基于大語言模型的多任務(wù)生成式重構(gòu)對話情緒識別

打開文本圖片集

中圖分類號：TP391 文獻(xiàn)標(biāo)志碼：A 文章編號：1001-3695（2025）07-006-1964-08

doi：10.19734/j. issn. 1001-3695.2024.12.0486

Abstract：Emotionrecognition inconversation（ERC）isakey task indialogue systems research.However，existing models oftensuferfromoverfittng tospecificdatasetsanddialoguepatersduetothecomplexityofpipelinedsign，hichlimitstheir generalizationability.Toaddress thisisse，thisstudyproposedamulti-task generativeemotionrecognitioninconversatin（MGERC）model basedonlarge language models.The model introduced two auxiliarytasksbasedon pre-trained large language models;speakeridentifcationandtopic-basedemotionprediction.Thespeakeridentificationtaskaimed toimplicitlymodelthe relationshipsbetweeconversationaloles，lpingthemodelbeterunderstandmotionalxchangesbetween diffrentparticipants.The topic-based emotionprediction task predicted theglobalthemeoftheconversation，capturing thepotentialconnectionbetweentopicsandemotions，thusimproving emotionrecognitionaccuracybyincorporatingcontextualinformation.Aditionaly，M-GERCintroducedaknowledgeretrieval modulethatretrieveddomain-specificknowledgeandintegratedexteral knowledge to further enhance the model’sunderstanding ofcontext.Experimentalresultsshow that M-GERC significantlyoutperforms existing mainstream ERC models，achieving W-F1 improvements of 3.1% ， 4.3% and 3.7% on the DailyDialog， MELD and EmoryNLP datasets，respectively.

Key Words：emotion recognition in conversation；large language models；topic；external knowledge

0 引言

“如果想要真正的智能機(jī)器，那么必須讓它們能夠理解情緒，而不僅僅是執(zhí)行特定的任務(wù)。（剩余22100字）