AI-Based Personalized Multilingual Course Recommender System Using

Large Language Models

Sourav Dutta, Florian Beier and Dirk Werth

August-Wilhelm Scheer Institut gGmbH, Uni-Campus D 5 1, 66123 Saarbr

ucken, Germany

{ﬁrstname.lastname}@aws-institut.de

Keywords:

Recommender System, Large Language Model, Natural Language Processing, Generative AI, AI in

Education.

Abstract:

This paper presents an AI-driven personalized course recommender system designed to enhance user engage-

ment and learning outcomes on educational platforms. Leveraging the EU DigComp competency framework,

the system constructs detailed user proﬁles through a chat assistant that guides users in identifying relevant

competency areas and completing tailored surveys. Course recommendations are generated based on a hybrid

scoring model that integrates semantic similarity and competency alignment, ensuring that course suggestions

are both contextually and skill-relevant. For users seeking structured guidance, the system offers a learning

path feature, utilizing a large language model to suggest subsequent courses that align with the user’s interests

and prior learning experiences. While traditional course recommenders often rely on simple keyword match-

ing, our system dynamically combines user interests and competencies for nuanced recommendations across

English and German courses. Screenshots of the system’s live demo showcase key functionalities, includ-

ing chatbot-led proﬁle creation, multilingual support, personalized learning paths. This paper highlights the

ongoing development of the recommender system and discusses future plans to further reﬁne and expand its

personalized learning capabilities.

1 INTRODUCTION

The rise of online learning platforms has revolu-

tionized access to education, allowing individuals to

learn at their own pace from a vast array of courses

(Pappano, 2012). However, the abundance of avail-

able resources can overwhelm learners, leading to the

need for personalized recommender systems to guide

users toward relevant courses that match their inter-

ests and skill levels. Recommender systems, tradi-

tionally based on collaborative ﬁltering and content-

based ﬁltering techniques, have shown promise in

various domains, including e-learning (Burke, 2002;

Manouselis et al., 2012). However, these methods of-

ten fall short when it comes to personalizing recom-

mendations based on a learner’s speciﬁc competency

proﬁle or learning goals (Adomavicius and Tuzhilin,

2005). The European Union’s Digital Competence

Framework for Citizens (DigComp) provides a struc-

tured approach to deﬁning digital skills and compe-

tencies. The framework outlines 21 key competen-

cies grouped into ﬁve dimensions, including informa-

tion literacy, communication, digital content creation,

safety, and problem-solving (Ferrari et al., 2014). By

integrating this competency framework into an edu-

cational platform, it becomes possible to generate a

proﬁle for each learner that reﬂects their strengths and

areas for improvement. This approach enables the de-

sign of personalized learning experiences that target

speciﬁc skills, offering users more relevant and effec-

tive learning paths.

Recent advancements in artiﬁcial intelligence (AI)

and natural language processing (NLP), particu-

larly with the introduction of large language models

(LLMs) such as BERT (Devlin et al., 2019), have

transformed the landscape of recommender systems.

LLMs have the capability to capture semantic nu-

ances in textual data, making them ideal for matching

course descriptions with user preferences and com-

petency proﬁles. These models, pre-trained on vast

amounts of multilingual text, allow for the devel-

opment of AI-based recommender systems that sur-

pass traditional keyword-based matching by leverag-

ing contextual understanding (Vaswani et al., 2017).

In this paper, we present an AI-based personalized

course recommender system for an educational plat-

form. Our system uses the DigComp framework to

assess user competency proﬁles and utilizes a ﬁne-

Dutta, S., Beier, F. and Wer th, D.

AI-Based Personalized Multilingual Course Recommender System Using Large Language Models.

DOI: 10.5220/0013260100003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 1069-1076

ISBN: 978-989-758-737-5; ISSN: 2184-433X

1069

tuned BERT model to compute semantic similarity

between user input and course content. Furthermore,

we introduce a novel learning path generation method

that builds customized course sequences for users,

ensuring a progressive and effective learning experi-

ence. This approach represents a signiﬁcant improve-

ment over traditional recommender systems, address-

ing both the need for personalization and the chal-

lenge of competency-based learning in the digital age.

2 RELATED WORK

Research on recommender systems spans multiple

domains, including e-commerce, entertainment, and

education (Ricci et al., 2010; Adomavicius and

Tuzhilin, 2005). In the context of education, the need

for personalized course recommendations has driven

innovation in both traditional and AI-based methods.

