Textbkoos remain one of the main methods of instruction, but—just like other educational tools—they have been evolving over the last several decades in many aspects (how they are created, published, formatted, accessed, and maintained). Most textbooks today have digital versions and can be accessed online. A large number of textbooks (and other structured instructional texts, such as tutorials) are freely available as open educational resources (e.g., Baraniuk et al., 2017). Many textbooks come with libraries of supplementary educational resources (Ericson & Miller, 2020; Gordon et al., 2021) or even distributed as parts of online educational services built on top of them (Ritter et al.,  2019). The transition of textbooks from printed copies to digital and online formats has facilitated numerous attempts to enrich them with various kinds of interactive functionalities including search and annotation, interactive content modules, automated assessments, question ans, chatbots, etc.

        As a result of these enrichments, new research challenges and opportunities emerge that call for the application of artificial intelligence (AI) methods to enhance digital textbooks and learners’ interaction with them. There are many research questions associated with this new area of research; examples include:

• How can we facilitate access to textbooks and improve the reading process?
• How can we process textbook content to infer knowledge underlying the text and use it to improve learning support?
• How can we process increasingly more detailed logs of students interacting with digital textbooks and extract insights on learning?
• How can we find and retrieve relevant content “in the wild”, i.e., on the web, that can enrich the textbooks?
• How can we leverage advanced language technology, such as chatbots, to make textbooks more interactive?
• How can we better understand both textbooks and student behaviors as they learn within the textbook and create personalized learner experiences?

Our previous analyses of the field of Intelligent Textbooks have identified several topical trends (Rodrigo et al., 2023) and historic milestones (Brusilovsky et al., 2022) that help us understand how the field has developed over the years and which challenges have remained the focus of the community. In this section, we present an updated overview of the evolution of the research on Intelligent Textbooks, including more recent developments in the field. We hope that this overview helps the readers understand the “big picture” of research in this fast-developing field.

Generation 1: Intelligence Engineered

The first generation of research on intelligent textbooks was born at the crossroads of intelligent tutoring systems (Reggia et al., 1980) and adaptive educational hypermedia (Brusilovsky et al., 1993). Interfaces of first prototypes did not utilize the familiar textbook layout; instead, they provided students with adaptive access to information items or hyper-cards as a response to students’ performance with accompanying exercises. After the Web (with its hypertext core) became the main platform for the deployment of educational tools, first-generation intelligent textbooks adopted the common structured textbook organization with its tree of sections, a table of contents, and glossaries. On the top of this common structure, these intelligent textbooks implemented a range of personalization techniques to support the readers, such as navigation support (Henze et al., 1999; Weber & Brusilovsky, 2016), adaptive page manipulation (Hohl et al., 1996), content recommendation (Kavcic, 2004), content sequencing (Ullrich, 2008), or a combination of different techniques (Sosnovsky, 2014). These textbooks were developed as closed-corpus environments that provided access to limited collections of carefully crafted of instructional texts and learning exercises (see Fig. 1). The intelligence in first-generation intelligent textbooks was based on traditional knowledge engineering approaches, which required engagement of human domain experts. As illustrated in Fig. 1, these experts helped structure the learning domain as a set of concepts (nodes), connect textbook sections and activities (documents) with domain concepts (edges), and develop personalization approaches based on these domain and content models (gears). Students’ interaction with the elements of the textbook content (texts and exercises) was traced to model the growth of student knowledge about domain concepts and use by its personalization technologies to help students master the textbook knowledge in an individually optimized manner.

