Modeling Temporal Context in EEG-to-Text Systems Using Self Attention

Shurid, Sadi Mahmud

Karoui, Balkis

Al-Radhi, Mohammed Sala

2026-05-27T09:22:09Z

2026

Brain-computer interface systems that translate electroencephalography (EEG) signals into text rely critically on the quality of learned neural representations. A key challenge in EEG-to-Text decoding is the effective modeling of temporal context, as neural activity related to language production is distributed over time and exhibits long-range dependencies. Many existing approaches employ convolutional encoders that primarily capture local temporal patterns, limiting their ability to represent global contextual information. In this work, we investigate the use of self-attention mechanisms to improve temporal context modeling in EEG-to-Text systems. We propose an encoder architecture that combines convolutional feature extraction with a Transformer-based self-attention module, enabling the integration of both local and long-range temporal information in EEG representations. The downstream text generation component is kept fixed, allowing a focused evaluation of the impact of encoder-level temporal modeling. Experiments are conducted on a publicly available EEG dataset for language decoding. The proposed self-attention-based encoder is evaluated against a convolutional baseline using standard text similarity metrics as well as an auxiliary semantic classification task. Results show consistent improvements in text decoding performance and representation quality, indicating that self-attention provides a more expressive temporal representation of EEG signals. These findings highlight the importance of temporal context modeling in EEG-to-Text systems and demonstrate that self-attention-based encoders offer a practical and effective enhancement for neural language decoding pipelines.

angol

Könyvfejezet

application/pdf

http://hdl.handle.net/10890/64939