基于层级图卷积网络的情绪识别模型  

Emotion Recognition Based on Hierarchical Graph Convoluation Networks

作　　者：聂小芳 谭宇轩 曾雪强[1] 左家莉[1] NIE Xiaofang;TAN Yuxuan;ZENG Xueqiang;ZUO Jiali(School of Computer&Information Engineering,Jiangxi Normal University,Nanchang,Jiangxi 330022,China)

机构地区：[1]江西师范大学计算机信息工程学院,江西南昌330022

出　　处：《中文信息学报》2024年第6期139-150,共12页Journal of Chinese Information Processing

基　　金：国家自然科学基金(62266021)。

摘　　要：细粒度情绪识别模型采用比传统方法更多的情绪类别,能更为准确地捕捉人们日常生活中经历和表达的情绪。然而,大幅增加的情绪类别以及细粒度情绪间存在的相互关联和模糊性,给细粒度情绪识别模型带来了挑战。已有情绪识别工作表明,引入情感词典等外部知识可以有效提升模型性能。但现有细粒度情绪识别模型引入情感知识的方式还较为简单,仍未考虑深层情感知识,例如,情感层级关系。针对上述问题,该文提出一种基于层级图卷积网络的情绪识别(Hierarchy Graph Convolution Networks-based Emotion Recognition,HGCN-EC)模型。HGCN-EC模型由语义信息模块、情绪层级结构知识模块和知识融合模块组成。语义信息模块提取文本的语义特征;情绪层级结构知识模块将细粒度情绪构建为树状层级结构并使用贝叶斯统计推断计算情绪之间的转移概率作为层级知识;知识融合模块采用图卷积网络将情绪层级知识与文本语义特征融合,用于实现情绪预测。在GoEmotions数据集上的对比实验结果表明,HGCN-EC模型具有相较于基线方法更优的细粒度情绪识别性能。The fine-grained emotion recognition involves more emotion categories and can accurately discriminate emotions that individuals experience and express in their daily lives.To consider deep sentiment knowledge like emotional hierarchy relationships,this paper proposes a Hierarchical Graph Convolution Networks based Emotion Recognition(HGCN-EC)model.The HGCN-EC model consists of three modules:a semantic information module,an emotion hierarchy knowledge module,and a knowledge fusion module.The semantic information module extracts the semantic features of texts.The emotion hierarchy knowledge module organizes fine-grained emotions into a tree-like hierarchy and calculates the transition probabilities between emotions as hierarchy knowledge based on Bayesian statistical inference.The knowledge fusion module uses graph convolutional network to fuse hierarchy knowledge of emotions with text semantic information for emotion prediction.Experiments on the GoEmotions dataset show that the proposed method outperforms baseline methods.

关 键 词：细粒度情绪识别 图卷积网络 情绪层级知识 GoEmotions 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]