Combining Deep Learning With Multiple Features For Answer Selection

In recent years, artificial intelligence is more and more popular in public, as it tries in many fields and achieves some success. It is always the goal of artificial intelligence that humans communicate with machine barrier-freely. And question answering system is a very significant entry point for achieving this vision. There are different kinds of question answering, such as community question answering, question answering based knowledge base and chatterbot. When answering a question, these question answering systems usually have big knowledge of storage, and will first generate a candidate answer set according to their background knowledge,then select one from the candidate set as the best answer after ranking all candidate answers.This paper studies the task of answer selection ranking, which scores each candidate answer and sorts all candidates by their scores. The semantic similarity between the questions and candidate answers is the core issue to score the candidate answers. The previous research on semantic similarity of two sentences include two categories: one is based on lexical features, syntactic features of question and answers to compute semantic similarity. The other is mainly based on deep learning which don't need the external linguistic tools or knowledge base resources. In this paper, we combine deep learning models with lexical features, topic features and alignment relations of the question and answers to solve the task of question answer selction and scoring. The contents of this paper are as follows:1. We extract lexical features with linguistic tools, and topic features with the Biterm Topic Model, and alignment relations with the IBMModel1, which will append to the LSTM.2. We construct the answer selection sort framework based on the CNN and LSTM respectively. Experiment results show that the LSTM performs higher than CNN in our task when no external features are added.3.we combine the extracted features with LSTM. We append the lexical features to the input of LSTM, and join the topic features with the output of LSTM, and add attention mechanism in LSTM based on alignment relations of the question and answers. Experiments show that either topic features or attention mechanism in LSTM, both of them have significant influence on our task.4. When combining topic and alignment features with LSTM together, the MAP and MRR on public datasets reach to 0.79 and 0.80 respectively, which performs higher than previous state-of-the-art works. Thus, these results demonstrate the effectiveness of our proposed models and the feasibility of the methods we used.