Cognitive salience features enhance multitask deep learning for pragmatic reasoning across cultures

Cross-cultural pragmatic reasoning requires computational systems to interpret implicit meanings and culturally-specific communication patterns beyond literal semantic comprehension. This study proposes a cognitive salience feature-driven multi-task deep learning model that integrates attention-based mechanisms mirroring human perceptual priorities with culture-conditioned gating for cross-cultural pragmatic inference. The model jointly optimizes four pragmatic reasoning tasks—implicature detection, speech act classification, politeness assessment, and cultural appropriateness evaluation—through dynamic task weighting and gradient conflict resolution strategies. Comprehensive experiments conducted on a multilingual corpus spanning eight cultural groups demonstrate that the proposed approach achieves 83.2% overall accuracy and 77.5% cross-cultural transfer accuracy, representing substantial improvements of 9.7 and 12.7% points respectively over standard multi-task learning baselines. Ablation studies confirm that cognitive salience mechanisms contribute 5.2% performance gain, while culture-conditioned processing enhances cross-cultural generalization. The model’s learned salience patterns exhibit strong correlation (0.81) with theory-driven pragmatic markers, validating interpretability and theoretical grounding. This work advances computational pragmatics by demonstrating that human-inspired cognitive mechanisms significantly enhance artificial systems’ cross-cultural reasoning capabilities.

Text 125266.pdf - Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (5MB) Official URL or Download Paper: https://www.nature.com/articles/s41598-026-40809-8

Actions (login required)

View Item