基于字形的中文零样本立场检测

打开文本图片集

关键词：零样本立场检测；象形特征；语义增强策略；多模态融合

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2026）04-007-1021-07

doi：10.19734/j.issn.1001-3695.2025.09.0279

Chinese zero-shot stance detection based on glyph

Fu Shufan，Wang Zhongqing†，Miao Yifei （SchoolofComputer Scienceand Technology，Soochow University，Suzhou Jiangsu 215o08，China）

Abstract：ZSSDaims toidentifytheatitudeexpressed ina giventext toward anunseentarget.Existing ZSSDmethods usualy follwapproachesoriginallydevelopedforEnglishdatasetsanddirectlyapplythem to Chinese datasets.Thispracticeignores cross-linguisticdiferences andthepictographic natureofChinese characters，leading topoor performance in Chinese ZD. To address these problems，this paper proposed a Chinese ZSSD method thatintegrated visual glyph features with semantic understanding.The methodconvertedtheinputtextintoglyph imageswith pictographic featuresandperformedcharacter-level segmentation.Subsequently，this modelappliedasemanticenhancement strategyin the imagespace，labeling sentimentand entitywords toconstructvisualcues.Finaly，itdesigndamultimodalfusionframeworktoaligntextualsemanticswithglyphbased visual features.Experimentsonseven domainsoftheC-STANCEdataset demonstrate thatthe proposed framework outperformsmultiplebaseline models.Experimentalresultsdemonstratetheeffectivenessofglyphinformationandunderscoreits potential in advancing Chinese ZSSD tasks.

Keywords：zero-shot stance detection（ZSSD）；pictographic feature；semantic enhancement strategy；multimodal

0 引言

立场检测旨在通过文本语义分析，确定作者对特定目标对象表达的主观倾向性（如支持、反对、中立）[1]。（剩余19472字）