Traditional explanatory resources, such as user manuals and textbooks, often contain content that may not cater to the diverse backgrounds and information needs of users. Yet, developing intuitive, user-centered methods to efectively explain complex or large amounts of information is still an open research challenge. In this paper we present ExplanatoryGPT, an approach we devised and implemented to transform textual documents into interactive, intelligent resources, capable of ofering dynamic, personalized explanations. Our approach uses state-of-the-art question-answering technology to generate on-demand, expandable explanations, with the aim of allowing readers to eficiently navigate and comprehend static materials. ExplanatoryGPT integrates ChatGPT, a state-of-the-art language model, with Achinstein's philosophical theory of explanations. By combining question generation and answer retrieval algorithms with ChatGPT, our method generates interactive, user-centered explanations, while mitigating common issues associated with ChatGPT, such as hallucinations and memory shortcomings. To showcase the efectiveness of our Explanatory AI, we conducted tests using a variety of sources, including a legal textbook and documentation of some health and financial software. Specifically, we provide several examples that illustrate how ExplanatoryGPT excels over ChatGPT in generating more precise explanations, accomplished through thoughtful macro-planning of explanation content. Notably, our approach also avoids the need to provide the entire context of the explanation as a prompt to ChatGPT, a process that is often not feasible due to common memory constraints.
The increasing demand for eficient explanatory resources, such as user manuals and textbooks,
calls for innovative approaches that can cater to the diverse backgrounds and information
needs of users. Traditional materials often contain static, predefined content, which may not
be suficient for users to efectively comprehend complex information [
Recent advancements in natural language processing, particularly in large-scale language
models like ChatGPT, have shown potential for enhancing explanatory capabilities [
To navigate these challenges, this paper proposes ExplanatoryGPT, a novel approach that harnesses the strengths of ChatGPT. This approach is dedicated to resolving the stated challenges and improving the quality of explanations in intelligent textbooks and documentation. For the purposes of this study, we interpret hallucination broadly as the generation of content that does not maintain fidelity to a given context or source.
ExplanatoryGPT combines ChatGPT with Achinstein’s philosophical theory of explanations. By integrating question generation and answer retrieval algorithms with ChatGPT, our method generates interactive, user-centered explanations, while mitigating common issues associated with ChatGPT, such as hallucinations. Our approach contributes to the field of intelligent textbooks and documentation by enhancing the explanatory power of textual information utilizing question-answering technology to generate on-demand, expandable explanations.
In particular, our method involves the following steps: First, the question-answering algorithms extract pertinent textual information found in textbooks or documentation. Then, this information is reorganized based on the user’s specific query, ensuring that the explanation is tailored to the user’s needs and context. Subsequently, we employ ChatGPT to refine the retrieved information, enhancing its readability, coherence, and cohesion. By leveraging ChatGPT’s text generation capabilities, we produce explanations that are not only more understandable but also more engaging and relevant to the user.
To demonstrate the efectiveness of our approach, we have applied it to a law textbook [
Our work contributes to the academic discourse on the importance of integrating advanced language models, such as ChatGPT, and philosophical theories like Achinstein’s framework, for the development of intelligent textbooks and documentation. This interdisciplinary approach can lead to the creation of more efective and versatile explanatory methods, which can be adapted to diferent domains and use cases. The benefits of our approach extend beyond the improvement of explanations themselves. By generating more accessible and comprehensible explanations, we can heighten users’ engagement and understanding of various subjects and technologies. This, in turn, can lead to increased adoption and more efective use of educational materials.
Our approach builds upon existing research on intelligent textbooks and natural language processing, particularly regarding advanced language models like GPT. In this section, we review related work and establish connections to our proposed approach.
