International Journal on Natural Language Computing (IJNLC) Vol.14, No.

1, February 2025

ENHANCING EDUCATIONAL QA SYSTEMS:

INTEGRATING KNOWLEDGE GRAPHS AND LARGE
LANGUAGE MODELS FOR CONTEXT-AWARE LEARNING
Ayse Arslan

USA

ABSTRACT
This study explores the integration of Knowledge Graphs (KGs) and Large Language Models (LLMs) to develop an
advanced question-answering (QA) system for educational purposes. The proposed method involves constructing a
KG using LLMs, retrieving contextual prompts from high-quality learning resources, and enhancing these prompts
to generate accurate answers to complex educational queries.

The technical framework presented in this paper, along with the analysis of results, contributes significantly to the
advancement of LLM applications in educational technology. The findings provide a robust foundation for
developing intelligent, context-aware educational systems that leverage structured knowledge to support
personalized learning and improve educational outcomes.

1. INTRODUCTION
The rapid advancement of generative large language models (LLMs) has sparked considerable interest in
leveraging their potential to revolutionize interactive question-answering (QA) processes, enabling deeper
exploration of conceptual knowledge across various academic disciplines. Initial efforts to integrate
LLMs into educational contexts have primarily focused on supporting learner dialogues and delivering
automated feedback through sophisticated text analysis, facilitating tailored and adaptive learning
experiences. These applications span a wide range of educational tools, including procedural QA systems
and individualized learning pathways.

Despite these advancements, research on utilizing LLM-based QA systems specifically to enhance

learners' conceptual comprehension remains in its nascent stages. This study addresses this gap by
proposing and validating a comprehensive technical framework designed to develop and implement
conceptual QA systems powered by state-of-the-art LLMs.

1.2. Brief Overview of LLMs

GenAI technologies such as LLMs (Large Language Models) can analyze the complex patterns and
structures of human language and generate human-like text and multimedia content.

LLMs have been described using a wide range of metaphors. While some emphasize its positive potential
as a supportive and empowering tool, likening it to a copilot (Risteff, 2023), a sorcerer’s apprentice (Liu
& Helmer, 2024), a form of co-intelligence (Mollick & Mollick, 2024), or an external brain (Yan et al.,
2024), others adopt a more cautious view that acknowledges both its promise and potential risks,

DOI: 10.5121/ijnlc.2025.14101 1
 International Journal on Natural Language Computing (IJNLC) Vol.14, No.1, February 2025
describing it as a double-edged sword (Furze, 2024), a kind of magic (Furze, 2024), or a powerful dragon
(Bozkurt, 2024a).

On the critical side, some have regarded LLMs as an autotune for knowledge (Cormier, 2023), a
colonizing loudspeaker (Gupta et al., 2024), a stochastic parrot (Bender et al., 2021), a dangerous “alien”
decision maker (Harari, 2024) or even a weapon of mass destruction (Maas, 2023).

Before delving into their specific use for QA systems, it might be useful to provide an overview of the
their general architecture and finetuning mechanism.

The most common ways to finetune language models are instruction finetuning and classification
finetuning (Raschka, 2024). Instruction finetuning involves training a language model on a set of tasks
using specific instructions to improve its ability to understand and execute tasks described in natural
language prompts.

In classification finetuning, the model is trained to recognize a specific set of class labels, such as "spam"
and "not spam." Examples of classification tasks extend beyond large language models and email
filtering; they include identifying different species of plants from images, categorizing news articles into
topics like sports, politics, or technology, and distinguishing between benign and malignant tumors in
medical imaging.

The key point is that a classification-finetuned model is restricted to predicting classes it has encountered
during its training—for instance, it can determine whether something is "spam" or "not spam”, but it can't
say anything else about the input text.

Classification finetuning is ideal for projects requiring precise categorization of data into predefined
classes, such as sentiment analysis or spam detection (Raschka, 2024).

In contrast to the classification-finetuned model, an instruction-finetuned model typically has the

capability to undertake a broader range of tasks. We can view a classification-finetuned model as highly
specialized, and generally, it is easier to develop a specialized model than a generalist model that works
well across various tasks.

Instruction finetuning improves a model's ability to understand and generate responses based on specific
user instructions. Instruction finetuning is best suited for models that need to handle a variety of tasks
based on complex user instructions, improving flexibility and interaction quality.

