Question Answering model

Tran Le, Stella (Seoyeon) Lee Grinnell College CSC 395 Information Retrieval

This repository contains the implementation of question-answering model which primary goals to understand sequential context.

Dependency

sys os pickle csv

numpy time re

tensorflow 2.0.0

tensorflow.keras

Github repository structure

script

This folder contains fetch_glove.sh and get_bAbi.sh which is used to download GloVe pre-trained word embeddings and Facebook bAbi dataset.

src

1. preprocessing contains python files to preprocess bAbi dataset
2. baseline_model contains python files to train/test lstm models
3. model contains python files of custom keras layers and models
4. train.py is a python file used for training our custom QA models

---

preprocessing

1. get_glove.py contains functions to load GloVe embeddings and embedding matrix
   • load_vectors(filepath): Load GloVe text file to dictionary of {word:embedding}
   • load_embedding_matrix(data_folder, dims): Load embedding vectors to an embedding matrix and create word-index mapping
2. process_bAbi.py contains functions to save each of tasks into Context, Question, and Answer
3. preprocessing.py
   • transform(input, max_len, tokenizer) : Returns padded sequence that is transformed from input by tokenizer.texts_to_sequences
   • main(dim, embedding_folder, data_folder) saves the embedding matrix and tokenizer

baseline_model

1. lstm_train.py and lstm_test.py is implementation for LSTM model

2. lstm_glove_train.py and lstm_glove_test.py is implementation for LSTM model using GloVe pre-trained word embeddings

   

model

1. AttentionModel contains our QA Attention, Attention+Mem models implementation

   model2.py contains

   • AttentionModel2

     

   • AttentionModel3

     

2. InputLayer.py contains PretrainedEmbedding(Layer) that is used for our custom models used for GloVe embedding layers