Studies suggest that using intelligent textbooks1 and interactive e-books can increase usage,
motivation, and learning gains compared to static e-books [
The use of Artificial Intelligence ( AI) for the automatic generation of interactive e-books seems to be under-explored. In one such project, [16] propose to automatically augment the sections of existing books with related YouTube videos by directly annotating the PDF, thus without breaking the structure of these textbooks. Unlike previous research, our approach fully automates the conversion of existing e-books into interactive versions by integrating theories of explanations, intelligent interfaces, and Explanatory Artificial Intelligence ( YAI). We explore how questions can practically organize and categorize explanation content.
By integrating ChatGPT’s strengths with Achinstein’s philosophical theory of explanations and question-answering algorithms, our proposed approach is a novel solution for enhancing intelligent textbooks and documentation. ExplanatoryGPT is aimed to generating more interactive, user-centered explanations while addressing challenges associated with advanced language models, fostering more efective and personalized learning experiences.
Our work aligns with current educational trends and the growing interest in the use of AI in
education, as evidenced by the emerging literature. For example, UNESCO’s Quick Start Guide2
provides an overview of ChatGPT in higher education and emphasizes challenges and ethical
implications. Similarly, Joyner [
To address the issues of hallucination and memory constraints commonly encountered in
large language models like GPT [
In this section, we provide an overview of the theoretical foundations that underpin our proposed approach for enhancing the explanatory power of AI systems. We discuss YAI and the role of ChatGPT as an advanced language model.
The concept of explanation in philosophy began to have a more precise role in the 20th century with the growth and development of the philosophy of science. Hempel’s deductive-nomological model [20] gave rise to the first theory of explanations, followed by many competing theories, such as Salmon’s Causal Realism [21], and Achinstein’s Ordinary Language Philosophy [22].
Achinstein’s theory emphasizes the communicative aspect of explanation, its usefulness in answering questions, and fostering understanding between individuals. Holland’s theory [23] frames the process of explaining as a purely cognitive activity, while Sellars [24] suggests a utilitarian process of constructing a coherent belief system.
According to Achinstein, explaining is an illocutionary act, born of a clear intention to produce new understandings in an explainee by providing a correct content-giving answer to an open-ended question. Illocution in explaining involves informed and pertinent answers to the main question and other implicitly relevant questions [25, 19].
Definition 1 (Illocution in Explaining). Explaining is an illocutionary act that provides answers to an explicit question on some topic along with answers to several other implicit or unformulated questions deemed necessary for the explainee to understand the topic properly. In the most generic case, no assumption can be made about the explainee’s knowledge and objectives, and the only implicit questions that can be exploited for illocution are the most generic ones, called archetypal questions.
For example, an answer like ‘I am happy because I just got a paper accepted at this important venue, and [...]’ would generally be considered an explanation because it answers other archetypal questions.
An YAI is an artificial intelligence program designed to generate user-centered, interactive explanations out of (possibly extensive) collections of explainable information [26]. An example of YAI based on Achinstein’s theory of explanation is YAI4Hu [25, 19].
YAI4Hu is a fully automatic explanatory tool used to explain (pre-existing) documentation about an AI-based system. In particular, the textual content of such documentation is algorithmically reorganized and represented as a special hypergraph where information can be either explored through overviewing or searched via open-ended questioning. On the one hand, open-ended questioning can be performed by asking open-ended questions in English through a search box that uses the knowledge graph for eficient answer retrieval. An example of open-ended questioning is shown in Figure 2.
On the other hand, overviewing can be performed iteratively from an initial explanation by clicking on automatically annotated words for which explanations exist. In particular, annotated words are visible because they have a unique format that makes them easy to recognize. After clicking on an annotation, a modal opens (see Figure 1), showing a navigation bar with tabs containing explanatory overviews of the clicked annotated words. The information shown in the overview includes: • A short description of the explained word (if available). • The list of other taxonomically connected words. • A list of predefined archetypal questions (e.g., why is this aspect/concept important, what is this aspect/concept, etc.) and their respective answers ordered by estimated pertinence.