While instruction finetuning is more versatile, it demands larger datasets and greater computational
resources to develop models proficient in various tasks. In contrast, classification finetuning requires less
data and compute power, but its use is confined to the specific classes on which the model has been
trained (Raschka, 2024).

Once a model is initialized with pretrained weights, a modification is required to transform a general
pretrained LLM into a specialized LLM for classification tasks. So, it is necessary to modify the
pretrained large language model to prepare it for classification finetuning.

It should be taken into account that it's not necessary to finetune all model layers. This is because, in
neural network-based language models, the lower layers generally capture basic language structures and
semantics that are applicable across a wide range of tasks and datasets. So, finetuning only the last layers
2
 International Journal on Natural Language Computing (IJNLC) Vol.14, No.1, February 2025
(layers near the output), which are more specific to nuanced linguistic patterns and task-specific features,
can often be sufficient to adapt the model to new tasks (Raschka, 2024).

2. A REVIEW OF RELATED WORK

The duality of generative AI—a mix of enthusiasm and skepticism—is well-documented in contemporary
literature [Lim et al., 2023; Stracke et al., 2024]. Advocates highlight its transformative potential in
generating human-like responses, while critics caution against issues such as misinformation and ethical
concerns. In educational contexts, QA systems leveraging LLMs aim to provide guided learning
experiences. However, challenges such as hallucinations, limited domain expertise, and lack of
explainability hinder their widespread adoption.

3. QA SYSTEMS AND LEARNING APPLICATIONS

QA systems designed for educational purposes typically integrate natural language understanding with
machine learning, enabling precise intent decoding, knowledge retrieval, and coherent response
generation. These systems align with pedagogical theories such as Vygotsky’s Zone of Proximal
Development (ZPD) and the concept of scaffolding, which emphasize structured support for learners to
achieve tasks beyond their independent capabilities.

Key components of QA systems include:

1. Question Analysis: Understanding the intent behind queries.

2. Information Retrieval: Accessing relevant knowledge from databases or KGs.
3. Answer Generation: Crafting contextually appropriate responses using LLMs.

Scaffolding in QA systems involves:

 Shared Understanding: Aligning learning objectives.

 Dynamic Assistance: Providing context-specific guidance.
 Progressive Independence: Reducing support as learners advance.

Despite their advantages, QA systems must address challenges such as LLM-induced hallucinations,
inadequate domain-specific knowledge, and the risk of cognitive overreliance on AI.

3.2. Retrieval-Augmented Generation (RAG): A Promising Solution

To overcome these limitations, Retrieval-Augmented Generation (RAG) combines external knowledge

sources, like KGs, with LLM workflows. This integration ensures higher accuracy and contextual
relevance by embedding retrieved knowledge into response generation.

Knowledge graphs (KGs) are pivotal in RAG implementations, offering semantic networks that map
relationships among concepts.

These structured representations enhance both the precision and depth of QA systems. KGs are built by
extracting entities and relationships from educational resources using natural language processing
techniques.
3
 International Journal on Natural Language Computing (IJNLC) Vol.14, No.1, February 2025

The effectiveness of RAG in educational contexts can be evaluated through two primary perspectives:

1. Semantic Similarity: Measuring the alignment between the retrieved knowledge and the context
of the generated response.
2. Contextual Groundedness: Assessing the relevance and coherence of the generated content
within the learning domain.

By leveraging these frameworks, researchers aim to optimize QA systems, ensuring that LLMs serve as
reliable, effective tools for enhancing learning experiences while mitigating their inherent limitations.

3.3. QA System Based on RAG and LLM

The integration of Retrieval-Augmented Generation (RAG) and Large Language Models (LLMs) in
Question-Answering (QA) systems represents a transformative approach to embedding disciplinary
knowledge into intelligent question-response mechanisms. This integration significantly enhances
learners' efficiency in self-directed study by optimizing the limited availability of their time and cognitive
resources. Notably, LLMs have been extensively employed in program tutoring and problem-solving
applications, including automated question-answering systems and knowledge recommendation platforms
[9], [10].

State-of-the-art natural language processing (NLP) models, such as BERT, Llama, and GPT, are
foundational in the development of open-domain chatbots. These models, initially trained on expansive
general-domain datasets, can be fine-tuned to specific tasks, thereby aligning more closely with the
nuanced requirements of their intended applications [8], [3]. Among these, advanced LLMs like GPT-4
and Llama-2 and -3 have garnered widespread recognition for their ability to generate coherent,
contextually appropriate text, underlining their profound impact on NLP advancements.

