计算机科学
鉴定(生物学)
人工智能
图形
对比分析
自然语言处理
心理学
语言学
分类学(生物学)
生态学
沟通
感知
任务分析
认知科学
出处
期刊:
日期:2026-09-01
卷期号:: 150-153
标识
DOI:10.1109/ichci71425.2026.11739796
摘要
To address limited labeled samples, semantic similarity among violation categories, and insufficient exploitation of inter-case relations in ecological and environmental violation cases, this paper proposes a few-shot intelligent identification method integrating graph attention and prototypical contrastive learning (GAPCL). Chinese RoBERTa-wwm-ext is employed to obtain contextual semantic representations, based on which a case relation graph is constructed. A multi-head graph attention network aggregates information from related cases to enhance case representations. Class prototypes are constructed from support samples, while prototypical contrastive learning improves intra-class compactness and inter-class separability. Query cases are classified according to their similarity to class prototypes. Experiments are conducted on EcoCase-12, CrossRegion-10, EnvJudgment-10, and CAIL-FSL under 1-shot, 3-shot, 5-shot, and 10-shot settings. On EcoCase-12, GAPCL achieves an average Accuracy of 85.47% and Macro-F1 of 84.31% under the 5-shot setting, improving Macro-F1 by 3.77 percentage points compared with RoBERTa-GAT. The results indicate that GAPCL improves few-shot identification performance on datasets with different case distributions and text characteristics.
科研通智能强力驱动
Strongly Powered by AbleSci AI