| With the rapid development of artificial intelligence,the dialogue systems have gradually entered the people's life.The dialogue system is a system that can communicate with people through natural language,such as Xiaobing developed by Microsoft,Siri launched by Apple Corp and so on.So,they can give us a lot of information and further help us in a much more comprehensible way.The multi-turn dialogue system is a kind of dialogue system who can continuously respond according to the previous conversation content and it's affecting people's lives.Compared with ordinary dialogue systems,through multi-turn communication,they provide us more convenient services and more useful information.In recent years,the dialogue systems have begun to evolve from previous search-based dialogue systems to neural-network-based generative dialogue systems.Although the performance of multi-turn dialogue systems has been greatly improved,there are still many problems to be solved.For example,some existing dialogue systems generate sentences containing grammatical errors or semantic ambiguities,and some dialogue systems produce perfunctory useless statements.In view of the problems mentioned above in current multi-turn dialogue systems,we carried on in-depth studies in three aspects,which are sentence perplexity evaluation,topic-word correlation and sentence generation.The main work and innovations of this paper are as follows.(1)A method for calculating the sentence perplexity is proposed through both the semantic and the grammar analysis of sentencesFirstly,we remove all the modifiers from an original sentence and thus extract its skeleton.Then,by analyzing the sentence dependency of the sentence skeleton and the original sentence,we get the semantic and grammatical information,which is used to calculate the perplexity of the original sentence relative to the reference sentence.Finally,we obtain the total perplexity of the whole sentence by the weighted calculation of the various perplexities.This method can be utilized to calculate the perplexity of the sentences generated by the dialogue system and further reduce the error rate of sentence generation.(2)A method of topic transfer detection in multi-turn dialogue systems based on word correlation is proposedWe first clarify the differences between the word correlation and the word similarity.Then,we propose a method of calculating the correlation in the word vector space.In the end,we use this method to detect the topic transfer in the dialogue process of multi-turn dialogue systems.This method can effectively discover the changes of the topic in the process of dialogue,which lays the foundation for the generation of appropriate response sentences.(3)A method of response generation for multi-turn dialogue systems based on topic correlation constraint is presentedFirst,we maintain a dynamic topic list during the dialogue process in a multi-turn dialogue system,which keeps topic information of the current dialogue.Then,we use the information in the dynamic topic word list to constrain the topic of sentences generated by the dialogue system,and construct the neural network model to generate statements.This approach enables the system to generate more contextual sentences,and thus,improves the quality of the generated responses and makes the dialogue process smoother. |