From a theoretical perspective, RAG enhances the reliability of text generated by LLMs by retrieving
information from external knowledge sources. This reliability—often referred to as "faithfulness"—is
assessed through the semantic correlation between the LLM-generated output and accurate reference
texts, with comparisons frequently drawn against responses provided by human tutors. Empirical
evidence suggests that RAG can significantly improve the performance of LLM-based QA systems in
educational contexts.

For instance, a QA system utilizing the Llama-3 model, which applies cosine similarity within the RAG
framework for text vector retrieval, demonstrated a 9.84% increase in accuracy on a test set comprising
non-graphical multiple-choice questions across various STEM disciplines [8].

The RAG framework is typically implemented using one of three primary methodologies:

1. Template-Based Retrieval: This approach employs keyword-matching algorithms to identify

and retrieve relevant textual materials [2].
2. Semantic Vector Retrieval: By extracting high-dimensional matrices (semantic vectors) through
NLP models, this widely adopted method leverages its simplicity and effectiveness for similarity-
based retrieval [8], [3].
3. Knowledge Graph-Based Retrieval: Distinct from text-based retrieval, this method organizes
data in a graph structure, forming a knowledge base represented as interconnected nodes (entities)
and edges (relations). Subgraphs are retrieved using semantic vector-based methods [2].
4
 International Journal on Natural Language Computing (IJNLC) Vol.14, No.1, February 2025

Knowledge graphs (KGs) are constructed through the assembly of "triples," in which a head entity is
linked to a tail entity via a predicate that defines their relationship [12]. These triples coalesce into a
multi-graph framework, with nodes representing entities and edges corresponding to relationships. NLP
techniques facilitate the extraction of knowledge entities and their interrelations from extensive
educational resources, enabling the creation of knowledge graphs as external knowledge repositories.

The construction of KGs in the educational domain has evolved from early rule-based and lexical
methods to modern machine learning and deep learning-based approaches. While lexical and rule-based
techniques rely on manual rule formulation by domain experts—limiting scalability—contemporary
methods leverage statistical models and neural architectures to manage vast datasets efficiently. For
example, convolutional neural networks (CNNs) have been used to extract KGs from MOOC resources,
including curricula and textbooks, to support instructional design and provide tailored course
recommendations [5].

Knowledge graphs further enable the elucidation of complex concepts by linking foundational ideas to
advanced topics through structured entity relationships. This interconnected representation facilitates a
progressive learning experience, allowing students to comprehend sophisticated subjects with greater ease
[4].

Once a knowledge graph is integrated into a QA system, the system can accurately respond to user queries
by analyzing the questions and retrieving relevant entities and their interconnections from the knowledge
base. The process involves extracting keywords, entities, and relationships from textual documents, often
guided by LLM-generated prompts. These entities and relationships are converted into text embeddings,
which are then used to search for related entities. If the similarity threshold (e.g., a cosine similarity of
0.8) is unmet or the entity appears for the first time, separate subgraphs are generated to maintain
coherence and integrity within the graph.

This workflow, illustrated in Figure 1, underscores the pivotal role of KG construction and retrieval in
enhancing QA system capabilities, thereby advancing the intersection of NLP, knowledge representation,
and educational technology.

5
 International Journal on Natural Language Computing (IJNLC) Vol.14, No.1, February 2025

Figure 1. KG Workflow

Relevant relationships within the knowledge graph (KG) are incorporated into the subgraph of the target
entity based on the required relational context as determined by the Large Language Model (LLM). At
this stage, the LLM primarily oversees the merging of entities and relationships. Each subgraph comprises
multiple foundational relationships, with each relationship defined as a pair of entities and their associated
connection. The process of merging entities and relationships relies on well-established NLP
disambiguation and extraction algorithms, which are fundamental tasks in the domain of computer
science.

In the retrieval query workflow, illustrated in the bottom-right corner of the diagram, the learner’s query
is processed to compute cosine similarity between its semantic vector and the semantic vectors of each
entity in the KG. This step identifies the entity most closely aligned with the query. Subsequently, the
system retrieves the most relevant relationships within the subgraph of the identified entity. This is
achieved by evaluating the similarity between the semantic vectors of entity pairs and relationships and
the query vector, using a predefined similarity threshold. The extracted relationships are then
concatenated to form a coherent context, which is input into the LLM to generate the KG-based response.

