提示学习和知识增强的对话文本情绪识别  

Emotion Recognition in Conversations through Prompt Learning and Knowledge Enhancement

作　　者：余本功[1] 邵梦露 YU Bengong;SHAO Menglu(School of Management,Hefei University of Technology,Hefei 230000,China)

机构地区：[1]合肥工业大学管理学院,安徽合肥230000

出　　处：《情报科学》2024年第11期1-10,21,共11页Information Science

基　　金：国家自然科学基金项目“基于在线评论的复杂产品用户个性化动态知识服务机制研究”(72071061)

摘　　要：【目的/意义】对话情绪识别是情感分析领域的一个新兴且重要的子领域,旨在从多种预定义情绪中准确识别每个话语的情绪状态。不仅适用于情感相关社交媒体的意见挖掘,还可以实时分析对话情绪,对舆情分析、用户情感洞察等领域有着积极的影响。【方法/过程】提出了一种名为ERC-PrKE的对话情绪识别模型,该模型通过整合从外部结构化数据中学习到的知识,为理解对话提供额外的语境信息,从而增强对话建模并丰富上下文表达。此外,在目标对话输入之后,引入了提示模板,使模型更好地理解目标话语并将其与上下文区分开。【结果/结论】实验结果表明,通过增加提示和引入外部知识使得对话情绪识别的性能提升,在MELD、EmoryNLP和IEMOCAP数据集上性能提升超过2%,在EmoryNLP数据集上更是达到了4.05%的增幅。【创新/局限】提出了一种融入提示与外部知识的对话情感分析模型,为在对话式人工智能迅猛发展的背景下改善用户情感体验提供了新的方向。未来将进一步考虑整合其他外部知识源,以增强情绪识别的鲁棒性和准确性。【Purpose/significance】Emotion recognition in conversations is a burgeoning and vital subfield within sentiment analysis.It is devoted to accurately discerning each utterance's emotional state from an array of predefined emotions.It applies to not only sentiment mining of emotion-focused social media but also real-time dialogue emotion analysis,positively impacting sectors such as public sentiment analysis and user emotion insight【.Method/process】This paper presents a dialogue emotion recognition model named ERCPrKE.The model integrates knowledge gleaned from external structured data,offering additional contextual information for dialogue comprehension.Such integration boosts the modeling of dialogues and enriches the representation of context.Moreover,we incorporated a prompt template after the target dialogue input to help the model distinguish between the target utterance and its context more effectively.【Result/conclusion】The experimental outcomes suggest that the performance of dialogue emotion recognition can be enhanced by implementing prompts and introducing external knowledge.The performance has improved by over 2%on the MELD,EmoryNLP,and IEMOCAP datasets,with a significant leap of 4.05%on the EmoryNLP dataset.【Innovation/limitation】We have proposed a dialogue sentiment analysis model that integrates prompts and external knowledge,paving a fresh path to enhancing users'emotional experience amid the swift advancement of dialogic artificial intelligence.Future research will consider integrating various sources of external knowledge to bolster the robustness and precision of emotion identification.

关 键 词：对话情绪识别 提示学习 知识增强 对话理解 情感分析 

分 类 号：G254[文化科学—图书馆学]