Brain-CLIPLM: Decoding Compressed Semantic Representations in EEG for Language Reconstruction

A research team from Zhejiang University has introduced Brain-CLIPLM, a novel framework that challenges conventional approaches to decoding language from non-invasive brain recordings. The system operates on a "semantic compression hypothesis"—the idea that EEG signals encode compressed semantic anchors rather than full linguistic structure. This perspective addresses fundamental limitations in brain-computer interfaces, where low signal-to-noise ratios and restricted information bandwidth have made direct sentence reconstruction impractical.

Brain-CLIPLM implements a two-stage process that aligns decoding complexity with neural information capacity. First, it extracts semantic anchors from EEG data using contrastive learning techniques. Then, it employs a retrieval-grounded large language model with Chain-of-Thought reasoning to reconstruct complete sentences from these compressed representations. This granularity matching principle represents a significant departure from previous direct decoding methods.

Evaluated on the Zurich Cognitive Language Processing Corpus, Brain-CLIPLM achieved 67.55% top-5 and 85.00% top-25 sentence retrieval accuracy, substantially outperforming direct decoding baselines. Cross-subject evaluations confirmed robust generalization, while control analyses including permutation testing demonstrated that EEG-derived representations carry sentence-specific information beyond language model priors. The framework's success suggests that EEG-to-text decoding is better framed as recovering compressed semantic content rather than reconstructing full sentences.

This work provides a biologically grounded and data-efficient pathway for non-invasive brain-computer interfaces, potentially enabling new communication methods for individuals with speech impairments. The semantic compression approach could revolutionize how we interpret neural signals, moving from literal reconstruction to intelligent inference based on compressed semantic representations.