Finally, the learner’s query and the contextual output from the KG (i.e., the retrieved information) are
jointly fed into the LLM. The LLM synthesizes this information to produce a comprehensive and
contextually accurate response to the learner's question, ensuring that the answer is both relevant and
grounded in the structured knowledge within the graph.

4. RECOMMENDATIONS

6
 International Journal on Natural Language Computing (IJNLC) Vol.14, No.1, February 2025
The integration of Generative AI (GenAI) into the data-to-wisdom continuum marks a transformative
shift in the dynamics of human wisdom, with synthetic information emerging as a significant counterpart
to human-generated organic information. While GenAI holds the potential to expand cognitive horizons
and facilitate complex tasks, its role must be carefully delineated to avoid fostering over-dependence. The
balance between support and over-reliance is particularly critical in educational contexts, where the
objective is to cultivate independent intellectual growth alongside technological augmentation.

GenAI’s capacity to revolutionize interdisciplinary learning lies in its ability to synthesize and integrate
knowledge across domains. However, such potential can only be realized if learners are guided toward
critical synthesis, rather than passive consumption of AI-generated outputs. Critical reflection is
indispensable; without it, learners risk accepting AI-derived connections at face value, thereby
undermining the deeper understanding that interdisciplinary learning demands

5. CONCLUSION
This study investigates the application of a Knowledge Graph (KG)-based Retrieval-Augmented
Generation (RAG) question-answering (QA) system, incorporating advanced retrieval enhancement
techniques to produce high-quality responses tailored for conceptual learning in educational contexts.

The technical framework and analytical findings presented herein offer valuable contributions to the
fields of conversational AI, intelligent tutoring systems, and Large Language Model (LLM) research. By
addressing the challenges of hallucination and domain-specific knowledge limitations, this approach
paves the way for more reliable and context-aware educational AI assistants.

Future research directions could explore the development of hybrid models that effectively merge the
natural language understanding capabilities of logical reasoning models with the structured knowledge
representation inherent in knowledge bases. Such advancements would further empower educational
systems to not only generate human-like text but also retrieve and reason with structured knowledge,
ultimately enhancing the learning experience and outcomes for students across various disciplines.

