On The Use Of Electroencephalography In Query Performance Prediction

Query Performance Prediction (QPP) enables information retrieval systems to estimate search effectiveness without requiring explicit relevance judgments. While traditional QPP research has focused exclusively on textual features, we investigate enhancing QPP through multimodal integration of electroencephalography (EEG) and eye-tracking signals captured during listening and reading. We utilize a specialised dataset where queries are represented as text alongside corresponding neurophysiological recordings, with graded relevance judgments for multiple documents. Our methodology employs an ensemble architecture with dedicated models for each modality, followed by a meta-learner that produces final predictions. Experimental evaluation across reading and listening tasks demonstrates that generalised models trained across subjects achieve consistent improvements over text-only baselines, with the trimodal configuration (EEG+eye-tracking+text) reaching Pearson correlations of 0.458 for the ZuCo reading dataset and bimodal EEG+text models achieving 0.417 for the Narrative listening dataset, representing 50-80% higher performance than personalised single-subject models. Query-level analysis reveals that neurophysiological signals substantially improve predictions for fragmentary or semantically ambiguous query cases, while the majority show neutral effects. These findings establish the feasibility of neurophysiological QPP under specific conditions and provide design principles for integrating brain-computer interfaces in information retrieval systems.