DEEPER: Dense Electroencephalography Passage Retrieval

2024-12-09

arXiv.org

AI Summary

To address the challenge of translating users’ intrinsic cognitive needs into textual queries for information retrieval, this paper proposes the first end-to-end EEG-to-passage dense retrieval paradigm, bypassing the conventional brain-signal-to-text decoding step. Methodologically, we construct a cross-modal unified semantic space, design a dedicated time-frequency EEG feature encoder, and adopt a dual-tower Transformer architecture enhanced with contrastive learning and multi-granularity negative sampling to achieve robust semantic alignment. On the ZuCo dataset, our approach achieves retrieval accuracy nearly five times higher than baseline methods; it demonstrates consistent performance across 30 subjects, indicating strong generalizability. This work provides the first empirical validation of direct passage retrieval from raw EEG signals, establishing a novel pathway for brain–computer interface–driven information retrieval.

Technology Category

• Cognitive Modeling & Cognitive Systems: Neural Spike Coding
• Humans and AI: Brain-Sensing and Analysis
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Querying, indexing, and retrieval in Web-related graphs

Abstract

A fundamental challenge in Information Retrieval (IR) is the cognitive burden of translating internal information needs into explicit textual queries. This translation barrier particularly affects users with undefined information needs or those who face physical constraints in traditional text input methods. While Brain-Machine Interfaces (BMIs) have emerged as a potential solution for direct neural query interpretation, existing approaches that attempt to convert brain signals into text queries have demonstrated limited success in capturing the complexity of neural semantic patterns. This paper introduces DEEPER Dense EEG Passage Retrieval, a novel framework that bypasses the need for explicit query translation by directly mapping electroencephalography (EEG) signals to relevant text passages. Our approach employs dense retrieval architectures to create a unified semantic space where both EEG signals and text passages can be effectively compared. Experimental evaluation on the ZuCo dataset shows that DEEPER substantially outperforms current EEG-to-text baselines, achieving nearly 5x improvement in retrieval precision while demonstrating robust performance across a diverse set of 30 participants. Through detailed ablation analysis, we identify key architectural components, including specialized neural encoders and strategic negative sampling techniques, that enable effective cross-modal semantic alignment. Our findings demonstrate the feasibility of direct EEG passage retrieval and suggest new possibilities for developing IR systems that can more naturally interface with users' cognitive processes.

Problem

• Brain Signals
• Information Retrieval
• Query Intention

Innovation

• Brain-Signal Matching
• Enhanced Information Retrieval
• User Intent Understanding