This section outlines key advancements in three ar-

eas: traditional recommender systems, NLP-based

approaches, and competency-based learning systems.

2.1 Traditional Recommender Systems

Traditional recommender systems fall into three pri-

mary categories: collaborative ﬁltering, content-

based ﬁltering, and hybrid approaches. Collaborative

ﬁltering, one of the earliest approaches, relies on user-

item interaction data to recommend items based on

the preferences of similar users (Schafer et al., 2007;

Ekstrand et al., 2011). This technique has been widely

used in various domains but often struggles with the

cold-start problem, where insufﬁcient data on new

users or items reduces its effectiveness.

Content-based ﬁltering, by contrast, matches users

with items based on item attributes, such as textual de-

scriptions in the case of educational courses (Pazzani

and Billsus, 2007; Lops et al., 2011). This method al-

lows for more personalized recommendations by con-

sidering the speciﬁc features of each course, but it

tends to lack diversity and novelty in the recommen-

dations, often leading to overspecialization. Hybrid

systems, which combine collaborative ﬁltering and

content-based approaches, have been developed to

overcome these individual limitations, improving rec-

ommendation accuracy and coverage (Burke, 2002;

Burke, 2007; C¸ ano and Morisio, 2017).

2.2 Course Recommendation in

Educational Platforms

In educational platforms, course recommendation

systems have traditionally relied on simple keyword-

based matching techniques (Manouselis et al., 2012;

Lu et al., 2015). Rule-based systems that use algo-

rithms like TF-IDF and cosine similarity to compare

user queries with course descriptions are common

(Colchester et al., 2017; Murtaza et al., 2022). While

such methods provide basic semantic matching, they

often fail to capture the full complexity of user intent

or course content, leading to recommendations that

may not fully align with the learner’s needs (Anand

and Mobasher, 2003; Zhang et al., 2020). Moreover,

these systems do not account for the progression in

a learner’s knowledge or provide personalized learn-

ing paths, making them less effective in guiding users

through a structured learning journey.

Some platforms have incorporated domain-

speciﬁc taxonomies or ontologies to improve the

matching process. For example, educational ontolo-

gies may categorize courses by subject or level of dif-

ﬁculty, but these approaches are often rigid and do not

adapt dynamically to changes in user preferences or

competencies (Manouselis et al., 2012). Educational

platforms like Coursera (https://www.coursera.org/)

and edX (https://www.edx.org/) employ comprehen-

sive skills taxonomy and learning objectives based on

different frameworks like Bloom’s Taxonomy (Bloom

et al., 1956), the Skills Framework for Information

Age (Foundation, 2015), and the Skills Network (An-

derson, 2017) to map out competencies and skills.

2.3 Advances in NLP-Based

Recommender Systems

Recent advancements in natural language processing

(NLP), particularly with the development of large lan-

guage models (LLMs), have opened new possibilities

for course recommendation systems. BERT (Bidi-

rectional Encoder Representations from Transform-

ers) (Devlin et al., 2019), a pre-trained transformer

model, has proven especially effective in understand-

ing the semantic context of text, enabling more accu-

rate matching between user queries and course con-

tent. BERT-based models capture bidirectional con-

text, making them more suitable for tasks like seman-

tic similarity, text classiﬁcation, and information re-

trieval (Vaswani et al., 2017).

In education, LLMs have been applied to generate

personalized learning plans, taking into account user

input, course descriptions, and the user’s progress.

Studies (Sun et al., 2019; Wu et al., 2023) have

demonstrated that BERT-based models signiﬁcantly

improve the accuracy of recommendations compared

to traditional approaches by leveraging deeper con-

textual understanding (Zhou et al., 2018). More-

over, these models can support multilingual plat-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1070

forms, broadening the applicability of the recom-

mender system to users from diverse linguistic back-

grounds.

2.4 Competency-Based Learning and

Frameworks

Competency-based learning frameworks have gained

traction as a way to personalize education by focusing

on the learner’s skills and competencies rather than

the content alone. The European Union’s DigComp

framework, for example, outlines 21 key digital com-

petencies across ﬁve dimensions namely, information

and data literacy, communication and collaboration,

digital content creation, safety, and problem solving

(Ferrari et al., 2014). These frameworks enable edu-

cational platforms to map courses to speciﬁc compe-

tencies, providing a structured and targeted learning

experience for users.

