Methods Of Sentence Relation Recognition For Interactive Question Answering

With the development of Internet technology and the rapid growth of information scale,people urgently need an accurate and efficient way to obtain information.From search engines to intelligent interactive question answering systems,the ways of searching information become more and more close to the natural interaction.Because of the emergence of massive Internet data and the progress of machine learning methods and natural language processing technology,the QA system for free text and heterogeneous information occurs in some fields,and it aims to be an intelligent system based on generative models.Different from search engines,the question answering system can understand the problem described in the form of natural languages better,and the system returns concise and accurate answers instead of some related documents.With the occurrences of Siri and Watson,the intelligent interactive question answering system has become a hot research topic in recent years.In the business field,it is also more and more potential to replace manual customer service.However,to construct an intelligent interactive QA system,it is very important to learn knowledge from the real customer service logs.And how to recognize the matching relations between questions and answers and the completion relations between continuous sentences from those complex interactive logs has become the key to construct the learning system.This paper mainly studies the question answer matching relation recognition and the completion relation recognition in the interactive question answering.For question answer matching,this paper constructs two kinds of models including the semantic matching model based on CNN and the generative model based on RNN.The word vector matrices of sentences are feed into the input layer of models,and the output are the question answer matching confidence.The comparative experiments of different deep learning models are carried out based on the Semeval-2016 community question answering data and the online customer service dialogue data.Furthermore,this paper analyses the experiment results based on the completeness of the questions,different structures of the generative model,the threshold selection and the methods of data extraction.It was found that on the community question answering data,the CNN model performs better than the RNN model.On the customer service dialogue data,the RNN based sequence to sequence model can learn scene information from the dialogues better.In the one round based data with a complete question,MAP reaches 84.41%.In view of the potential semantic completion relations between continuous sentences in the interactive questions answering,this paper studies the recognition of sentence completion relations.The paper constructs the parallel convolutional neural networks and series LSTM to extract high-level semantic features and model the sentence pairs.And this paper uses the Support Vector Machine,the CNN based model and the RNN based model to recognize the completion relations between sentences respectively.Experimental results show that the CNN based method outperforms other comparative methods and achieves the best F1 value of 67.8%.Finally,the question completion relation recognition and the matching relation recognition are combined to be applied to the interactive question answer semantic matching.