融合多策略论元识别的框架语义角色标注模型OA

Frame Semantic Role Labeling(FSRL)is a key task in the field of natural language processing.Its core objective is to identify the arguments governed by a target word within the evoked frame,and assign accurate semantic role labels to these arguments,thereby providing semantic support for downstream tasks.Currently,the performance bottleneck of FSRL mainly lies in the argument identification stage,specifically manifested in two issues:argument boundary errors and argument false positives.To address these challenges,we propose a multi-argument identification strategy that integrates label prediction,boundary awareness,and argument enu-meration.Concretely,the label prediction module jointly models argument identification and role labeling through end-to-end sequence labeling,enabling globally optimal label prediction.The boundary awareness module is optimized collaboratively with label prediction via multi-task learning,which enhances the model's ability to locate argument boundaries.The argument enumeration module generates a candidate argument set to filter out false-positive arguments.The integration of these three strategies achieves complementarity across three dimensions:global semantic association,local boundary localization,and constrained filtering output.Experimental results on two standard datasets(FN1.5 and FN1.7)demonstrate that the proposed model outperforms the state-of-the-art baseline model(AGED)by1.35 percentage points and 0.24 percentage points in terms of F1-score,respectively.Notably,the model also achieves a significant improvement in precision.In addition,the proposed method does not rely on external semantic resources,providing an effective solution for resource-constrained scenarios.