Previous research (Justesen et al., 2019) has ex-

plored integrating competency frameworks into edu-

cational recommender systems, but many implemen-

tations are limited to matching courses based on pre-

deﬁned categories rather than dynamically analyzing

user competencies and needs. Our approach builds

upon this work by incorporating both the DigComp

framework and AI-based semantic analysis, offering

a more sophisticated method for matching user pro-

ﬁles with relevant course content.

2.5 AI-Driven Learning Path

Generation

Another emerging area in educational recommender

systems is the generation of personalized learning

paths. Traditional recommenders typically suggest a

list of courses without considering the order in which

learners should complete them. However, recent AI-

driven approaches are addressing this gap by using

LLMs to dynamically generate learning sequences

that align with the learner’s progress and goals (Zhou

et al., 2018). These systems can provide not only

course recommendations but also a structured path

that optimizes the learning experience by guiding

users through progressively advanced material.

In our work, we extend this concept by employ-

ing LLMs like the mistralai/Mistral-7B-Instruct-v0.2

model (Jiang et al., 2023) to generate personalized

learning paths that account for the user’s past learning

experiences and future goals. This approach enables

the creation of tailored, goal-oriented learning paths,

enhancing the overall user experience in navigating

complex educational ecosystems.

3 SYSTEM ARCHITECTURE

AND METHODOLOGY

The system we developed for personalized course rec-

ommendations is composed of several interconnected

components, each responsible for a speciﬁc aspect

of the recommendation and learning path generation

process. The architecture includes a Rasa (Bock-

lisch et al., 2017) intent-based chatbot for initial user

interaction, a user competency proﬁle based on the

EU DigComp framework, a rule-based recommender,

and an AI-based course recommendation engine us-

ing large language models (LLMs). In this section,

we detail the design and functioning of each of these

components.

Figure 1: Flowchart of the important components in the sys-

tem. (Zoom in to read).

3.1 User Interaction and Competency

Proﬁling

The ﬁrst step in the system is user interaction, fa-

cilitated by a Rasa intent-based chatbot (Bocklisch

et al., 2017). The chatbot prompts users to describe

what they are interested in learning. By analyzing

the user’s responses using intent detection and entity

extraction techniques, the chatbot selects the relevant

competency areas from the DigComp framework that

the user should focus on. This process ensures that

the user receives recommendations aligned with their

learning needs and goals. This is a key advantage

for users who may not be familiar with the DigComp

competences, as it guides them through a tailored se-

lection process.

Once the relevant competency areas are identiﬁed,

the chatbot asks the user to complete corresponding

surveys based on these competencies. The surveys

are structured on a Likert scale, allowing users to self-

assess their proﬁciency in various competency areas.

The results of these surveys are then used to create

a user competency proﬁle, represented as a 21-length

vector corresponding to the 21 DigComp competency

areas (Ferrari et al., 2014).

Rasa was chosen for its customizable, open-

source framework that allows detailed intent recog-

AI-Based Personalized Multilingual Course Recommender System Using Large Language Models

1071

Figure 2: Flowchart showing the decision-making of the

chatbot (in German). (Zoom in to read).

Figure 3: Chatbot interface (in German) providing user as-

sistance with selected competences.

nition and seamless integration with external data

sources. Since Rasa supports nuanced conversational

ﬂows, it was ideal for building a bot that could effec-

tively guide users through selecting their competen-

cies, improving the onboarding process and ensuring

accurate proﬁling. Again, the DigComp framework

was chosen because it is well-established for access-

ing digital competencies and is structured into 21 ar-

eas, allowing for a granular approach to competency-

based learning. By using DigComp, our system

can provide recommendations closely aligned with

industry-recognized competencies, which increases

relevance for both users and educational providers.

Figure 4: Survey form (in German) presented to the user for

the Information and Data competence. (Zoom in to read).

Figure 5: User competence proﬁle (in German) shows the

overall competence levels of the user with a guide.

3.2 Course Data Collection and

Annotation

The system’s course database comprises approxi-

mately 800 courses from various online providers,

with 374 of them manually annotated. These courses

follow the MoocHub data schema (https://moochub.

org/), which includes metadata such as course ti-

tle, description, and provider information. However,

these courses do not have DigComp-related anno-

tations, requiring an additional step for competency

mapping.

