文献类型：期刊文献

英文题名：HyGST: a graph-integration enhanced generative framework for end-to-end aspect-based sentiment analysis

作者：Ma, Zhiyuan[1];Gu, Chenxi[1];Wang, Nan[2];Yang, Ling[1];Wang, Yongjie[1]

机构：[1]Univ Shanghai Sci & Technol, Inst Machine Intelligence, 580 Jungong Rd, Shanghai 200093, Peoples R China;[2]East China Univ Sci & Technol, Sch Informat Sci & Engn, 130 Meilong Rd, Shanghai 200237, Peoples R China

年份：2026

卷号：17

期号：8

外文期刊名：INTERNATIONAL JOURNAL OF MACHINE LEARNING AND CYBERNETICS

收录：;EI(收录号：20263021162499);Scopus(收录号：2-s2.0-105045399625);WOS:【SCI-EXPANDED(收录号:WOS:001829113700003)】；

基金：This work is supported by National Local Joint Engineering Laboratory of Next Generation Internet Data Processing Technology under grant No. ZYGX2025K00802. The authors would like to thank Meiqi Pan for her previous investigation and contributions to this work, and also thank the anonymous reviewers for their valuable comments and helpful suggestions.

语种：英文

外文关键词：Aspect-based sentiment analysis; Generative language model; Graph convolutional network; Semantic representation

摘要：End-to-End Aspect-Based Sentiment Analysis (EABSA) is a pivotal task in fine-grained sentiment mining, which requires the joint extraction of aspect terms and prediction of their associated sentiment polarities from unstructured text. Despite the remarkable progress of generative models in ABSA, existing approaches still suffer from two critical limitations: inadequate capture of task-specific semantic associations and a prevalent over-prediction. These flaws not only hinder the accurate detection of implicit aspect terms but also compromise the reliability of sentiment polarity inference. To address these issues, we propose a novel Hybrid framework with Generation-guided Sequence Tagging (HyGST), which innovatively enhances task-relevant semantics by exploiting the intrinsic correlation between words and labels. HyGST encompasses two primary modules, specifically a generative T5 model and a sequence tagging module. A pre-trained T5 model can provide a context encoder, predicting potential aspect-sentiment pairs to guide subsequent fine-grained modeling. Meanwhile, the sequence tagging module constructs a token-label semantic graph based on the T5 encoder's outputs, explicitly modeling the relational dependencies between tokens and labels. Finally, a graph convolutional network aggregates contextual information from the constructed semantic graph, while leveraging global label-aware representations to deliver enhanced task-specific semantic supervision for EABSA. Preliminary experiments demonstrate that the proposed framework outperforms current EABSA methods, verifying its effectiveness in addressing the core challenges of EABSA.