The paper discusses a new approach to target sentiment classification (TSC) leveraging pre-trained language models (PTLMs). The researchers argue that existing PTLM-based models have limitations. They propose a solution that combines the strengths of language modeling and explicit target-context interactions. They design a domain- and target-constrained cloze test to generate target attributes related to the review context. These attributes are used to construct a heterogeneous information graph. A novel heterogeneous information gated graph convolutional network is then used to model interactions among attribute, syntactic, and contextual information. The model outperforms others on three benchmark datasets.
Publication date: 21 Dec 2023
Project Page: https://arxiv.org/abs/2312.13766
Paper: https://arxiv.org/pdf/2312.13766
