Journal of King Saud University Computer and Information Sciences Компьютерные науки (общие)
Span-level emotion cause extraction with Gaussian-enhanced boundary-aware conditional fast gradient method
Аннотация
Abstract Span-Level Emotion Cause Extraction (SECE) aims to identify and extract cause spans that trigger specific emotion in documents. Existing studies overlook the structural asymmetry in the SECE task caused by heterogeneous expression patterns and structural sparsity, making it difficult to construct effective cross-instance shared features at the model level. Although adversarial training can improve generalization by introducing controllable perturbations, standard FGM applies uniform spatial perturbations across the entire sequence, implicitly assuming that all positions contribute homogeneously to the loss, which contradicts the significant differences in gradient sensitivity and decision uncertainty between causal boundary regions and background regions in the SECE task. To this end, we explicitly condition adversarial perturbations on the SECE structural prior, proposing the Gaussian-Enhanced Boundary-Aware Conditional Fast Gradient Method (GEBA-C-FGM). The method employs C-FGM to impose differentiated perturbation intensity constraints on boundary and non-boundary tokens, redistributing adversarial budgets at the embedding layer to enhance discriminative robustness in boundary regions. The boundary-aware module dynamically amplifies the discriminative dimensions of boundary features, converting adversarial robustness into precise boundary localization accuracy. Upper-layer Gaussian noise penetrates the semantic abstraction layers of Transformers, imposing structural regularization constraints on cross-clause causal reasoning. Experiments have shown that GEBA-C-FGM outperforms state-of-the-art baseline approaches with an improvement of 2.2%, 3.4%, and 1.9% F1 score on SECE, SECCE, and English ECA benchmark dataset.