ChatGPT is a cutting-edge language model based on the GPT architecture [27]. GPT has been used in various applications, including question-answering systems [28], summarization [29, 30], and chatbots [31, 32], showcasing its versatility and potential for generating contextually relevant and coherent text. However, ChatGPT and similar models have been known to sufer from hallucinations and memory constraints, which can lead to the generation of text that is plausible-sounding but factually incorrect or ungrounded [33]. Our proposed methodology addresses these issues by integrating ChatGPT with question-answering algorithms, grounding the generated explanations in relevant and accurate information, and refining the readability, coherence, and cohesion of the retrieved information.
In this section, we introduce our proposed approach, which combines answer retrieval algorithms with ChatGPT to produce high-quality explanations. Our main objective is to leverage ChatGPT’s ability to provide clear explanations while addressing common issues such as hallucinations and lack of control over its outputs, which can utilize any learned information regardless of its relevance.
Our strategy builds upon the open-questioning and overviewing mechanisms from Section 3.2. These mechanisms use a concept called illocution, where key archetypal questions are answered to produce valid explanations. Our goal is to enhance these mechanisms by utilizing ChatGPT to merge the answers, thus taking advantage of the strengths of both techniques.
Our approach combines information retrieval algorithms with ChatGPT in a complementary manner to generate high-quality explanations. On one hand, the information retrievers extract and reorder critical information from textbooks or software documentation, designed to address a wide range of user questions, explicit or implicit. By aligning explanations with the user’s specific query, we ensure relevance and contextual appropriateness.
On the other hand, ChatGPT is employed to refine the retrieved information. ChatGPT possesses good text generation capabilities, producing human-like content (see Section 3.3). Moreover, it enables the generation of high-quality and customizable text [34]. For these reasons, after the algorithms extract and reorganize relevant information, we use ChatGPT to refine the information retrieved, aiming to enhance readability, coherence, and cohesion of the explanations, making them more accessible to a wider audience.
In the refinement process, ChatGPT rephrases, restructures, and condenses the information gathered by the answer retrieval algorithm, ensuring clear, concise, and easily understandable explanations. This approach also addresses potential issues of hallucinations or memory limitations in ChatGPT by grounding it in the information retrieved by the algorithms. To guide ChatGPT in the refinement process, the following instruction (or any similar instruction) can be given:
Using exclusively the information provided, create a coherent and clear explanation of ’{question}’ for a student with no prior knowledge on the subject. Provide both jargon and simpler synonyms. Your exposition should follow this structure: Short Answer, Technical Details, Conclusion.
Information provided: • {retrieved_answer_1} • {retrieved_answer_2} • ...
In this instruction: • ‘{question}’ is a placeholder for the actual question that the user (implicitly or explicitly) asks, which will be replaced with the specific question that requires an explanation. • ‘{retrieved_answer_1}’, ‘{retrieved_answer_2}’, and so on, are placeholders for the answers that the question-answering algorithms extract. These answers come from reliable sources or databases and provide the raw material for ChatGPT to structure the explanation. The instruction guides ChatGPT to create an explanation using the information provided, i.e., the retrieved answers. The explanation should be clear and understandable to a student with no prior knowledge of the subject. It asks ChatGPT to use both technical terms (jargon) and their simpler synonyms to make the information accessible to a wide range of users.
The requested structure (i.e., Short Answer, Technical Details, Conclusion) helps ChatGPT structure the information logically. The ‘Short Answer’ provides a brief response to the question, the ‘Technical Details’ delve deeper into the explanation, and the ‘Conclusion’ summarizes the main points. By following these guidelines, ChatGPT is expected to generate explanations that are not only technically accurate but also clear, well-structured, and suitable for users with diferent levels of understanding. The focus on using exclusively the information provided helps keep ChatGPT on track and reduces potential hallucinations.
In the next section, we will delve into the details of how this approach can be employed, its limitations, and the outcomes it can yield.
