کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

BrainTranslator: Decoding Coherent Text from EEG Signals with Scheduled Sampling

نویسندگان :

Fatemeh Majidi¹ Mohammad Bagher Khodabakhshi² Mohammad Reza Rezaeian³

1- hamedan university of technology 2- hamedan university of technology 3- hamedan university of technology

کلمات کلیدی :

EEG-to-Text،Transformer،BrainTranslator،Scheduled Sampling،Brain–Computer Interface (BCI)

چکیده :

In recent years, converting EEG signals directly into natural text has become a major challenge in brain–computer interfaces. This study presents BrainTranslator, a Transformer-based model that generates coherent sentences directly from neural activity. The architecture includes an EEG encoder and a BART-based decoder. The encoder produces word-level and sentence-level representations; the sentence-level vector is derived through temporal mean pooling as an 840-dimensional feature and is placed at the start of the sequence. The decoder generates text after a linear projection layer (840 to 1024). Training follows a two-stage procedure. First, the EEG encoder is trained while the decoder remains fixed. Then, the full network is optimized end-to-end. Scheduled Sampling is applied to create realistic generation conditions and reduce Exposure Bias. The model is evaluated in an inter-subject setting using the ZuCo dataset. Results indicate that even under fully generative conditions (without Teacher Forcing), BrainTranslator achieves BLEU-1 of 0.4372, BLEU-4 of 0.0911, and a CER of 1.21. Accuracy is higher for key lexical items and short syntactic patterns than for longer sequences. Overall, BrainTranslator offers an effective framework for direct EEG-to-text decoding and supports progress in assistive communication technologies and next-generation brain–computer interfaces.