Generation 2: Intelligence Integrated

For the next generation of intelligent textbooks, the focus initially shifted to more practical aspects of implementing textbook environments, such as open architectures (De Bra & Calvi, 1998; Brusilovsky, 2004), integration of textbooks with external educational resources (Dolog et al., 2004a), and standardization of semantic models that support such architectures and integration (Dolog et al., 2004b; Melis et al., 2006). Illustrated in Fig. 2, this integration enabled researchers to address one of the recognized problems of first-generation intelligent textbooks: their closed-box nature (once designed, they were impossible to extend, modify, and scale). The development of Semantic Web technologies offered new standardization formats and modeling languages for shareable and expressive knowledge representation along with new architectural solutions for intelligent software. This development enabled researchers to implement knowledge models for intelligent textbooks as ontologies, develop educational material as interlinked and annotated learning objects, and seamlessly integrate shareable knowledge models and content within textbooks. Although adaptive and intelligent access to textbook content remained the ultimate goal of second-generation intelligent textbooks, the expressive power of employed semantic models supported some new functionalities. The most advanced second-generation textbooks were able to “understand” their content and engage in new forms of interactions with learners such as meaningful question answering and concept mapping (Chaudhri et al., 2013).

Generation 3: Intelligence Extracted

The next set of intelligent textbook projects were characterized by a fundamental shift in the perspective on the role of textbooks and the AI toolkit used to “make them intelligent”. While the first and second generations were driven by classical AI approaches for symbolic knowledge representation and logic-based knowledge inference, the third generation employed Machine Learning (ML) and Natural Language Processing (NLP) methods focusing on knowledge extraction. Researchers realized that textbooks are not only educational tools, but also sources of high-quality educational content that is written and structured by domain experts (textbook authors) with the purpose of explaining the subject of the textbook to students. Illustrated in Fig. 3, third-generation intelligent textbook researchers treated digital textbooks less as objects to add “intelligence” to, and more as sources of “intelligence” that can be harvested automatically and at scale. A wide range of newly-developed NLP, semantic web, and ML methods were applied to the content of digital textbook with the purpose of extracting topics (Sosnovsky et al., 2012), concepts (Wang et al., 2015), relations (Adorni et al., 2019; Alzetta et al., 2020), and even complete knowledge models (Alpizar-Chacon & Sosnovsky, 2021). Turning the focus of knowledge extraction research back to education, novel NLP approaches used textbooks as knowledge sources to generate individual learning objects (Yarbro & Olney, 2021), explanations (Sovrano et al., 2022), and assessment material (Dresscher et al., 2021; Van Campenhout et al., 2021). Together, the new ability to automatically extract knowledge and learning objects from digital textbooks means that the textbooks themselves could be automatically transformed into intelligent textbooks (Alpizar-Chacon et al., 2020), linked to external interactive content (Rahdari et al., 2020) and connected to each other (Guerra et al., 2013; Alpizar-Chacon & Sosnovsky, 2019).

Generation 4: Intelligence Datamined

As the adoption of digital textbooks in real educational context expanded, another source of “intelligence” became increasingly available: usage data generated by students interacting with textbooks. These data included latent traces of students’ progress with textbooks through reading, question-answering, explicitly provided annotations (Farzan & Brusilovsky, 2008), highlights (Winchell et al., 2018), and tags (Zhang et al., 2017). Illustrated in Fig. 4, the fourth-generation research on intelligent textbooks applied a range of data-mining methods to leverage student-textbook interaction logs in several ways. Most importantly, the large volumes of student data allowed researchers to better understand reading behavior and connect it to learning progress. On a coarse-grained level, extracted patterns of navigation through textbook sections and pages along with annotation activity were used to predict readers’ overall success or failure (Kim et al., 2021), which then could be used for a timely intervention (Winchell et al., 2018; Yin et al., 2019). On a more fine-grained level, student interactions with textbook sections were used to model their knowledge of domain concepts (Thaker et al., 2019). The resulting models could be used to recommend the most relevant internal (Thaker et al., 2020) or external (Barria-Pineda et al., 2019) content. User interaction data could also be used to improve textbook organization or representation. For example, log data combined with the analysis of textbook content could be used to obtain more reliable prerequisite relationships (Liu et al., 2016). Finally, interaction data could be revealed directly to the readers through the textbook interface. This approach could be used to support open learner modeling, social awareness, and self-regulation (Papanikolaou et al., 2003; Barria-Pineda et al., 2019).

