Utilizing Cognitive Signals Generated during Human Reading to Enhance Keyphrase Extraction from Microblogs

Automatic Keyphrase Extraction (AKE) is essential for organizing the massive volume of unstructured content on microblogging platforms. While previous research has used eye-tracking data to capture reader attention and improve AKE, these signals are limited by physiological and technical constraints. This study investigates whether electroencephalogram (EEG) signals—which directly reflect brain electrical activity—can complement or outperform eye-tracking data to enhance AKE performance.

The researchers utilized the ZuCo cognitive language processing corpus, extracting 8 EEG features and 17 eye-tracking features. To integrate these signals into AKE models without distorting the underlying cognitive data, they injected the features into the input of the soft-attention layer and the query vectors of the self-attention layer. They evaluated these signals across six different model architectures, including BiLSTM, BiLSTM+CRF, and various attention-based variants, and later extended the analysis to include Pre-trained Language Models (PLMs) like BERT and T5.

The study demonstrates that cognitive signals generated during human reading consistently improve AKE performance across all tested model architectures. EEG signals provided the most significant performance boost, particularly those in the beta and gamma frequency bands, which are associated with cognitive workload. Interestingly, while both signal types are beneficial, combining EEG and eye-tracking features resulted in performance levels between the two individual signal types, indicating that their integration may involve complex synergistic or redundant effects that are not yet fully optimized.

This research highlights the potential of neurophysiological data to improve Natural Language Processing (NLP) tasks. By demonstrating that EEG signals offer more comprehensive cognitive evidence than eye-tracking alone, the study provides a roadmap for future multimodal approaches to information extraction. It suggests that as signal acquisition technology becomes more accessible, incorporating direct brain activity measurements could significantly enhance the accuracy of automated content analysis systems.