Attention-based memory network for sentence-level question answering

Abstract

Sentence-level question answering (QA) for news articles is a promising task for social media, whose task is to make machine understand a news article and answer a corresponding question with an answer sentence selected from the news article. Recently, several deep neural networks have been proposed for sentence-level QA. For the best of our knowledge, none of them explicitly use keywords that appear simultaneously in questions and documents. In this paper we introduce the Attention-based Memory Network (Att-MemNN), a new iterative bi-directional attention memory network that predicts answer sentences. It exploits the co-occurrence of keywords among questions and documents as augment inputs of deep neural network and embeds documents and corresponding questions in different way, processing questions with word-level and contextual-level embedding while processing documents only with word-level embedding. Experimental results on the test set of NewsQA show that our model yields great improvement. We also use quantitative and qualitative analysis to show the results intuitively.