Generation 5: Intelligence Generated

Illustrated in Fig. 5, many recent works on intelligent textbooks have started to incorporate large language models (LLMs), which are perfectly suited to extract and synthesize information from textbooks. Evidence suggests that most existing LLMs are pre-trained on a large amount of textual data collected from books, textbooks, and even synthetic textbook-like data (CNN, 2023; Gunasekar et al., 2023). As a result, LLMs already capture abundant knowledge contained in textbooks during pre-training and can be used to help learners. For example, the work in Levonian et al. (2023) examines using OpenStax textbooks to power automated question answering tools for learners, showing that retrieval augmented generation techniques are effective at reducing the intrinsic hallucination problem in LLMs. One challenge in this direction is that the context length of LLMs, i.e., the amount of information they can effectively take in during deployment, is limited and not enough to cover long documents (Fernandez et al., 2024). Another important direction is to use LLMs to power more interactive experiences around textbooks, e.g., chatbots: Khan academy leverages their access to GPT-4 by OpenAI to power a tutoring chatbot around their content through prompt engineering (Khan Academy, 2023), while others use GPT-4, a proprietary LLM to simulate tutor-learner conversations and use this data to train smaller, open-source LLMs as chatbots (Sonkar et al., 2023).

Powered by advancing AI technologies, new learner support tools around intelligent textbooks are being developed at a rapid pace. Through the review of existing work, we find that new technology, especially that in the third, fourth, and fifth generations of intelligent textbooks as we categorized above, is enabled by extracting large-scale data from textbooks. For example, chatbots need to be carefully calibrated based on real teacher-learner interaction around learning content (Macina et al., 2023; Wang et al., 2024). These developments reveal a mutually beneficial cycle between the development of new tools and data extraction and analysis: new tools open up new ways of learner-textbook interaction, which generates new types of data that can, in turn, power the development of new tools. This cycle had already existed before in the field of intelligent educational systems. Indeed, the development of richer learning interfaces led to the increased volume of learner data being collected to power user models, which then enabled new personalization approaches. However, with new AI technologies, this cycle accelerated and became more evident. With this vision in mind, it is important for future work on intelligent textbooks to develop new modes of interaction between learners and textbooks, e.g., through drawing, sketching, highlighting (Kim et al., 2020), annotations, and conversations, powered by advanced human-computer interaction, smart interfaces, and generative AI (OpenAI, 2024), which will open up new fronts in data analysis to help us better understand learners.

This special issue was motivated by the success of the International Workshop on Intelligent Textbooks, which served as the main forum for reporting progress in this research field over the last five years. Altogether, 54 contributions have been presented at the workshop and published in its proceedings from 2019 to 2023 (Sosnovsky et al., 2019, 2020, 2021, 2022, 2023). Our analysis of these contributions identified eleven main research topics:

• Intelligent and Adaptive Interfaces
• Smart Content Integration
• Content Analysis and Knowledge Extraction
• Learning Content Extraction and Generation
• Intelligent Textbook Authoring and Generation
• Interaction Mining and Behavior Modeling
• Social Interaction and Crowdsourcing
• GenAI-based Textbooks
• Intelligent Textbook Pedagogy
• Domain-Specific Textbook Prototypes
• Bridging Physical and Digital Textbooks

        Figure 6 attempts to visualize chronologically how these topics were represented in the papers published in the workshop proceedings in different years and how the community interests evolved. The numbers in the table indicate how many papers that year covered this particular topic. The area of a topic zone in the diagram visualizes the corresponding ratio of this topic. Since many papers covered more than one topic, these areas do not reflect the direct percentage of papers covering the topic but rather the relative focus of the topic compared to all other topics covered that year. For example, in 2019, 10 of 14 papers described different forms of adaptive and intelligent interaction between the interfaces of intelligent textbooks and their users. However, the same papers covered many other topics as well; hence, the area visualizing this topic’s contribution that year corresponds to 27% (instead of 71%). On a relevant note, we have noticed that with time, papers have become more focused on a single technology rather than on a combination of those (while in 2019, an average paper could be attributed to 2.4 topics, these numbers gradually decreased: to 1.7 in 2023). This might reflect the growing maturity of the research presented at the workshop, as the authors gradually started to spend more effort exploring individual technologies rather than presenting projects and prototypes that attempted to do many things at once. It is important to note that the identified topics might differ in granularity, which would influence their representation in the papers. A more detailed analysis could show additional trends; however, even on this level we can see the evolution of the intelligent textbook research over the 5 years.

