context.

Fig 3: Sample Question 1 deep-dive

Fig 4: Sample Question 2 deep-dive