Emotion Recognition On Multi-Turn Dialogue Text

Emotion is usually defined as the individual's psychological state related to subjective cognition,emotion and behavior.Therefore,emotion recognition is a very important part of the artificial intelligence.Emotion recognition in multi-turn conversation is receiving more and more attention recently,and it has become a relatively new research direction in the field of natural language processing.There are more and more conversation data in current social media platforms.If you can dig into the emotional perspectives contained in these conversations,you can analyze the current trend of public opinion.In addition,emotion recognition in conversation can also be used as a secondary tool for psychological analysis in medical systems.Another important application is that in the chat bot,it can help generate emotionalrelated answers by analyzing the user's emotional information.However,there are many difficulties in emotional recognition in the dialogue.For example,fine-grained emotion classification requires more accurate classification,the mining of semantic information in conversation context,the continuation or shift of emotions,the reasons for emotion in context and the interlocutor's speaking habits are different.The focus of this paper is to solve these problems by analyzing these difficulties and building a reasonable model.The research content of this paper is mainly summarized into the following three points:(1)Emotion recognition of dialogue text based on feature engineering.Several different machine learning classifiers are selected,and the features of text,emotion and vocabulary are extracted based on the emotion-related influence in the text.Experiments were carried out with feature selection and feature combination,and the effects of different features on emotion recognition were analyzed.(2)Emotion recognition in conversation based on deep learning.Emotion recognition in conversation can be divided into emotional semantic encoding of a single round of dialogue and context extraction in multiple turns of dialogue.Recurrent neural network(RNN),convolutional neural network(CNN)and Transformer were used to study the semantic extraction of texts.In the part of the context information extraction,three different collaboration parts are utilized to extract the emotion coherence and transformation,the speaker state and the deep semantic information in the context information.(3)Emotion recognition in conversation based on pre-training and transfer learning.We overcome the shortcomings of insufficient dialogue corpus are not enough to capture information by training context-sensitive word vector representations,emoji vector representations,and emotional vector representations on other large-scale corpora.We also utilized method of transfer learning to learn the emotional representation of text in other corpora,so that the emotional semantic mining in the multi-turn dialogue in the target task can be completed faster and better.