To bridge this gap, we employed ChatGPT based

on the GPT-4 model (Radford et al., 2019; Brown

et al., 2020) to annotate the 374 courses with difﬁ-

culty levels (on a scale of 1 to 5) and competency

areas from the DigComp framework. The annotation

process involved extracting keywords from course de-

scriptions and assigning appropriate difﬁculty ratings

and competency areas based on the course content.

This use of language models enabled efﬁcient and

consistent annotation across multiple providers, en-

suring that the courses could be aligned with user

competency proﬁles for personalized recommenda-

tions. ChatGPT based on GPT-4 is used here for

annotation because of its advanced natural language

understanding and adaptability in generating contex-

tually relevant annotations. Without access to do-

main experts or teachers for manual annotation, GPT-

4 provides an efﬁcient alternative that leverages ex-

tensive pre-trained knowledge to identify course difﬁ-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1072

culty and competencies accurately. This choice al-

lows for high-quality annotation at scale, address-

ing the need for a robust, consistent annotation pro-

cess that would otherwise require considerable hu-

man expertise and resources. GPT-4’s capabilities

ensure consistent labeling across a large volume of

courses, which is essential for maintaining annota-

tion quality when scaling up the system. Addition-

ally, using an automated model minimizes the time

and cost associated with human annotation, making

it possible to achieve comprehensive coverage across

all courses without delays. This approach supports

our goal of quickly developing a personalized rec-

ommendation system based on accurately classiﬁed

and well-annotated course data. GPT-4’s capacity for

contextual understanding allows it to map course con-

tent to the DigComp competency areas effectively,

ensuring that recommendations align well with the

skills deﬁned in this widely accepted framework. This

alignment is key to ensuring that each recommenda-

tion supports relevant skill-building, which strength-

ens the educational value of the system for users.

3.3 Course Recommender System

Initially, a rule-based recommender system was im-

plemented to offer basic course recommendations

based on user-entered search queries. This sys-

tem used term frequency-inverse document frequency

(TF-IDF) and cosine similarity to calculate semantic

similarity between the search query and course de-

scriptions. A weighted score was computed for each

course, which determined its rank in the recommen-

dation list (Sch

utze et al., 2008). Although effective

for simple query matching, the rule-based approach

lacked the capability to adapt to individual user com-

petencies and did not fully leverage semantic infor-

mation embedded in course content.

To improve recommendation quality and incor-

porate personalization, an AI-based recommender

system was developed using the pretrained Distil-

BERT (“distiluse-base-multilingual-cased-v1”) (De-

vlin et al., 2019) model from HuggingFace. BERT’s

transformer architecture allows it to capture deep con-

textual relationships within text, making it ideal for

calculating semantic similarity between user queries

and course descriptions. We trained course embed-

dings on the course titles and descriptions for all

374 annotated courses and stored them in a search-

able database. DistilBERT, especially its multilingual

variant, offers a powerful yet computationally efﬁ-

cient approach for text embeddings, ideal for real-

time course recommendation scenarios. Since the

system needs to handle both English and German in-

puts, a multilingual transformer is necessary to ensure

consistent quality across languages. DistilBERT’s

lightweight architecture provides an optimal balance

between model performance and computational efﬁ-

ciency.

When a user enters a search query, the system cal-

culates the semantic similarity between the query and

the course embeddings, selecting only those courses

that meet a predeﬁned threshold of 78%. Courses

above this threshold are then ranked based on the

reverse Euclidean distance between the user’s 21-

length competency vector and the annotated compe-

tency vector for each course. This step ensures that

the recommendations are not only semantically rele-

vant but also aligned with the user’s competency pro-

ﬁle. The hybrid scoring model addresses limitations

in both purely semantic and purely rule-based sys-

tems by combining contextual similarity with person-

alized competency alignment. This approach ensures

that recommendations aren’t only relevant in terms

of content but are also tailored to each user’s skill

level, increasing the likelihood that recommendations

will be meaningful and actionable for the user. The

threshold of 78% is empirically chosen to strike a bal-

ance between relevance and inclusivity in recommen-

dations. This level is set based on preliminary testing

to ensure that users receive high-quality suggestions

without overly limiting course options, allowing for a

more diverse set of learning opportunities.

