Semantically Guided Cross-Subject EEG Emotion Recognition with Pretrained Language Model

Abstract

EEG emotion recognition typically follows the standard signal-to-label paradigm. It maps EEG signals to discrete labels via feature engineering and machine learning, yet neglects the rich semantic information inherent in the task. This purely data-driven approach easily overfits to subject-specific physiological patterns, severely degrading its generalization on new subjects. Here, we propose a Semantically Guided Domain Adaptation (SGDA) method for cross-subject EEG emotion recognition. Specifically, SGDA projects non-stationary EEG features into an emotional semantic space, leveraging the discriminative power of structured linguistic embeddings to enhance class separability. To construct this unified space, a pretrained language model is utilized to extract linguistic embeddings from emotion text labels, which serve as invariant semantic prototypes to guide cross-domain feature alignment. Moreover, to facilitate robust inference, an adaptive distribution-aware feature fusion strategy dynamically aggregates multi-source knowledge based on the distances between target samples and source centroids. Extensive experiments on five benchmarks demonstrate that SGDA achieves state-of-the-art performance in cross-subject scenarios. The source code is publicly available at https://github.com/EDH777/SGDA.