Research On Semantic Relevance Calculation For Question Answering

The question answering system is derived from the need for quickly and accurately acquiring information,which is the advanced form of the information retrieval system.Its core is the semantic relevance Calculation between question and answers.The traditional computational semantic relevance methods are mostly based on feature engineering,using syntactic parsers,knowledge bases and other resource libraries,which need high cost.Deep learning technology has been widely used in the fields of computer vision,speech recognition,sentence classification and question answering system because of its advantages such as automatic extracting features and good generalization performance among languages.Because of the good application prospect of question answering system and the efficient computing ability of deep learning,this paper designs an algorithm model which uses the deep learning to calculate the semantic relevance between question and answers.Among the deep learning models,Long-Short Term Memory(LSTM)can "remember" the context information,Convolutional Neural Network(CNN)can extract local features.Therefore this paper combines the advantages of LSTM and CNN,which first uses LSTM to extract the semantic information of QA sentence,and then uses CNN to extract abstract features.In order to compare the sentences semantics more comprehensively,this paper presents a combination of semantic features and statistical features to compute the relevance of QA pairs.The semantic features this paper presented include CED features,projection features and parallel features.Statistical features include overlap features and BM25 features.One can effectively calculate the relevance score between question and answers by combining the two types of features.To keep consistent with the related work,this paper uses the MAP,MRR and Top-1 measures to evaluate the effect of the answer ranking.In the experimental section,this paper use three QA dataset: TrecQA,WikiQA and Insurance QA.I first compare the influence of word vector,model structure,feature selection and optimizer on the experimental results,and select the optimal parameter to test the effect of the datasets.The results show that the neural network in this paper has achieved good results on these three data sets.Combining the semantic features and the statistical features can effectively calculate the relevance between QA pairs.Although the results of this paper can not achieve state-of-the-art performance compared with the related work,the model structure of this paper is simple and easy to implement.It does not require feature engineering and grammar analysis,which is suitable for all language processing,and it has practical significance.