Figure 6: Course catalog (in German) showcasing the list of

available courses. (Zoom in to read).

3.4 Personalized Learning Path

Generation

In addition to recommending individual courses, the

system generates personalized learning paths con-

taining a sequence of 2–3 courses. After select-

ing the ﬁrst course from the AI-based recommender,

the system generates a customized prompt for

the mistralai/Mistral-7B-Instruct-v0.2 model (Jiang

et al., 2023) from HuggingFace. The Mistral-7B-

Instruct model is chosen for its capability to gen-

erate contextually relevant prompts that guide users

AI-Based Personalized Multilingual Course Recommender System Using Large Language Models

1073

to their next steps in learning. By leveraging this

model’s generative capacity, the system can create a

custom, adaptive learning path that considers users’

past courses. This approach aligns with the goal of

providing not only relevant but also sequential learn-

ing recommendations. The prompt includes informa-

tion about the user’s interests, previously completed

courses, and goals, requesting the model to suggest

the next step in the user’s learning journey (Zhou

et al., 2018).

Prompt: “A person searches for [user search input].

The person has taken the following courses: [course

names with course descriptions].”

Question: “Write in a paragraph which topics this

person should learn next.”

The model’s output provides key topics or skills

for the next course, which is used as input to the

AI recommender. This process repeats until the sys-

tem constructs a complete learning path, allowing

users to follow a structured, goal-oriented sequence

of courses.

Figure 7: Flow of information in the learning path genera-

tion. (Zoom in to read).

Figure 8: The learning path “Artiﬁcial Intelligence and Ma-

chine Learning” for a user proﬁle. (Zoom in to read).

The system architecture integrates various compo-

nents—from a chatbot-driven user interface to ad-

vanced AI-based recommendation models—to de-

liver personalized course suggestions and learning

paths. The system’s ability to dynamically generate

competency-aligned recommendations and tailored

learning paths represents a signiﬁcant advancement

over traditional rule-based recommenders, offering a

more engaging and effective user experience.

4 FUTURE WORK

While the current system leverages content-based

methods and language models to provide personal-

ized recommendations, the recommendation quality

could be enhanced by integrating a collaborative ﬁl-

tering approach. Collaborative ﬁltering has been

widely adopted in recommendation systems to iden-

tify patterns in user behavior, recommending items

based on the preferences of similar users (Koren et al.,

2009). Apart from real datasets, simulated datasets

including user data can be used to initially validate

the collaborative ﬁltering algorithm (Herlocker et al.,

2004).

Collaborative ﬁltering has the potential to com-

plement the existing content-based recommendation

system by incorporating user-item interactions, which

will help overcome limitations like the “cold start”

problem inherent in purely content-based systems

(Schein et al., 2002). By leveraging user similari-

ties, courses can be recommended based not only on

course content and competency proﬁles but also on

the learning patterns of other users with similar inter-

ests and skill levels. This hybrid system—combining

content-based ﬁltering, AI-based matching, and col-

laborative ﬁltering—has the potential to signiﬁcantly

improve recommendation relevance and engagement

(Burke, 2002; Burke, 2007).

Using real user data at a later time point, the col-

laborative ﬁltering model can be ﬁne-tuned to work

with live data. User feedback can also be inte-

grated into the system, allowing for further reﬁnement

of recommendations based on explicit (e.g., course

ratings) and implicit (e.g., click-through rates) sig-

nals. Additionally, reinforcement learning techniques

could contribute to continuously adapt the learning

paths based on user progress and outcomes (Zheng

and Wang, 2022).

Beyond the education sector, the personalized rec-

ommender and adaptive learning path approach de-

veloped in this project are important assets which

hold signiﬁcant potential for other domains requir-

ing tailored content delivery and skill progression. In

corporate training, such a system could guide em-

ployees through customized learning paths aligned

with career goals, role requirements, or skill gaps,

ensuring that development resources are both rele-

vant and impactful. Similarly, in healthcare, this

approach could support personalized patient educa-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1074

tion by recommending articles, videos, or courses

tailored to individual health conditions or treatment

plans, thereby enhancing patient engagement and ad-

herence to health protocols. Additionally, in sectors

like e-commerce, personalized recommenders could

suggest products or services based on past purchases

or browsing behavior, while an adaptive path model

could guide customers through complementary prod-

ucts or bundles in a curated sequence. The ﬂexibil-

ity and contextual adaptability of this recommender

system make it valuable across various ﬁelds where

user-speciﬁc recommendations enhance engagement

and satisfaction.

