On May 2, 2023, PhD student Xiu Li will present her work “Towards Building an Intelligent AI-based Content Recommender System in Education”. The midterm seminar takes place at the Department of Computer and Systems Sciences (DSV), Stockholm University.

Respondent: Xiu Li, DSV
External reviewer: Robert Östling, Department of Linguistics, Stockholm University
Internal reviewer: Jaakko Hollmén, DSV
Main supervisor: Aron Henriksson, DSV
Supervisors: Jalal Nouri and Martin Duneld, DSV

Contact Xiu Li

About Xiu Li’s work

The digitalization of education has made massive and diverse learning materials available. Figure 1 shows the learning material types and their relations in a learning platform. The curriculum as the central educational guide determines what is essential for teaching and learning. Teachers design courses according to the curriculum and adopt learning contents from different textbooks, exercise books, and supporting materials inside and outside the learning platform, while also creating learning contents of their own such as lecture slides, exercises, instructions etc.

AI-enhanced educational recommender systems have emerged to tackle the big data challenge of filtering contents, finding items of interest, and suggesting useful learning resources that can meet curriculum goals, users’ needs and profiles. Such systems play an important role for both educators and learners. For educators, they can contribute to pedagogical practices through recommendations that improve their planning and assist in educational resource filtering such as in a Learning Content Management System and Educational portals [1, 2, 3]. For learners, they can provide relevant and personalized learning contents to facilitate knowledge acquisition in self-regulated learning, for example, by suggesting personalized remedial learning materials [4] and building a personalized review module within the learning objectives [5], or recommending Wikipedia pages and Youtube videos for external augmentation reading and learning [6, 7, 8]. Developing such an AI-enhanced service in a learning platform can thus be highly valuable to assist the teaching and learning processes.

From a technical perspective, content recommenders can be based on heuristic rules, semantic similarity, or user behavior. Rule-based recommenders usually involve sophisticated human logic and content attributes such as difficulty level, finishing time, types, etc. for pre-configuration. Semantic similarity-based recommenders focus on the content itself. They project contents to numeric representations in a feature space and calculate the distance of semantic meanings between texts. The semantics of learning material can be represented by the extracted educational concepts (ontology-based approach) [6, 4] or the contextual semantic meaning of the whole sentence [9, 10] (semantic-based approach). After extracting representative concepts for each learning content unit, learning materials can be linked by applying an exact concept phrase match, or using word embeddings to encode the extracted concepts or the whole sentence in a Semantic Textual Similarity (STS) approach, or through a combination of the aforementioned approaches.

A user behavior-based recommender is a personalized content recommender, focusing on students’ needs and building on the analysis of student profiles in terms of, for instance, knowledge level, preferences, learning style and focusing on what is “needed” by the student, such as recommending a learning object for practice that the student has not mastered well enough. A user behavior-based recommender can also be a “collaborative filtering” (CF), i.e. putting the students in a social context and grouping them with similar patterns of behaviors and then recommending the learning resources that most students in your group have engaged with but you have not [11, 12]. This is also called behaviour similarity.

Content recommendation systems can utilize information retrieval and semantic search technologies to cope with the content part, and accommodate rule-based and behavior-based personalized perspectives to further enhance the relevancy of recommended learning resources. Therefore, a hybrid approach is also common [4, 13, 14, 6]. Figure 2 shows how databases are designed to accommodate different data resources in a learning system. The user profile data and user behavior data can be used together with content data to enhance content recommendations to serve the end users.