We have developed three proof-of-concept demonstrations of the ExplanatoryGPT. These examples showcase the potential and versatility of the ExplanatoryGPT system. Two of these demonstrations focus on the software documentation of AI-based systems, specifically those used for Credit Approval and predicting heart diseases. We selected these examples because they delve into intricate, technical details and concepts that could easily confuse an AI model. More information about these AI-based systems can be found in previous work [19].
Our third demonstration involves the ExplanatoryGPT’s application to a textbook designed to teach the writing of legal memorandums in the US legal system. This example showcases the model’s potential to be used in educational settings, specifically for teaching complex topics like law. Details about this textbook and educational setting can be found in [35].
The reason we chose these diverse examples is because they involve information and technical concepts that may not be easily understood by ChatGPT. Software documentation, for instance, often contains technical information that needs to be interpreted in context, while textbooks may contain specific information that may need to be carefully controlled or hidden during the learning process.
An overview of the ExplanatoryGPT mechanism in the Credit Approval system is presented in Figure 1. As seen in this figure, ExplanatoryGPT synthesizes both retrieved and generated information to deliver an explanation. This mechanism is expected to be especially efective in technical areas such as credit approval, where information needs to be clearly communicated to avoid any misunderstanding. Similarly, Figure 2 illustrates the open-questioning mechanism of ExplanatoryGPT in the educational scenario of teaching legal memorandum writing.
Table 1 presents a side-by-side comparison of the explanatory outputs produced by ChatGPT and ExplanatoryGPT. This comparison highlights the diferences in the quality and precision of explanations provided by the two models. Specifically, the explanations generated by ChatGPT are obtained through the following instruction which (unlike ExplanatoryGPT) does not include any retrieved answer but only the given question: “Create a coherent and clear explanation of ’question’ for a student with no prior knowledge on the subject. Provide both jargon and simpler synonyms. Your exposition should follow this structure: Short Answer, Technical Details, Conclusion.”
As shown in Table 1, our ExplanatoryGPT mechanism has the potential to efectively control what content should be part of the explanations. This control is crucial for several reasons: 1. It mitigates the ‘hallucination’ problem of ChatGPT, where the model sometimes generates information that is irrelevant or incorrect. This is evident in the ‘CEM’ example shown in Table 1. 2. It improves contextualisation of explanations which may otherwise be generic and not specifically focused. For example, the ‘Satisfactory Trades’ explanation in Table 1 is more accurate and targeted due to the control provided by ExplanatoryGPT. 3. It provides a mechanism to control which information should be shown or hidden. This feature is particularly useful in educational settings, where certain information might be withheld as part of a learning exercise.
Instead of employing ChatGPT as a stand-alone generator of end-to-end explanations, we incorporate an YAI mechanism for the macro-planning phase, which involves the selection and ordering of information [36]. Subsequently, ChatGPT is utilized for micro-planning and for surface realization, i.e., the coherent amalgamation of information into sentences.
The main task of ExplanatoryGPT is to avoid overly broad, incorrect, or out-of-context explanations. This becomes especially clear when considering the examples in Table 1, such as the first two questions. The answers provided by ExplanatoryGPT are much more specific and indicate a greater ‘awareness’ of the explanatory context.
Notably, in the absence of an YAI mechanism, one might still formulate a prompt incorporating the full context necessary for generating an appropriate response. However, in the broadest of scenarios, executing this prompt without a mechanism (i.e., an YAI) for filtering redundant information would not be feasible, efective, or eficient. This is primarily due to the stringent memory constraints associated with ChatGPT. Therefore, the use of a YAI allows for more accurate and context-specific outputs.
In this paper, we presented ExplanatoryGPT, an approach that combines ChatGPT with Achinstein’s philosophical theory of explanations for intelligent textbooks and documentation. Our proposed YAI methodology has the potential to enhance the quality and usability of explanations, making them more interactive and user-centred.
We showcased the potential of ExplanatoryGPT to generate better and less hallucinated explanations in a law textbook and software documentation related to healthcare and finance. Our research highlights the importance of integrating advanced language models and philosophical theories for the future development of intelligent textbooks and documentation.
