개방형 어휘 환경에서의 EEG-To-Text 디코딩 성능 향상을 위한 잠재 확산 기반 모델

초록

Driven by recent advances in deep learning and the growing need to decode EEG into text under openvocabulary settings, we address the persistent issues of weak EEG–text alignment, noise sensitivity, and poor crosssubject generalization that limit prior EEG-to-Text systems. We propose Latent Diffusion with DeepConvNet (LDMDConv), which embeds EEG into a probabilistic latent space and pretrains an encoder via a latent diffusion process with a UNet denoiser conditioned on text embeddings, thereby explicitly modeling uncertainty while strengthening cross-modal alignment. During decoding, the pretrained encoder and EEG-specialized DeepConvNet extract spatial–spectral features that are passed to a pretrained BART decoder to generate text. Experiments on the public Zurich Cognitive Language Processing (ZuCo) dataset—EEG recorded while participants read sentences—demonstrate that LDMDConv achieves state-of-the-art performance and improved robustness to inter/intra-subject variability and noise. Quantitatively, the model attains a BLEU-1 of 42.40% and a ROUGE-1 F1 of 40.64%, surpassing strong baselines, with consistent gains across higher-order BLEU scores. These results indicate that combining diffusion-based uncertainty modeling with EEG-aware feature extraction materially improves open-vocabulary EEG-to-Text decoding. We conclude that LDM-DConv offers a robust foundation for this task and advocates future work on larger cross-subject corpora and evaluation metrics that capture semantic fidelity and generative diversity beyond exact n-gram overlap.

키워드