On the surface, the Question-Answering task seems straightforward—given a question and (optionally) some related facts, the model must produce an answer.

Traditional approaches to QA include rule-based models or information retrieval methods based on word overlap or tf-idf scores. However, training a model to understand the input as well as the facts in terms of both syntax and semantics can be challenging due to the inherent complications of natural language. Deep neural networks can learn to model these complexities without handcrafting or feature engineering, and have emerged as the state-of-the-art networks for these tasks.