For future work, we plan to conduct extensive user studies to empirically evaluate our approach, as our current evaluation was limited in scope and rigor. Other research directions include expanding the application of our methodology to other domains and industries.
In conclusion, our proposed YAI methodology ofers a promising avenue for improving learning and understanding experiences for a diverse audience, ultimately contributing to more inclusive and eficient educational materials and software documentation.
Acknowledgements. We thank the copyright holders of [
Author Contributions. F. Sovrano: conceptualization, methodology, software, data curation, original draft preparation, visualization, investigation, validation, formal analysis. K. Ashley: conceptualization and supervision. A. Bacchelli: review, editing, and supervision. [9] J. P. Beier, M. A. Rau, Embodied learning with physical and virtual manipulatives in an intelligent tutor for chemistry, in: Artificial Intelligence in Education - 23rd International Conference, AIED 2022, Durham, UK, July 27-31, 2022, Proceedings, Part I, volume 13355 of Lecture Notes in Computer Science, Springer, 2022, pp. 103–114. [10] I. A. Chacon, J. Barria-Pineda, K. Akhuseyinoglu, S. A. Sosnovsky, P. Brusilovsky, Integrating textbooks with smart interactive content for learning programming, in: Proceedings of the Third International Workshop on Intelligent Textbooks 2021, volume 2895 of CEUR Workshop Proceedings, CEUR-WS.org, 2021, pp. 4–18. [11] M. Mohammed, C. A. Shafer, Increasing student interaction with an etextbook using programmed instruction (short paper), in: Proceedings of the Third International Workshop on Intelligent Textbooks 2021, volume 2895 of CEUR Workshop Proceedings, CEUR-WS.org, 2021, pp. 40–44. [12] N. Matsuda, M. Shimmei, PASTEL: evidence-based learning engineering method to create intelligent online textbook at scale, in: Proceedings of the First Workshop on Intelligent Textbooks, volume 2384 of CEUR Workshop Proceedings, CEUR-WS.org, 2019, pp. 70–80. [13] Z. Wang, J. Valdez, D. B. Mallick, R. G. Baraniuk, Towards human-like educational question generation with large language models, in: Artificial Intelligence in Education - 23rd International Conference, AIED 2022, Durham, UK, July 27-31, 2022, Proceedings, Part I, volume 13355 of Lecture Notes in Computer Science, Springer, 2022, pp. 153–166. [14] M. Shimmei, N. Matsuda, Automatic question generation for evidence-based online courseware engineering, in: Proceedings of the Fourth International Workshop on Intelligent Textbooks 2022, volume 3192 of CEUR Workshop Proceedings, CEUR-WS.org, 2022, pp. 18–25. [15] B. Kluga, M. S. Jasti, V. Naples, R. Freedman, Adding intelligence to a textbook for human anatomy with a causal concept map based ITS, in: Proceedings of the First Workshop on Intelligent Textbooks, volume 2384 of CEUR Workshop Proceedings, CEUR-WS.org, 2019, pp. 124–134. [16] J. Barria-Pineda, A. B. L. Narayanan, P. Brusilovsky, Augmenting digital textbooks with reusable smart learning content: Solutions and challenges, in: Proceedings of the Fourth International Workshop on Intelligent Textbooks 2022, volume 3192 of CEUR Workshop Proceedings, CEUR-WS.org, 2022, pp. 77–91. [17] Y. Bang, S. Cahyawijaya, N. Lee, W. Dai, D. Su, B. Wilie, H. Lovenia, Z. Ji, T. Yu, W. Chung, et al., A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity, arXiv preprint arXiv:2302.04023 (2023). [18] Y. Mass, H. Roitman, Ad-hoc document retrieval using weak-supervision with BERT and GPT2, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, Association