| 摘 要: 情绪原因识别是一项具有挑战性的细粒度情感分析任务。针对现有方法对情绪语义特征的分析不够充分、对情绪及其原因的内在因果关系挖掘不够深入等问题,提出了一种基于多粒度语义融合和交互注意的情绪原因识别方法。该方法构建了词级、短语和子句三级编码网络,用于建模不同语言粒度下子句的上下文信息,每级网络均设计了交互注意机制,用于挖掘子句之间的深层语义和依赖关系。使用多维注意捕获词语的不同特征对于子句的重要程度。在基准语料库上的实验表明,该方法能够提高识别的性能。 |
| 关键词: 情绪原因识别 语义融合 交互注意 相对位置 多维注意 |
|
中图分类号: TP183
文献标识码: A
|
| 基金项目: 中国高校产学研创新基金(2021BCE02013) |
|
| Emotion Cause Recognition Based on Multi-granularity Semantic Fusionand Interactive Attention |
|
FAN Binfeng, CHEN Jiong
|
(Department of Artificial Intelligence, Shanxi Polytechnic College, Taiyuan 030006, China)
158663705@qq.com; 1851142388@qq.com
|
| Abstract: Emotion cause recognition is a challenging fine-grained sentiment analysis task. To address the limitations of existing methods—such as insufficient analysis of emotional semantic features and inadequate exploration of the intrinsic causal relationships between emotions and their causes—this paper proposes an emotion cause recognition approach based on mult-i granularity semantic fusion and interactive attention. The method constructs a three-level encoding network (word, phrase and clause) to model contextual information of clauses at varying linguistic granularities. An interactive attention mechanism is designed at each level to mine deep semantic and dependency relationships between clauses. Mult-i dimensional attention is employed to capture the significance of
diverse word-level features for clauses. Experiments on benchmark corpora demonstrate that the proposed method enhances recognition performance. |
| Keywords: emotion cause recognition semantic fusion interactive attention relative position mult-idimensional attention |