5 CONCLUSION

In this paper, we have presented an AI-based per-

sonalized course recommender system grounded in

the EU DigComp competency framework. By using

a combination of natural language processing tech-

niques, large language models, and semantic similar-

ity algorithms, our system provides tailored course

recommendations based on users’ competencies and

interests. Additionally, a learning path generation

module offers structured course sequences, further

enhancing the personalized learning experience. The

integration of a Rasa chatbot allows for an intuitive

and interactive user interface, improving engagement

by guiding users through competency-based assess-

ments. The annotation of courses with DigComp

competency areas, facilitated by LLMs, ensures that

the recommendations are competency-aligned and

relevant to individual learning goals.

As the project progresses, we plan to incorporate

collaborative ﬁltering algorithms to augment the rec-

ommendation engine. By leveraging both simulated

and real user data, we aim to create a hybrid system

that combines the strengths of content-based and col-

laborative ﬁltering techniques. Ultimately, this sys-

tem will enable more effective and personalized ed-

ucational experiences, catering to a wide variety of

learners and their evolving needs.

ACKNOWLEDGEMENTS

Both Sourav Dutta and Florian Beier are supported

by the “Bundesministerium f

ur Wirtschaft und Kli-

maschutz” within the project “MERLOT” which was

funded under the project reference 68GX21008K.

REFERENCES

Adomavicius, G. and Tuzhilin, A. (2005). Toward the

next generation of recommender systems: A sur-

vey of the state-of-the-art and possible extensions.

IEEE transactions on knowledge and data engineer-

ing, 17(6):734–749.

Anand, S. S. and Mobasher, B. (2003). Intelligent tech-

niques for web personalization. In IJCAI Workshop on

Intelligent Techniques for Web Personalization, pages

1–36. Springer.

Anderson, K. A. (2017). Skill networks and measures of

complex human capital. Proceedings of the National

Academy of Sciences, 114(48):12720–12724.

Bloom, B. S., Engelhart, M. D., Furst, E. J., Hill, W. H., and

Krathwohl, D. R. (1956). Taxonomy of educational

objectives: Cognitive and affective domains. New

York: David McKay, pages 20–24.

Bocklisch, T., Faulkner, J., Pawlowski, N., and Nichol,

A. (2017). Rasa: Open source language under-

standing and dialogue management. arXiv preprint

arXiv:1712.05181.

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,

Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,

Askell, A., Agarwal, S., Herbert-Voss, A., Krueger,

G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.,

Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E.,

Litwin, M., Gray, S., Chess, B., Clark, J., Berner,

C., McCandlish, S., Radford, A., Sutskever, I., and

Amodei, D. (2020). Language models are few-shot

learners. In Advances in Neural Information Process-

ing Systems, volume 33, pages 1877–1901.

Burke, R. (2002). Hybrid recommender systems: Survey

and experiments. User modeling and user-adapted in-

teraction, 12:331–370.

Burke, R. (2007). Hybrid web recommender systems. The

adaptive web: methods and strategies of web person-

alization, pages 377–408.

C¸ ano, E. and Morisio, M. (2017). Hybrid recommender

systems: A systematic literature review. Intelligent

data analysis, 21(6):1487–1524.

Colchester, K., Hagras, H., Alghazzawi, D., and Aldab-

bagh, G. (2017). A survey of artiﬁcial intelligence

techniques employed for adaptive educational systems

within e-learning platforms. Journal of Artiﬁcial In-

telligence and Soft Computing Research, 7(1):47–64.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2019). BERT: Pre-training of deep bidirectional

transformers for language understanding. In Proceed-

ings of the 2019 Conference of the North American

Chapter of the Association for Computational Lin-

guistics: Human Language Technologies, Volume 1

(Long and Short Papers).

Ekstrand, M. D., Riedl, J. T., Konstan, J. A., et al. (2011).

Collaborative ﬁltering recommender systems. Foun-

dations and Trends® in Human–Computer Interac-

tion, 4(2):81–173.

Ferrari, A., Bre

cko, B. N., and Punie, Y. (2014). Digcomp:

a framework for developing and understanding digital

competence in europe. eLearning Papers, (38):1.