for Computational Linguistics, 2020, pp. 4191–4197. [19] F. Sovrano, F. Vitali, Generating user-centred explanations via illocutionary question answering: From philosophy to interfaces, ACM Trans. Interact. Intell. Syst. 12 (2022). [20] C. G. Hempel, P. Oppenheim, Studies in the logic of explanation, Philosophy of Science 15 (1948) 135–175. doi:10.1086/286983. [21] W. Salmon, Scientific Explanation and the Causal Structure of the World, Book collections on Project MUSE, Princeton University Press, 1984. URL: https://books.google.it/books? id=2ug9DwAAQBAJ. [22] P. Achinstein, The Nature of Explanation, Oxford University Press, 1983. URL: https: //books.google.it/books?id=0XI8DwAAQBAJ. [23] J. Holland, K. Holyoak, R. Nisbett, P. Thagard, Induction: Processes of Inference, Learning, and Discovery, Bradford books, MIT Press, 1986. URL: https://books.google.it/books?id= Z6EFBaLApE8C. [24] W. Sellars, Science, Perception and Reality, New York: Humanities Press, 1963. [25] F. Sovrano, F. Vitali, From philosophy to interfaces: an explanatory method and a tool inspired by achinstein’s theory of explanation, in: IUI ’21: 26th International Conference on Intelligent User Interfaces, College Station, TX, USA, April 13-17, 2021, ACM, 2021, pp. 81–91. [26] F. Sovrano, F. Vitali, Explanatory artificial intelligence (yai): human-centered explanations of explainable ai and complex data, Data Mining and Knowledge Discovery (2022). [27] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al., Language models are unsupervised multitask learners, 2019. [28] B. Sievers, Question answering for comparative questions with GPT-2, in: Working Notes of CLEF 2020 - Conference and Labs of the Evaluation Forum, Thessaloniki, Greece, September 22-25, 2020, volume 2696 of CEUR Workshop Proceedings, CEUR-WS.org, 2020. [29] Z. Yang, Y. Dong, J. Deng, B. Sha, T. Xu, Research on automatic news text summarization technology based on GPT2 model, in: AIAM 2021: 3rd International Conference on Artificial Intelligence and Advanced Manufacture, Manchester, United Kingdom, October 23 - 25, 2021, ACM, 2021, pp. 418–423. [30] B. Chintagunta, N. Katariya, X. Amatriain, A. Kannan, Medically aware GPT-3 as a data generator for medical dialogue summarization, in: K. Jung, S. Yeung, M. P. Sendak, M. W. Sjoding, R. Ranganath (Eds.), Proceedings of the Machine Learning for Healthcare Conference, MLHC 2021, 6-7 August 2021, Virtual Event, volume 149 of Proceedings of Machine Learning Research, PMLR, 2021, pp. 354–372. [31] T. Teubner, C. M. Flath, C. Weinhardt, W. M. P. van der Aalst, O. Hinz, Welcome to the era of chatgpt et al, Bus. Inf. Syst. Eng. 65 (2023) 95–101. [32] M. Cascella, J. Montomoli, V. Bellini, E. Bignami, Evaluating the feasibility of chatgpt in healthcare: An analysis of multiple clinical and research scenarios, J. Medical Syst. 47 (2023) 33. [33] H. Alkaissi, S. I. McFarlane, Artificial hallucinations in chatgpt: implications in scientific writing, Cureus 15 (2023). [34] P. P. Ray, Chatgpt: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope, Internet of Things and Cyber-Physical Systems (2023). [35] F. Sovrano, K. Ashley, P. Brusilovsky, F. Vitali, Yai4edu: an explanatory AI to generate interactive e-books for education, in: Proceedings of the Fourth International Workshop on Intelligent Textbooks 2022, volume 3192 of CEUR Workshop Proceedings, CEUR-WS.org, 2022, pp. 31–39. [36] R. Dale, H. Moisl, H. Somers, Handbook of Natural Language Processing, Taylor & Francis, 2000. URL: https://books.google.it/books?id=VoOLvxyX0BUC.