Leveraging Human Reading Behavior for Keyphrase Extraction: A Webcam-based Eye-tracking Corpus

Traditional keyphrase extraction (KPE) models rely heavily on textual features like term frequency and semantic context, often ignoring how humans actually read and process information. This study investigates whether incorporating human reading behavior—captured via cost-effective, webcam-based eye-tracking—can provide auxiliary signals to improve the accuracy of KPE in Chinese academic abstracts.

The authors developed a lightweight data collection platform using the open-source SearchGazer library to capture eye-tracking data from readers of Chinese Library and Information Science (LIS) abstracts. They constructed the Chinese LIS Eye-Tracking Corpus (CLIS-ET), which includes three character-level metrics: first fixation duration (FFD), fixation number (FN), and total fixation duration (TFD). These features were then integrated into neural KPE architectures, specifically the Att-BiLSTM+CRF model, to measure their impact on extraction performance compared to text-only baselines.

The integration of eye-tracking features consistently enhanced KPE performance across the tested models. The researchers found that combining fixation number (FN) and total fixation duration (TFD) yielded the best results. This suggests that the cognitive effort and attention patterns reflected in these metrics serve as valuable indicators for identifying keyphrases, effectively bridging the gap between human cognitive processing and automated text analysis.

This research demonstrates that high-precision, laboratory-grade equipment is not strictly necessary to leverage human reading behavior for NLP tasks. By providing a cost-effective methodology and a new domain-specific corpus (CLIS-ET), the authors lower the barrier for researchers to incorporate cognitive signals into information retrieval and text-processing systems. This approach offers a promising path toward building more cognitively grounded models that better align with human attention.