To make a connection to the previous section, Figure 6 associates (combinations of) topics with the key goals explored by the five generations of intelligent textbooks. It is clear that adaptive support of interaction with textbook content (Generation 1) remains an important aspect of what is considered to be intelligence in intelligent textbooks. At the same time, semantic integration of external interactive resources with textbooks (Generation 2) has gradually given way to approaches focusing on the extraction of knowledge and interactivity from textbooks using methods of ML, symbolic NLP, and content analysis (Generation 3). In turn, very recently, this line of research itself gave way to a new wave of papers driven by the recent success of Large Language Models (Generation 5). We think that this trend will continue and expect to observe both development of new prototypes fully powered by generative AI and enrichment of already developed successful tools with new generative AI technologies. Finally, we see that the stream of research focusing on data-driven and social aspects of intelligent textbooks (Generation 4) has remained stable over the years. The remaining three topics could be classified as miscellaneous. They do not constitute a particular trend, yet provide unique perspectives and ideas on how intelligent textbooks can be enhanced, organized, or orchestrated. The remainder of this section discusses the identified topics in more detail.

It was our experience with the Intelligent Textbook workshop that motivated the special issue of the International Journal of Artificial Intelligence in Education on Intelligent Textbooks. The five years of workshops demonstrated an increased interest in the topic as well as a growing maturity of research in this area. The special issue was proposed to the journal, accepted, and announced to the community. The editorial process for this special issue has been organized in two phases. First, interested authors submitted extended abstracts that describe the main contributions of the intended papers and their relevance to the theme of the special issue. Nineteen of such abstracts were received and reviewed by guest editors; authors of nine of them were invited to submit full versions of their papers. Each of the submitted papers was reviewed by at least three reviewers and a guest editor. The final set of papers accepted for publication after several rounds of review includes four submissions.

        The papers that form this special issue provide new contributions to the key topics presented in Fig. 6 and follow the trajectory of the development of the state of the art in the area of Intelligent Textbooks. The papers represent a healthy combination of methods and approaches from the third, fourth, and fifth generations of intelligent textbooks.

        Sovrano et al. (2024) focuses on providing on-demand explanations to learners using intelligent textbooks, doing so by extracting a knowledge graph from a collection of textbooks and other learning resources. The authors conducted an experiment in the context of legal writing education to show that the enriched textbook leads to better learning outcomes than the baselines. This approach fits into the third and fifth generation of intelligent textbooks, using both more traditional technology such as knowledge graphs and more recent technology such as generative question-answering models. These technologies cover the topics of knowledge extraction, textbook authoring and generation, and genAI-based textbooks, while the end product contributes to the topic of domain-specific textbook prototypes.

        Jonnson et al. (2024) follows up on a series of papers in the International Workshops on Intelligent Textbooks on question generation, using approaches that were commonly used before the emergence of LLMs. This work enriches textbooks with generated questions, with an added twist, parallel construction: Questions are generated in English and translated into other languages for existing parallel corpora (textbooks and their translated versions). This approach fits into the fifth generation of intelligent textbooks, also covering the topics of learning content generation and intelligent textbook authoring and generation.

        Morris et al. (2024) uses LLMs to generate feedback on student-written summaries of textbook content. This paper isolates two aspects from the scores given to student-written summaries: content and writing, and develops an LLM-based classifier to score these aspects and provide feedback to students. It combines the technologies of the fourth and fifth generations of intelligent textbooks through the analysis of student-written summary scoring data and the use of LLMs for a classification task. These approaches cover the topics of learning content extraction and generation, interaction mining and behavior modeling, and genAI-based textbooks.

