A shallow convolutional neural network architecture for open domain question answering

Abstract

This paper addresses the problem of answering a question by choosing the best answer from a set of candidate text fragments. This task requires to identify and measure the semantical relationship between the question and the candidate answers. Unlike previous solutions to this problem based on deep neural networks with million of parameters, we present a novel convolutional neural network approach that despite having a simple architecture is able to capture the semantical relationships between terms in a generated similarity matrix. The method was systematically evaluated over two different standard data sets. The results show that our approach is competitive with state-of-the-art methods despite having a simpler and efficient architecture.