MBGFF Model for EEG-Text Generation Under Multi-Task Parallel Imagination

Brain-computer interface (BCI)-based text generation provides a promising communication pathway for individuals with severe motor or speech impairments. However, most existing EEG-based text generation studies rely on a single imagery modality, which limits information richness, decoding robustness, and practical communication applicability. To address these issues, this study proposes a text generation framework based on a Multi-Branch Global Feature Fusion (MBGFF) model under a multi-task parallel imagery paradigm that jointly incorporates visual, speech, and motor imagery. To reduce inter-subject variability, a personalized channel layout strategy based on one-way analysis of variance is designed for subject-specific EEG acquisition. The proposed MBGFF model further introduces a tri-branch architecture that combines multi-scale local feature extraction and global feature fusion through convolutional blocks, multi-head attention, and a global convolution block, thereby enhancing multi-character EEG decoding. In addition, an online EEG-to-text platform with real-time signal acquisition, decoding, and dual-model language correction is developed to support practical communication scenarios. Experiments on a 29-character EEG dataset collected from 10 subjects show that the proposed method achieves an average offline decoding accuracy of 71.48%. In online experiments, the average decoded character accuracy rate reaches 64.70%, and the average corrected character accuracy rate further improves to 77.39% after language correction. These results demonstrate that the proposed framework is effective for multi-character EEG decoding and shows strong potential for practical assistive EEG-based text generation.