REFERENCES

[1] Ansari, A. N., Ahmad, S., & Bhutta, S. M. (2024). Mapping the global evidence around the use of ChatGPT
in higher education: A systematic scoping review. Education and Information Technologies, 29(9), 11281–
11321. https://doi.org/10.1007/s10639-023-12223-4
[2] Bayne, S., Evans, P., Ewins, R., Knox, J., Lamb, J., Macleod, H., O’Shea, C., Ross, J., Sheail, P., Sinclair, C.,
& Johnston, K. (2020). The manifesto for teaching online. MIT
Press. https://doi.org/10.7551/mitpress/11840.001.0001
[3] Bayne, S., & Ross, J. (2016). Manifesto redux: Making a teaching philosophy from networked learning
research. In S. Cranmer, N. B. Dohn, M. de Laat, T. Ryberg, & J. A. Sime (Eds.), Proceedings of the 10th
International Conference on Networked Learning 2016 (pp. 210–128). Lancaster: University of
Lancaster. https://doi.org/10.54337/nlc.v10.8821
[4] Bender, E. M., Gebru, T., McMillan-Major, A., &Shmitchell, S. (2021, March). On the dangers of stochastic
parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness,
accountability, and transparency (pp. 610–623). https://doi.org/10.1145/3442188.3445922
[5] Birt, L., Scott, S., Cavers, D., Campbell, C., & Walter, F. M. (2016). Member checking: A tool to enhance
trustworthiness or merely a nod to validation? Qualitative Health Research, 26(13), 1802–
1811. https://doi.org/10.1177/1049732316654870
[6] Bozkurt, A., Xiao, J., Lambert, S., Pazurek, A., Crompton, H., Koseoglu, S., Farrow, R., Bond, M., Nerantzi,
C., Honeychurch, S., Bali, M., Dron, J., Mir, K., Stewart, B., Costello, E., Mason, J., Stracke, C. M., Romero-
7
 International Journal on Natural Language Computing (IJNLC) Vol.14, No.1, February 2025
Hall, E., Koutropoulos, A., Toquero, C. M., Singh, L., Tlili, A., Lee, K., Nichols, M., Ossiannilsson, E.,
Brown, M., Irvine, V., Raffaghelli, J. E., Santos-Hermosa, G. Farrell, O., Adam, T., Thong, Y. L., Sani-
Bozkurt, S., Sharma, R. C., Hrastinski, S., & Jandrić, P. (2023a). Speculative futures on ChatGPT and
generative artificial intelligence (AI): A collective reflection from the educational landscape. Asian Journal of
Distance Education, 18(1), 53–130. https://doi.org/10.5281/zenodo.7636568
[7] Burns, T., Sinfield, S., & Abegglen, S. (2023). Postdigital academic writing. In Encyclopedia of Postdigital
Science and Education (pp. 1–7). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-35469-
4_27-1
[8] Clayton, M. J. (1997). Delphi: a technique to harness expert opinion for critical decision-making tasks in
education. Educational Psychology, 17(4), 373–386. https://doi.org/10.1080/0144341970170401
[9] Cormier, D. (2023, Jan 20). ChatGPT search – Autotune for knowledge. Dave’s Educational
Blog. https://davecormier.com/edblog/2023/01/20/chatgpt-search-autotune-for-knowledge/comment-page-1/
[10] Costello, E. (2024). ChatGPT and the educational AI chatter: Full of bullshit or trying to tell us
something? Postdigital Science and Education, 6(2), 425–430. https://doi.org/10.1007/s42438-023-00398-5
[11] Floridi, L. (2023). AI as agency without intelligence: On ChatGPT, large language models, and other
generative models. Philosophy & Technology, 36(1), 1–7. https://doi.org/10.1007/s13347-023-00621-y
[12] Furze, L. (2024, July 19). AI metaphors we live by: The language of artificial
intelligence. https://leonfurze.com/2024/07/19/ai-metaphors-we-live-by-the-language-of-artificial-
intelligence/
[13] Gale, K., &Bowstead, H. (2013). Deleuze and collaborative writing as method of inquiry. Journal of
Learning Development in Higher Education, 6. https://doi.org/10.47408/jldhe.v0i6.222
[14] Gupta, A., Atef, Y., Mills, A., & Bali, M. (2024). Assistant, parrot, or colonizing loudspeaker? ChatGPT
metaphors for developing critical AI literacies. Open Praxis, 16(1), 37–
53. https://doi.org/10.55982/openpraxis.16.1.631
[15] Harari, Y. N. (2024). Nexus: A brief history of information networks from the stone age to AI. Random House
Publishing Group. https://www.ynharari.com/book/nexus/
[16] Hicks, M. T., Humphries, J., & Slater, J. (2024). ChatGPT is bullshit. Ethics and Information Technology,
26(2), 38. https://doi.org/10.1007/s10676-024-09775-5
[17] Jandrić, P., Hayes, D., Truelove, I., Levinson, P., Mayo, P., Ryberg, T., Monzó, L. D., Allen, Q., Stewart, P.
A., Carr, P. R., Jackson, L., Bridges, S., Escaño, C., Grauslund, D., Mañero, J., Lukoko, H. O., Bryant, P.,
Fuentes-Martinez, A., Gibbons, A., … Hayes, S. (2020). Teaching in the Age of Covid-19. Postdigital
Science and Education, 2(3), 1069–1230. https://doi.org/10.1007/s42438-020-00169-6
[18] Jandrić, P., Luke, T. W., Sturm, S., McLaren, P., Jackson, L., MacKenzie, A., Tesar, M., Stewart, G. T.,
Roberts, P., Abegglen, S., Burns, T., Sinfield, S., Hayes, S., Jaldemark, J., Peters, M. A., Sinclair, C., &
Gibbons, A. (2023). Collective writing: The continuous struggle for meaning-making. Postdigital Science
and Education, 5(3), 851–893. https://doi.org/10.1007/s42438-022-00320-5
[19] Koutropoulos, A., Stewart, B., Singh, L., Sinfield, S., Burns, T., Abegglen, S., Hamon, K., Honeychurch, S.,
& Bozkurt, A. (2024). Lines of flight: The digital fragmenting of educational networks. Journal of Interactive
Media in Education, 2024(1), 1–13. https://doi.org/10.5334/jime.850
[20] Latour, B. (2010). An attempt at a “compositionist manifesto”. New Literary History, 41(3), 471–
490. https://doi.org/10.1353/nlh.2010.a408295
[21] Lim, W. M., Gunasekara, A., Pallant, J. L., Pallant, J. I., &Pechenkina, E. (2023). Generative AI and the
future of education: Ragnarök or reformation? A paradoxical perspective from management educators. The
International Journal of Management Education, 21(2), 100790. https://doi.org/10.1016/j.ijme.2023.100790
[22] Liu, D., & Helmer, E. (2024, Feb 9). The sorcerer’s apprentice: Applied AI for data. Perspectives,
Substack. https://debliu.substack.com/p/the-sorcerers-apprentice-applied
[23] Maas, M. M. (2023). AI is like… A literature review of AI metaphors and why they matter for policy. SSRN
Electronic Journal. https://doi.org/10.2139/ssrn.4612468
[24] Mack, E. (2014, Oct 26). Elon Musk: ‘We are summoning the demon’ with artificial
intelligence. CNET. https://www.cnet.com/science/elon-musk-we-are-summoning-the-demon-with-artificial-
intelligence/
[25] Merriam, S. B. (2001). Qualitative research and case study applications in education. Jossey-Bass.
[26] Mollick, E., & Mollick, E. (2024). Co-Intelligence. Random House UK.

