Training a QA model using Facebook's Babi Dataset for Single Supporting Fact.

• Read and parse the stories in the bAbi task format from the downloaded file.
• Vectorize the Story, Query and Answer and add padding.
• Embedd the Input Sequence and Question into a sequence of vectors.
• After all RNN is used.
• RMSProp optimizer is used and loss function is categorical_crossentropy.
• Model is fit on the training and validation context, question and answers.
• We pass in plot callbacks to see how training cost and validation cost varies as the number of epochs increases.
• Model is saved and then loaded and our test examples are vectorized and passed to the model to predict the answer for the question.

Below is an example:

Story: Daniel went back to the hallway . Mary went back to the office .

Question: Where is Mary ?

Original Ans: office

Predicted Ans: office

References: https://towardsdatascience.com/building-an-ai-chat-bot-e3a05aa3e75f