Consistencies Of Personal Pronouns And Focal Points In Conversation Systems

Conversation systems stand significantly in natural language processing.With the development of artificial intelligence,conversation systems begin to be applied to various areas.What's more,it is a long-standing goal to build an intelligent conversation system.The study of conversation systems can be divided into two categories: task-oriented dialogue systems and open-domain conversation systems.In this paper,we adopt the paradigm of editing retrieved responses and complete two works.These two works are about two different problems in open-domain conversation systems.(1)The problem of inconsistent personal pronouns in conversation systems.Retrievalbased methods use retrieval models to select responses from pre-defined repositories.With the differences between input contexts and the prepared conversations increase,these methods tend to return more and more inappropriate responses.According to our observations,there are a number of inappropriate responses suffering from inconsistent personal pronouns.Responses with inconsistent personal pronouns may be not understandable or not fluent.It may have negative effects to user experience.To solve this problem,we propose a Sequential Editing Model(SEM).SEM can edit retrieved responses based on conversation contexts to correct inconsistent personal pronouns.Besides,we further propose two extensions of SEM.Our experiment results show that SEM significantly outperforms existing editing models in editing inconsistent personal pronouns.More importantly,the two extensions can further improve the performance of SEM.(2)The problem of out of conversation focal point in conversation systems.In human conversations,conversation focal points play an important role in keeping the relevance between responses and their contexts.Currently,researchers propose an editing model to edit retrieved responses so that their model can generate more informative responses.However,focal points of retrieved responses may be not completely same with conversation contexts and the existing editing model does not have designs to identify and consider the focal points in conversation contexts.Thus,this editing model can not correct the focal points and tend to generate out of focal point responses.These responses are always unrelated to the contexts.To solve this problem,we propose an Autofocal Response Editing(ARE)model.ARE can identify the focal points of conversation contexts and edit retrieved responses based on the points.When identifying conversation focal points,we propose and use a self-constraint mechanism to replace the widely used self-attention mechanism.The experiment results show that the self-constraint mechanism can help ARE to identify focal points more accurately.What's more,ARE significantly outperforms mainstream models in various aspects.