Research And Implementation Of Conversation Structure Modeling Based On Multi-aspect Semantics

As instant messaging software such as QQ,wechat and Dingding gradually become an important tool for communication,the group chat function is used more and more frequently.It is of great significance to analyze massive group chat messages and capture valuable content.For example,by analyzing hot social topics in group chats,emergencies can be reasonably prevented,and by identifying premeditated crimes,spreading pornography,gambling,drugs and other illegal acts,the police can be reported to reduce social harm.In order to deeply understand the group chat,conversation structure discovery technology came into being,which can automatically identifying the reply relationship between the messages in the group chat,it assists various downstream content analysis tasks.Due to a large number of phenomena such as short length,pronouns,and irregular expressions in messages,it brings great challenges to structural semantic modeling.Most of the existing conversation structure discovery technology only focus on the content-level semantics,ignoring user-level semantics,which can help conversation structure discovery tasks.This paper proposes two conversation structure discovery techniques based on multi-semantic fusion:(1)A conversation structure discovery technique based on the gated recurrent unit.First,the research found that thread content semantic,discourse structure semantic and user attention semantic in group chats have complementary relationships;then based on gated recurrent units,three encoders are designed to capture the three semantics and fuse them to identify the conversation structure.A conversation structure discovery technology based on heterogeneous graph neural network.Firstly,a context encoder based on heterogeneous graph is designed to integrate content semantics and user semantics.Then based on the encoder,tracking the change of dialog state to judge the dialog structure.In addition,based on the "heter-order and homo-geneity" of implicit semantic features in group chat,a contrastive learning method is introduced to enhance the context encoder.The experimental results show that the above two techniques can effectively improve the indicators of precision,recall and F1-value on the task of conversation structure discovery.Finally,this paper applies the above technology to a group chatoriented public opinion monitoring system.By reconstructing the conversation structure,it supports downstream analysis tasks such as topic division and key person identification.The case analysis shows that the technology has important application value.