Label knowledge-guided heterogeneous graph contrastive learning for semi-supervised short text sentiment classification
Year: 2025
Type: article
Abstract: Semi-supervised classification demonstrates effective performance in categorizing short-length texts, such as social media posts and online reviews, through the utilization of limited labeled data. Consequently, semi-supervised short text sentiment classification has emerged as a significant research domain within semi-supervised short text classification. However, existing sentiment classification methods predominantly rely on extensive labeled datasets for implementation and typically treat textual labels as discrete symbolic representations (e.g., categorical identifiers for classification tasks). This conventional method results in oversight of two critical linguistic dimensions: the inherent linguistic characteristics embedded within labels themselves and the underlying semantic correlations between labels and textual content. To address the limitations above, this study proposes a novel Label Knowledge-guided Heterogeneous Graph Contrastive Learning (LKG-HGCL) framework for semi-supervised short text sentiment classification. Specifically, we incorporate both label-related terms and their semantic expansions as label knowledge to construct a short text-label knowledge heterogeneous graph, explicitly modeling dynamic interactions between label semantics and short texts. By performing heterogeneous graph contrastive learning through multi-relational edge augmentation, adaptive feature augmentation, heterogeneous graph encoding, and the various contrastive learning modes, the model significantly enhances its capability to capture critical label semantics while generating optimized short text embeddings. The framework establishes robust associations between label knowledge and limited labeled and large amounts of unlabeled short texts, thereby effectively improving semi-supervised learning performance in sentiment analysis. Extensive experiments on three benchmark datasets demonstrate that the proposed LKG-HGCL method outperforms state-of-the-art semi-supervised approaches in classification accuracy and Macro-F1 metrics.
Source: Journal Of Big Data
Author: Mingqiang Wu
Institution: Hunan University of Arts and Science
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,290
Funder: Hunan University
OpenAlex ID: W4414597476