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

AI summary

Overview

Research area: Natural language processing, specifically keyphrase extraction (KPE), combined with human-computer interaction research on reading behavior (webcam-based eye-tracking).

Technical level: Intermediate. The work combines corpus construction, a lightweight gaze-estimation tool (SearchGazer), character-level eye-tracking features, and neural KPE models (Att-BiLSTM+CRF).

Scope: A single study that asks whether cheap, webcam-derived reading-behavior signals can improve keyphrase extraction from Chinese academic abstracts in Library and Information Science.

What This Paper Is About

Keyphrase extraction research has largely concentrated on better textual representations of documents, treating keyphrases as purely linguistic objects. This paper argues that keyphrases are also units that draw a reader's attention during comprehension, so signals about where and how long a reader looks could help identify them. The authors build a Chinese academic eye-tracking corpus using only a webcam, then test whether three fixation-based features improve keyphrase extraction.

Key Contributions

1. A lightweight webcam-based data collection platform. The authors built a reading-data platform on top of the open-source SearchGazer library, offering a low-cost alternative to conventional eye-tracking hardware for collecting reading behavior at scale.

2. CLIS-ET, a Chinese academic eye-tracking corpus. Because eye-tracking data for Chinese academic reading is scarce, the authors constructed the Chinese LIS Eye-Tracking Corpus, covering Chinese academic abstracts in Library and Information Science.

3. Three character-level eye-tracking features defined and evaluated. First fixation duration (FFD), fixation number (FN), and total fixation duration (TFD) were extracted at the character level and injected into KPE models to measure their effect on extraction performance.

4. Public release of dataset and code. The abstract points to a GitHub repository for the dataset and code and a second repository for the reading eye-tracking system, making the approach reproducible.

Main Findings

• Eye-tracking features consistently helped. Adding the reading-behavior features improved keyphrase extraction performance across the models tested, rather than helping only in isolated cases.

• Fixation count plus total fixation duration was the strongest combination. Of the feature settings examined, combining FN and TFD produced the best results, and this was observed on the Att-BiLSTM+CRF model.

• Readers' fixation behavior carries usable signal. The authors interpret the improvement as evidence that how long and how often a reader fixates on a character is informative for deciding whether it belongs to a keyphrase in academic abstracts.

• No numeric results are reported in the abstract. The abstract states the direction and the best-performing configuration but does not give evaluation scores, model comparisons beyond the best case, corpus size, or baseline numbers; those details are not available from the abstract alone.

Methodology in Plain English

The researchers started from a practical obstacle: proper eye-tracking equipment is expensive and produces little data for Chinese academic reading. They therefore built their own reading-data collection setup using SearchGazer, an open-source library that estimates gaze from an ordinary webcam, and used it to record how people read Chinese academic abstracts in Library and Information Science. That collection became the CLIS-ET corpus.

From the recorded gaze data, they computed three measurements for each character: how long the reader looked at it the first time (first fixation duration), how many times the reader looked at it (fixation number), and how long the reader looked at it in total (total fixation duration). These three signals were then added to keyphrase extraction models as extra information alongside the text. By comparing models that used the eye-tracking features with models that did not, they could judge whether reading behavior actually helps identify keyphrases. Their best-performing setup combined fixation number and total fixation duration within an Att-BiLSTM+CRF extraction model.

Why This Matters

Impact on research. The paper opens a line of work that treats keyphrase extraction as a reading-comprehension problem rather than only a text-representation problem. It also addresses a data gap: eye-tracking resources for Chinese academic reading are limited, and CLIS-ET plus the released platform lower the barrier for others to collect similar data. Because the collection method relies on webcams rather than laboratory hardware, it makes reading-behavior research feasible at a much larger scale.

Real-world applications:

• Academic search and recommendation systems that surface keyphrases from abstracts, where reader-attention signals could refine which terms are highlighted.
• Digital library and reference-management tools that automatically tag or index Chinese-language papers.
• Reading-assistance interfaces that adapt summaries, highlights, or navigation to what a reader actually attends to.
• Low-cost user-studies setups for publishers, educators, or researchers who need reading data but cannot afford dedicated eye-tracking equipment.

Industry relevance. Search engines, scholarly databases, publishing platforms, and recommendation services all depend on accurately extracting key terms from documents. A webcam-based signal that improves extraction accuracy without specialized hardware is attractive to organizations that cannot deploy eye-tracking labs but can ask users to sit in front of a standard camera.

Future Directions

• Test whether the approach generalizes beyond Chinese LIS abstracts to other languages, domains, and document types, which the abstract does not address.
• Evaluate other eye-tracking measures and granularities. Only FFD, FN, and TFD at character level were examined, leaving open whether additional gaze features or word-level aggregation would help more.
• Assess the accuracy limits of webcam-based gaze estimation. Since the platform is deliberately lightweight, how much estimation error it introduces and how that error affects extraction performance remains an open question.
• Move from demonstrating improvement to characterizing it. The abstract reports a consistent positive effect but no magnitudes or comparisons against other methods, so follow-up work establishing how large the gains are and when they hold would be a natural next step.

Target Audience

Researchers and graduate students in natural language processing, information retrieval, and digital libraries, particularly those working on keyphrase extraction or on incorporating user behavior into text models. It is also relevant to human-computer interaction and reading-research groups interested in low-cost webcam-based eye-tracking, and to practitioners building indexing, search, or reading-support tools for Chinese-language academic content.

Authors’ abstract

Purpose: Keyphrases are statistically and semantically important textual units that can also attract readers' attention during comprehension. However, existing keyphrase extraction (KPE) studies mainly focus on improving textual representation while largely overlooking human reading behavior. This study examines whether lightweight webcam-based eye-tracking features can improve KPE from Chinese academic abstracts in Library and Information Science (LIS). Methodology: To address the limited availability of eye-tracking data for Chinese academic reading, we developed a lightweight webcam-based data collection platform using the open-source SearchGazer library and constructed the Chinese LIS Eye-Tracking Corpus (CLIS-ET). Three character-level eye-tracking features, first fixation duration (FFD), fixation number (FN), and total fixation duration (TFD), were incorporated into KPE models to evaluate their effects on extraction performance. Findings: Eye-tracking features consistently improved KPE performance. The combination of FN and TFD achieved the best results on the Att-BiLSTM+CRF model, indicating that readers' fixation behavior provides useful signals for identifying keyphrases in academic abstracts. Originality/value: This study introduces a cost-effective webcam-based eye-tracking approach for KPE and presents CLIS-ET, a Chinese academic eye-tracking corpus containing FFD, FN, and TFD features. The results demonstrate the value of incorporating human reading behavior into keyphrase extraction. Dataset and code: https://github.com/yan-xinyi/ET_AKE and https://github.com/yan-xinyi/Reading_ET_System.