Hello everyone, this is Li.
I would like to introduce a paper I read at the recent English seminar, along with my thoughts on it.

Paper Title: LECTOR: Summarizing E-book Reading Content for Personalized Student Support
Year of Publication: 2025
Authors: Erwin Daniel López Zapata, Cheng Tang, Valdemar Švábenský, Fumiya Okubo, Atsushi Shimada
Journal: International Journal of Artificial Intelligence in Education
Pages: 1-39

Background and Purpose
In recent years, the use of digital learning materials has expanded rapidly in many countries, including Japan, South Korea, and the United States. By using digital materials, it becomes possible to record learners’ reading behavior logs—detailed information such as “which slides” they read, “when,” and “for how long.” This allows instructors and researchers to visualize student learning behaviors and utilize them for grade prediction and instructional support.

However, many previous studies have focused solely on behavioral data, with insufficient attention paid to semantic information, such as “what was read” or “what kind of meaningful content was learned.” It has been pointed out that this leads to superficial analysis that ignores context, limiting the design of educational interventions and the realization of personalized support.

The purpose of this study is to propose a new model, “LECTOR,” which handles slide content information and reading behavior data in an integrated manner, and to demonstrate its effectiveness. The specific objectives are the following three points:

Automatically extracting keyword information from slide materials
Structuring the extracted content data and integrating it with reading behavior data
Predicting learners’ academic risks and providing a basis for personalized support

LECTOR utilizes “BERT (Bidirectional Encoder Representations from Transformers)” and the “Attention mechanism,” which are core technologies in natural language processing, to construct a “slide-topic relationship matrix” that represents the relationship between slides and topics. This model makes it possible to quantify students’ semantic understanding and interests.
Furthermore, this study sets the following two research questions:
Is a content extraction model that incorporates educational context more effective than traditional NLP methods?
Does integrating topic preference features improve the accuracy of academic risk prediction?

Text Data Processing
In previous studies, vocabulary statistical models such as TF-IDF (Term Frequency–inverse Document Frequency) have been used for keyword extraction from materials, summarization, and comparing similarities between materials. However, these methods take almost no account of contextual meaning, making it difficult to sufficiently capture educationally important information.

Keyphrase extraction technology has evolved from first-generation statistical models to second-generation graph models, and now to the currently mainstream third-generation embedding models (such as BERT). LECTOR, proposed in this study, belongs to this third generation and possesses uniqueness by leveraging the specific structure of educational slides. By introducing higher-performance language models, it attempts to overcome the limitations of conventional methods.

LECTOR is mainly composed of the following three modules: topic extraction; importance score; and similarity score.
The text of the slides is encoded by BERT, and semantic relevance is analyzed. Furthermore, by incorporating structural information such as slide titles and subheadings, deeper contextual understanding is made possible. Finally, a “slide-topic relationship matrix” is generated, and by multiplying this with the student’s reading behavior data, a “topic preference vector for each student” is calculated. This allows for the grasping of learning tendencies and risks.

Experimental Results and Discussion
In this study, to verify the effectiveness of LECTOR, the following two experiments were conducted.
In Experiment 1, we verified whether LECTOR could extract effective keywords from teaching slides compared to conventional methods (such as TF-IDF). As a result, LECTOR outperformed other models in all evaluation metrics. Conventional TF-IDF and AttentionRank tend to overlook educationally important keywords because they give excessive weight to frequent words or have a shallow understanding of context. In contrast, LECTOR utilizes the structural information of slides and extracts words while considering contextual meaning, thus enabling the extraction of more appropriate keywords. Additionally, the point that semantically important words were extracted without omission or excess due to weight correction by Smooth Inverse Frequency (SIF) and the introduction of Attention is also evaluated. However, LECTOR also has some challenges. For example, accuracy decreases when the slide structure is ambiguous, it cannot sufficiently capture the relationships between slides, and it cannot handle non-text information such as figures and mathematical formulas. It will be necessary to consider addressing these points in the future.

In Experiment 2, we verified whether the “topic preference vector” generated by LECTOR is effective for predicting students’ academic risk. As a result, prediction accuracy improved significantly, especially in contexts such as self-directed learning. This is thought to be because it was possible to visualize “interest in content,” which cannot be fully grasped by mere reading time or click counts. Thus, LECTOR is a flexible model that captures both behavioral data and semantic information in an integrated manner, and it can be said to have high potential in educational applications.

Reflections
The main reason I introduced this paper is that I felt the method of analyzing slide content data would be useful in my own research in the future. In particular, I thought this study provides significant implications for future research design in terms of integrating reading behavior data with semantic information. In addition, the richness of the experimental design was impressive. In Experiment 1, the effectiveness of content extraction in an educational context was verified against traditional NLP methods, and in Experiment 2, it was shown that integrating topic preference features could predict and explain students’ academic risk with higher accuracy. Furthermore, I felt that the LECTOR model has high potential for application in actual educational practice. For example, I believe it can be expected to be deployed in various educational applications, such as reading behavior analysis in electronic materials, construction of automatic summarization and recommendation systems for materials, and personalization and automation of learning support and interventions.