Research on Answer Selection Based on LSTM
Yangsen Zhang, Yuanyuan Peng · 2018
Answer selection is a critical task in question answering systems. Based on deep learning frameworks, this paper proposes an answer selection method without using external semantic resources and manual features. Firstly, the word vector is used to represent a question and its candidate answers. Secondly, deep learning frameworks are used to calculate semantic similarities between them. Finally, the answer which has the highest semantic similarity is selected as the question's right answer. In this method, we construct two semantic similarity calculation models, one of which is based on bidirectional long short-term memory (BLSTM), and the other applies attention mechanism to enhance the representation of the semantic information which is related to the question. Experimental results on the public Data Set InsuranceQA show that the semantic similarity calculation model based on attention mechanism has achieved a good performance, and the proposed method can reduce the impacts of the similarity evaluation methods and the imbalanced data.