AI-Based Personalized Multilingual Course Recommender System Using Large Language Models

1075

Foundation, S. (2015). Sﬁa6 the complete reference guide.

Herlocker, J. L., Konstan, J. A., Terveen, L. G., and Riedl,

J. T. (2004). Evaluating collaborative ﬁltering recom-

mender systems. ACM Transactions on Information

Systems (TOIS), 22(1):5–53.

Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C.,

Chaplot, D. S., Casas, D. d. l., Bressand, F., Lengyel,

G., Lample, G., Saulnier, L., et al. (2023). Mistral 7b.

arXiv preprint arXiv:2310.06825.

Justesen, N., Bontrager, P., Togelius, J., and Risi, S. (2019).

Deep learning for video game playing. IEEE Trans-

actions on Games, 12(1):1–20.

Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factor-

ization techniques for recommender systems. Com-

puter, 42(8):30–37.

Lops, P., De Gemmis, M., and Semeraro, G. (2011).

Content-based recommender systems: State of the art

and trends. Recommender systems handbook, pages

73–105.

Lu, J., Wu, D., Mao, M., Wang, W., and Zhang, G. (2015).

Recommender system application developments: a

survey. Decision support systems, 74:12–32.

Manouselis, N., Drachsler, H., Verbert, K., and Duval, E.

(2012). Recommender systems for learning. Springer

Science & Business Media.

Murtaza, M., Ahmed, Y., Shamsi, J. A., Sherwani, F., and

Usman, M. (2022). Ai-based personalized e-learning

systems: Issues, challenges, and solutions. IEEE ac-

cess, 10:81323–81342.

Pappano, L. (2012). The year of the mooc. The New York

Times, November, 2.

Pazzani, M. J. and Billsus, D. (2007). Content-based rec-

ommendation systems. In The adaptive web: methods

and strategies of web personalization, pages 325–341.

Springer.

Radford, A., Wu, J., Child, R., Luan, D., Amodei, D.,

Sutskever, I., Dean, J., and Ghemawat, S. (2019). Lan-

guage models are unsupervised multitask learners. In

OSDI’04: Sixth Symposium on Operating System De-

sign and Implementation, pages 137–150.

Ricci, F., Rokach, L., and Shapira, B. (2010). Introduction

to recommender systems handbook. In Recommender

systems handbook, pages 1–35. Springer.

Schafer, J. B., Frankowski, D., Herlocker, J., and Sen, S.

(2007). Collaborative ﬁltering recommender systems.

In The adaptive web: methods and strategies of web

personalization, pages 291–324. Springer.

Schein, A. I., Popescul, A., Ungar, L. H., and Pennock,

D. M. (2002). Methods and metrics for cold-start rec-

ommendations. In Proceedings of the 25th annual in-

ternational ACM SIGIR conference on Research and

development in information retrieval, pages 253–260.

Sch

utze, H., Manning, C. D., and Raghavan, P. (2008). In-

troduction to information retrieval, volume 39. Cam-

bridge University Press Cambridge.

Sun, F., Liu, J., Wu, J., Pei, C., Lin, X., Ou, W., and

Jiang, P. (2019). Bert4rec: Sequential recommenda-

tion with bidirectional encoder representations from

transformer. In Proceedings of the 28th ACM interna-

tional conference on information and knowledge man-

agement, pages 1441–1450.

Vaswani, A., Shazeer, N. M., Parmar, N., Uszkoreit, J.,

Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin,

I. (2017). Attention is all you need. In Neural Infor-

mation Processing Systems.

Wu, C., Wu, F., Huang, Y., and Xie, X. (2023). Personal-

ized news recommendation: Methods and challenges.

ACM Transactions on Information Systems, 41(1):1–

50.

Zhang, Y., Chen, X., et al. (2020). Explainable recommen-

dation: A survey and new perspectives. Foundations

and Trends® in Information Retrieval, 14(1):1–101.

Zheng, Y. and Wang, D. X. (2022). A survey of rec-

ommender systems with multi-objective optimization.

Neurocomputing, 474:141–153.

Zhou, Y., Huang, C., Hu, Q., Zhu, J., and Tang, Y.

(2018). Personalized learning full-path recommenda-

tion model based on lstm neural networks. Informa-

tion sciences, 444:135–152.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1076