Image credit: Medium

Open-domain Question Answering with End-to-End Memory Networks

Please click on the ‘Video’ and ‘Code’ buttons displayed above to view project source code and model in action!


The task is to build a statistical or deep neural model for Question Answering either utilizing both the provided Wikidata Knowledge Graph and Wikipedia Text Corpus or just a single knowledge source. Any existing methods/libraries could be utilized. In this project, I have implemented an End-to-End Memory Net using Keras. Later, trained the model for 100 epochs on Facebook’s bAbI dataset which has 1000 questions for training and 1000 for test. Also, the dataset has an array of text passages ranging from Single-fact, Multi-fact, Multihop Reasoning to Agent Decision based corpus.

Execution Instructions

  • Clone the repository
  • Install the project dependencies: “pip install -r requirements.txt”
  • In the “” file set the hyperparameters (lstm_size, epochs, batch size, model type) accordingly.
  • Run the model: “python”
  • Assess model performance using the test questions in Test Questions file

Available Models

  1. LSTM
  2. Bi-Directional LSTM
  3. GRU
  4. End-to-End Memory Net
Akshay Joshi
M. Sc. Data Science & Artificial Intelligence