[COL 764 Project] Knowledge-based Question Answering using Link Prediction technique

Overview of the Approach Used

Environment Requirements

Python 3.7, PyTorch 1.6

Set data file paths and folder

1. For main_train_vocab.py, provide the following paths in line 124-128.

datasetName = "MetaQA"			#dataset folder name
data_path = "./Data/MetaQA/1-hop/vanilla/qa_train.txt"			# training QA pairs path
valid_data_path = "./Data/MetaQA/1-hop/vanilla/qa_dev.txt"		# validation QA pairs path
kbPath ="./Data/MetaQA/kb.txt"									# KG path

For main_train_roberta.py, provide the following paths in line 166-171.

datasetName = "MetaQA"			#dataset folder name
data_path = "./Data/MetaQA/1-hop/vanilla/qa_train.txt"			# training QA pairs path
valid_data_path = "./Data/MetaQA/1-hop/vanilla/qa_dev.txt"		# validation QA pairs path
test_data_path = "./Data/MetaQA/1-hop/vanilla/qa_test.txt"		# test QA pairs path (only to store RoBERTa embeddings beforehand)
kbPath ="./Data/MetaQA/kb.txt"									# KG path

2. For main_test_vocab.py, provide the following paths in line 129-132.

datasetName = "MetaQA"			#dataset folder name
data_path = "./Data/MetaQA/1-hop/vanilla/qa_train.txt"			# training QA pairs path (Used just to calculate vocab size)
test_data_path = "./Data/MetaQA/1-hop/vanilla/qa_test.txt"		# test QA pairs path

For main_test_roberta.py, provide the following paths in line 131-133.

datasetName = "MetaQA"			#dataset folder name
test_data_path = "./Data/PQ-2/test_q.txt" 						# test QA pairs path

3. The Models folder should be available before running the code.

Instructions to run code

1. pip install -r requirements.txt 
2. To train model using LSTM + RoBERTa: python main_train_roberta.py 
3. To train model using only LSTM: python main_train_vocab.py 
4. To test the model using LSTM + RoBERTa: python main_test_roberta.py
5. To test the model using only LSTM: python main_test_vocab.py

References

1. Improving Multi-hop Question Answering over Knowledge Graphs using Knowledge Base Embeddings (ACL 2020).
2. Complex Embeddings for Simple Link Prediction (ICML 2016).
3. An Interpretable Reasoning Network for Multi-Relation Question Answering (COLING 2018)