Decoding Intent from Fragments: Structure-Aware Late Interaction for Concept-Level EEG-to-Text Retrieval

While retrieval-based paradigms have emerged as a more viable alternative to generative approaches for EEG-to-Text decoding, existing methods typically compress continuous EEG signals into holistic vectors. However, real-world brain-computer interaction is fundamentally driven by discrete conceptual units, and current methods that compress continuous EEG signals into holistic vectors struggle to capture the granularity of these sparse, fragmented intents. To address this limitation, we introduce a novel task: Brain Concept-to-Sentence Retrieval (BCSR), designed to decode intent from fragmented EEG signals. We propose SparkRetriever, a structure-aware retrieval model tailored for aligning discrete EEG concepts with natural language sentences. Departing from traditional dual-tower architectures, SparkRetriever incorporates a fine-grained late interaction mechanism. By preserving EEG tokens as independent semantic anchors, it enables precise multi-to-multi matching with text tokens. Furthermore, to tackle signal discreteness and cross-subject variability, we construct a robust semantic manifold leveraging Multi-Positive Contrastive Learning and Subject Embeddings. Extensive experiments on the ZuCo benchmark demonstrate that SparkRetriever significantly outperforms state-of-the-art baselines, achieving relative improvements ranging from 23.4% to 33.7% across multiple metrics. Our analysis reveals that multi-positive supervision acts as the primary driver of performance, effectively mitigating the bias inherent in single-sentence alignment.