Revisiting Neurological Aspects of Relevance: An EEG Study

Abstract

Relevance is a key topic in Information Retrieval (IR). It indicates how well the information retrieved by the search engine meets the user's information need (IN). Despite research advances in the past decades, the use of brain imaging techniques to investigate complex cognitive processes underpinning relevance is relatively recent, yet has provided valuable insight to better understanding this complex human notion. However, past electrophysiological studies have mainly employed an event-related potential (ERP) component-driven approach. While this approach is effective in exploring known phenomena, it might overlook the key cognitive aspects that significantly contribute to unexplored and complex cognitive processes such as relevance assessment formation. This paper, therefore, aims to study the relevance assessment phenomena using a data-driven approach. To do so, we measured the neural activity of twenty-five participants using electroencephalography (EEG). In particular, the neural activity was recorded in response to participants' binary relevance assessment (relevant vs. non-relevant) within the context of a Question Answering (Q/A) Task. We found significant variation associated with the user's subjective assessment of relevant and non-relevant information within the EEG signals associated with P300/CPP, N400 and, LPC components, which confirms the findings of previous studies. Additionally, the data-driven approach revealed neural differences associated with the previously not reported P100 component, which might play important role in early selective attention and working memory modulation. Our findings are an important step towards a better understanding of the cognitive mechanisms involved in relevance assessment and more effective IR systems.

Comparing point‐wise and pair‐wise relevance judgment with brain signals

“…The above analysis shows that the significant difference in each component of ERP is consistent with the overall trend difference between the two paradigms. All the EEG signal patterns analyzed of point-wise relevance judgment above are aligned with previous works on relevance judgment (Moshfeghi et al, 2016;Pinkosova et al, 2020Pinkosova et al, , 2022. However, when we shift our focus to the pair-wise paradigm, we find that all conclusions are, in fact, the opposite (addressing RQ1).…”

“…This ideal image meets all the given requirements and aligns with the assessor's personal aesthetic preferences. Consequently, highly relevant images that closely match this ideal standard will be evaluated more effortlessly, as previously discussed (Pinkosova et al, 2020(Pinkosova et al, , 2022. In contrast, during the pair-wise relevance judgment, assessors are exposed to an initial image before making their judgment.…”

“…These commonly used ERP components capture distinct stages of neural processing, such as early sensory processing (P100), semantic processing (N400), attention allocation and decisionmaking (P300), and language syntax and grammar processing (P600). Previous studies have shown that ERP components also exhibit some fixed patterns in the procedure of relevance judgments (Pinkosova et al, 2020(Pinkosova et al, , 2022. We apply standard ERP analysis methods to the recorded data, including preprocessing, dividing time window, and identifying region of interest (ROI) (Luck, 2014).…”

“…Furthermore, studies by Ye et al (2022) utilize the ERP method to investigate the process of reading comprehension. In the context of point-wise relevance judgment, Pinkosova et al (2020Pinkosova et al ( , 2022 conduct a user study and observe significant differences in ERP of highrelevance, low-relevance, and no-relevance answers. Pinkosova et al (2023) suggest that self-perceived knowledge (SPK) play an important role in relevance assessment, and Michalkova et al (2024) explore the neurological mechanisms behind feeling-of-knowing.…”