融合VAD知识的情感分布增强细粒度情绪识别方法

VAD Emotion Distribution Augmented Fine-grained Emotion Recognition

摘要

细粒度情绪识别模型采用几十种情绪类别对人类情绪进行建模，能够比传统模型更准确地捕捉人们表达的细微情绪。然而，现有的情绪预测模型并未充分考虑到数量众多的细粒度情绪之间存在的复杂相关性。针对这一问题，本文提出了一种融合VAD(Valence-Arousal-Dominance，效价-唤醒-支配)知识的情感分布增强细粒度情绪识别方法（VAD Emotion Distribution Augmented Fine-grained Emotion Recognition,EDAER）。EDAER模型采用情感分布建模VAD空间中情绪的相关性，结合文本语义信息与心理学先验知识进行细粒度情绪识别。EDAER模型包括语义信息模块、情感分布信息模块和融合预测模块。语义信息模块通过预训练的BERT(Bidirectional Encoder Representation from Transformers)模型提取文本的语义特征；情感分布信息模块基于VAD距离度量情绪间的相似度，为情感词生成情感分布；融合预测模块则利用注意力机制将文本语义信息与情感分布信息整合，并实现情绪预测。在GoEmotions数据集上的实验结果表明，EDAER模型的宏平均F1值达到51.75%，优于使用情感词典作为外部知识的KEA(Knowledge-Embedded Attention)模型和采用情绪层级关系作为外部知识的HGCN-EC(Hierarchy Graph Convolution Networks based Emotion Recognition)模型。特别是在三个样本量较少的情绪类别上，EDAER的F1值显著高于其他模型。实验结果验证了通过情感分布建模VAD空间中情绪的相关性，可以有效学习罕见情绪的相关知识，从而提升模型对细粒度情绪的识别能力。

Abstract

Fine-grained emotion recognition models, which employ dozens of emotion categories to model human emotions, are capable of capturing subtle emotional expressions more accurately than traditional models. However, existing emotion prediction models have not fully considered the complex correlations that exist among the numerous fine-grained emotions. To address this issue, this paper proposes an VAD (Valence-Arousal-Dominance) Emotion Distribution Augmented Fine-grained Emotion Recognition (EDAER). EDAER models the emotional correlations in the VAD space using emotion distributions and combines textual semantic information with psychological priors for fine-grained emotion recognition. The EDAER model consists of three modules: a semantic information module, an emotion distribution information module, and a fusion prediction module. The semantic information module extracts textual semantic features using a pre-trained BERT (Bidirectional Encoder Representation from Transformers) model; the emotion distribution information module generates emotion distributions for emotion words based on VAD distance metrics to measure the similarity between emotions; and the fusion prediction module integrates textual semantic features and emotion distribution information through an attention mechanism to predict emotions. Experimental results on the GoEmotions dataset demonstrate that the macro-average F1 score of the EDAER model reaches 51.75%, outperforming both the KEA (Knowledge-Embedded Attention) model, which uses emotion lexicons as external knowledge, and the HGCN-EC (Hierarchy Graph Convolution Networks based Emotion Recognition) model, which utilizes hierarchical emotion relationships as external knowledge. Notably, for three emotion categories with fewer samples, EDAER significantly outperforms other models in terms of F1 score. These results validate that modeling emotional correlations in the VAD space through emotion distributions can effectively capture knowledge related to rare emotions, thus improving the model's ability to recognize fine-grained emotions.

Graphical abstract

关键词

VAD情绪空间 / 外部知识 / 情绪分类 / GoEmotions

Key words

VAD emotion space / external knowledge / emotion classification / GoEmotions

基金资助

国家自然科学基金(62266021)

江西省教育厅科学技术研究(GJJ2200330)