A Heterogeneous Network fused with Context-aware Contrastive Learning for Sarcasm Topic-Target Pair Identification
Minjie Yuan, Mengyu Xiang, Yuxuan Song, Qiudan Li, Jinye Fu, Daniel Dajun Zeng · 2024
Sarcastic comments are often used to express dissatisfaction with products or events. Mining the topics and targets can provide clues for analyzing the underlying reasons behind the sarcasm, which helps understand user demands and improve products service. Existing research mainly focuses on mining single facet of sarcasm, such as topic or target, ignoring the complex interrelations between them. To overcome the above challenges, this paper proposes a Heterogeneous Information Network fused with Context-Aware Contrastive Learning (HINCCL) method. This approach aims to model multi-view features including syntactic style, domain knowledge, and textual semantics through a hierarchical attention aggregation mechanism. Furthermore, a context-aware negative contrastive training strategy is designed to learn the differentiated representations between different topic-target pairs. The effectiveness of the proposed method is validated on a dataset constructed in the digital domain.