Reasearch On Coherence Evaluation In The Multi-turn Dialogue Threads Base On Community Forums

Recently,dialogue System has attracted more and more attention because of its potential huge commercial value.With the accumulation of data and the development of machine learning and deep learning technologies,it has become possible to create an automatic dialogue system for a chat partner.In daily life,there are many demands for open-domain dialogue systems,such as the emotional companionship of empty nest elderly,the acquisition of hot news events,and even helping parents complete target of communication and education with their children.Although there have been many applications of open domain dialogue systems,such as yellow chicken chat-bot,Microsoft Ice,etc.,the effect is not satisfactory.Most users use them in an early adopter attitude rather than actual using them.So the open domain dialogue system still has huge room for improvement in the future.The existing research work tends to be more of a single-turn dialogue,but the actual dialogue usually has a corresponding context and needs to take into account the dialogue history.Therefore,this article will start from multi-turn dialogues and study the technologies which are needed for the open-domain dialogue system.The content of this article is summarized in the following three points:(1)Coherence of multi-turn dialogue.Using deep learning methods to build dialogue systems requires large-scale dialogue corpus.Therefore,it is a crucial task to extract large-scale and semantically coherent multi-turn conversations from online forums.In this paper,we uses deep learning methods to train and evaluate the coherence of the multi-turn conversations in order to obtain high quality multi-turn conversation corpus.(2)Research on multi-turn conversation semantic matching model.Considering that the evaluation of multi-turn conversation involves more complex factors,we have further discussed the semantic matching of the responses in the multi-turn conversations and we hope to use the deep learning model to evaluate the property of response in a certain context,whether the response matches the context.This is also one of the key technologies for constructing a retrieval-based multi-turn dialogue system.This paper conducts experiments on different semantic similarity calculation methods,as well as representation-based and interaction-based methods,and attempts to extract the semantic features of different granularities in multi-turn conversation through the deep learning model,and finally obtains a good result in the training set and test set we used.(3)The construction of multi-turn dialogue system.In the last chapter of this paper,the construction of multi-turn dialogue system is studied.Based on the content of the first two chapters of this paper,a simple open-domain dialogue system dialogue system is constructed.The actual operation has got good results.