Research On Question Analysis And Answer Extraction Methods Of Chinese Question Answering Systems

With the rapid development of the Internet and the rise of natural language processing technology,question answering systems have entered a period of development oriented towards open fields and based on free text data.How to get useful information from amount of information is a hot spot in academic and industry.Compared with traditional search engines,question answering systems allow users to ask questions in natural language and better meet the user's need for fast,efficient and accurate getting information.This paper further analyzes question answering systems from two aspects: question analysis and answer extraction.In question analysis,this paper proposed a dynamic topic model—CTM,which is based on Extended Latent Dirichlet Allocation(Extended LDA)and Incremental Biterm Topic Model(IBTM)from the perspectives of long text,short text and seriality of data.The model not only captures the semantic information in the user's question texts,but also captures the word pairs information in the sliding window,classifies the real-time data and analyzes the user's intentions.Due to the error generated by Chinese word segmentation tools and the flexibility of Chinese grammar,traditional methods of extracting keywords from Chinese text does not completely capture the head words in the user's questions.We use the combination of part-of-speech taggings and thesaurus to extend keywords.Experiment results show that the two methods with our dynamic topic model help to mine the head words in user questions.In terms of answer extraction,this paper we studied is similar to the answer selection and rank answers.In order to maintain coherence in context,we still use the answer extraction to represent the answer selection and rank answers.Due to the current neural network-based answer extraction models do not fully consider the intrinsic relationship between questions and answers,we propose a Bi-directional Long Short Term Memory Network(Bi-LSTM)-based answer extraction model.The model directly learns question vector with the Bi-LSTM,and further extracts the question features by the Convolutional Neural Network(CNN).Then it uses the attention mechanism to weight answer vector according to the correlation between the representation vector of the answer and the final representation vector of the question.Finally,the model effectively captures therelationship between questions and answers.The experimental results show that our proposed Bi-LSTM-based answer extraction model has achieved good experimental results on the Q&A dataset based on search engines.