8
 International Journal on Natural Language Computing (IJNLC) Vol.14, No.1, February 2025
[27] Pelletier, K., Brown, M., Brooks, D. C., McCormack, M., Reeves, J., Arbino, N., Bozkurt, A., Crawford, S.,
Czerniewicz, L., Gibson, R., Linder, K., Mason, J., & Mondelli, V. (2021). 2021 EDUCAUSE horizon report
teaching and learning edition. EDUCAUSE. https://www.learntechlib.org/p/219489/
[28] Peters, M. A., Besley, T., Tesar, M., Jackson, L., Jandrić, P., Arndt, S., & Sturm, S. (2021). The methodology
and philosophy of collective writing: An educational philosophy and theory reader volume X.
Routledge. https://doi.org/10.4324/9781003171959
[29] Risteff, M. (2023, Apr 24). AI is an extraordinary copilot when used responsibly. K-12
Dive. https://www.k12dive.com/spons/ai-is-an-extraordinary-copilot-when-used-responsibly/647644/
[30] Saban, A., Kocbeker, B. N., & Saban, A. (2007). Prospective teachers’ conceptions of teaching and learning
revealed through metaphor analysis. Learning and Instruction, 17(2), 123–
139. https://doi.org/10.1016/j.learninstruc.2007.01.003
[31] Sharples, M., & Pérez y Pérez, R. (2022). Story machines: How computers have become creative writers.
Routledge. https://doi.org/10.4324/9781003161431
[32] Slagter van Tryon, P. J., & Bishop, M. J. (2006). Identifying “e-mmediacy strategies for web-based
instruction: A Delphi study.” The Quarterly Review of Distance Education, 7(1) 49–62.
[33] Stracke, C. M., Burgos, D., Santos-Hermosa, G., Bozkurt, A., Sharma, R. C., Swiatek Cassafieres, C., dos
Santos, A. I., Mason, J., Ossiannilsson, E., Shon, J. G., Wan, M., Obiageli Agbu, J.-F., Farrow, R., Karakaya,
Ö., Nerantzi, C., Ramírez-Montoya, M. S., Conole, G., Cox, G., & Truong, V. (2022a). Responding to the
initial challenge of the COVID-19 pandemic: Analysis of international responses and impact in school and
higher education. Sustainability, 14(3), 1876. https://doi.org/10.3390/su14031876
[34] Stracke, C. M., Chounta, I.-A., & Holmes, W. (2024). Global trends in scientific debates on trustworthy and
ethical artificial intelligence and education. Artificial Intelligence in Education. Communications in
Computer and Information Science, 2150, 254–262. https://doi.org/10.1007/978-3-031-64315-6_21
[35] Tlili, A., Shehata, B., Adarkwah, M. A., Bozkurt, A., Hickey, D. T., Huang, R., & Agyemang, B. (2023).
What if the devil is my guardian angel: ChatGPT as a case study of using chatbots in education. Smart
Learning Environments, 10(1), 1–24. https://doi.org/10.1186/s40561-023-00237-x
[36] Yan, Y., Sun, W., & Zhao, X. (2024). Metaphorical conceptualizations of generative artificial intelligence use
by Chinese university EFL learners. Frontiers in Education, 9. https://doi.org/10.3389/feduc.2024.1430494
[37] Ziglio, E. (1996). The Delphi method and its contribution to decision-making. In M. Adler & E. Ziglio
(Eds.), Gazing into the oracle: The Delphi method and its application to social policy and public health (pp.
3–33). Jessica Kingsley Publishers