The goal of the project is to build a model that understands a set of factual sentences to be able to answer questions based on the facts. For example,
Mary is in the bathroom. Mary moved to the Bedroom. John traveled to the hallway. Where is Mary?
The model should know that Mary is in the Bedroom.
Data scientists at Facebook reported an impressive work using a special kind of recurrent neural networks (RNN) called long short-term memory (LSTM) here.
Later, Stephen Merity used Keras deep learning library to produce this. His success is due to the clever idea of his which is splitting the factual statements from the question and merging them back together at the end. Can I combine the two clever ideas with the use of pre-trained GloVe word vectors for word ebedding to train a general question answering system? It turns out combining two great ideas results in a better performance.
The dataset I trained my model comes from facebook's bAbI project.
I completed this project in just over two weeks time, so not everything is as perfect as I would have liked it to be.