End-to-end memory network implementation for Question answering

This is an implementation of a chatbot that can answer questions based on a "story" given to the bot.

Dataset

• Babi dataset released by facebook research.
• A particular subset of the dataset which has stories, questions and answers is used as data (One of 20 tasks in the bAbI project).
• Training set(10000) and test set(1000) are seperated and each sample is in a tuple format (story,question,answer)

Model

a) Single Layer case which implements a single memory hop operation

b) Multiple Layer implementation (using RNNs) with multiple hops in memory

3 main sub-components of network:

• Input Memory Representation
• Output Memory Representation
• Generating Final Prediction

Full model : Using LSTMs with multiple layers on top of sub-components. Network produces a probabilty for every single word in the vocabulary. In this implementation, there will be high probablity on either yes or no.

Code accompanying the End-To-End Memory Networks paper: https://arxiv.org/pdf/1503.08895.pdf

For further understanding : https://www.youtube.com/watch?v=ZwvWY9Yy76Q