Research On Emotion Recognition In Conversation Based On Deep Learning

With the development of human-machine dialogue technology,we expect that the dialogue system can not only complete specific interactive tasks,but also be capable of modeling and analyzing emotion in conversation,which makes the interactive process more humanized.Thus,the dialogue system is required to understand emotion of human and make a reasonable response.To meet this need,Emotion Recognition in Conversation(ERC),as a basic research in human-machine dialogue system,is dedicated to identify the emotions conveyed in conversation.Since ERC is helpful for the in-depth emotional communication between human and machine,it is of great significance in research.There are following difficulties in ERC:(1)in different conversation contexts,the same dialogue text may contain different emotions.(2)Deep learning method requires a large amount of training data,but there is only a small amount of annotated data available.In view of the above problems,this paper explores how to combine the conversational context with conversational text,while introducing additional data to better complete the ERC task.The main work of this paper is as follows:(1)An ERC model based on multi-task learning(ERC-MTL)is proposed.We argue that the conversational context is of great importance to emotion recognition,moreover the emotional connection between adjacent dialogue sentences is particularly important.Based on this point of view,this paper introduces a sub-task named Judging the Emotional Consistency in Conversation(JECD).JECD is utilized to analyze whether the emotion shifts between the previous dialogue text and the current dialogue text.The model proposed in this paper adopts the multi-task learning method to train ERC and JECD tasks simultaneously by combining loss functions.The model can synthesize the contextual information to learn the possible emotional shifts,thus increasing the accuracy of emotion recognition.Finally,experimental results show that the ERC-MTL model has achieved significant improvement in ERC.(2)A method based on transfer learning(ERC-TL)is proposed.In order to alleviate the problem of data sparse,two transfer learning tasks are added on the basis of the pre-trained language model TransBERT.Firstly,the model learns the connection between the conversational text and the background information,so that the model can obtain better semantic representation.Secondly,through sentence-level emotion classification training procedure,the model learns the emotion-related representation related to sentence-level sentiment analysis and ERC.In this paper,emotion-related supervised knowledge is transferred to the target model by parameter sharing.With the help of these transfer learning tasks,the model can obtain better initialization parameters.Experimental results show that the ERC-TL method can identify emotions more accurately than BERT baseline,which verifies the effectiveness of the proposed method.