        Blobstein et al. (2024) continues the fourth-generation trend of extracting useful knowledge from a large volume of interaction data generated by textbook readers. The interesting source of learner data leveraged by this work are students’ annotation in the form “emojies” or hashtags that students use to represent their emotions and other types of reactions to the text they read. The paper demonstrated that aggregating student annotation data could help teachers understand which parts of the text have been challenging or interesting. It also explored the feasibility of using LLMs to predict student annotations.

Despite its long history, Intelligent Textbooks remain a field of innovations where new approaches are introduced and explored in various studies (Kukulska-Hulme et al., 2024). Intelligent textbook researchers eagerly adopt technologies developed outside of the field and come up with use-cases and application scenarios that expand the capabilities and potential effectiveness of textbook prototypes. Textbooks continue to be one of the primary sources of knowledge for students, despite more interactive, non-textual content gaining popularity in many other tasks and contexts.

        It is time for textbooks to enhance their functionality with interactive features that improve the overall effectiveness of textbooks as learning tools (Clinton-Lisell et al., 2023). Adaptive textbook interfaces, smart content, intelligent services, and community-focused tools to engage with other users will transform learners’ study with textbooks into a more interactive and social learning experience. The new wave of generative AI that already fuels the current generation of intelligent textbooks opens up a wide range of possibilities to make learning with textbooks even more interactive and engaging. LLMs are highly capable of effectively transferring what they learn from pre-training on web-scale text to different contexts, resulting in high levels of fluency and consistency of the generated text. Currently, the most popular direction for these works is to use LLMs to automatically produce assessment questions for intelligent textbooks, with promising early results (Olney, 2023). In addition, new intelligent textbook prototypes are being built that employ generative AI to support new meaningful forms of interaction (Crossley et al., 2024). We expect more interesting works that will use the potential of these models in various forms of learner support around intelligent textbooks, especially if it builds on the success of the previous research. We hope that this special issue will be helpful for all researchers in the field of intelligent textbooks and will guide the future development of the field.

可结合某一主题的相关知识设计新形态教材/智能教材：例如，在线学习交互（等效交互定理、教学交互层次塔等）、探究社区（认知存在感、教学存在感、社会存在感等）、在线学习的网络支持（知识网络、社会网络、认知社会网络等）等等。

思路引导：

PS：可基于自身与小组对新形态教材/智能教材的思考与想法，通过智能体（https://www.coze.cn/s/E7-IUPuAgiA/）对话，查找相关文献或案例，以启发思考。

可聚焦讨论主题，结合以下问题贴例加以思考，或对同伴的观点进行评价或质疑：（每位学生需至少参与3个不同问题帖的讨论）

思考点/问题贴例：

1.直接用通用大模型问答替代教材内容可行吗？这会不会削弱学生深度阅读和系统化构建知识体系的传统优势？
2.将智能教材/新形态教材与MOOC/社交媒体（新元素）合并，能否真正促进协作学习，还是仅仅制造了更多碎片化的信息干扰？
3.理科强调逻辑推演，文科强调发散思维，目前的智能教材模板是否需要针对不同学科的认知规律进行深度改造？
4.传统的线性教材结构是否应调整（修改/放大/缩小）为完全的非线性网状结构？这种调整对学生的系统思维培养是利大于弊还是反之？
5.智能教材产生的大规模学情数据，除了用于教学，能否用于预测区域教育趋势或辅助公共教育政策制定？
6.如果学生过度依赖教材的自适应推荐，我们是否应该刻意“消除”部分便捷功能，强制培养其信息检索能力？
7.教材评价通常由专家制定，能否逆向由学生的实际学习效果来反向评判教材算法的优劣，并据此自动迭代？

PS：在讨论过程中，可通过与智能体展开对话（https://www.coze.cn/s/E7-IUPuAgiA/），协助思考，或与同伴开展观点辩论。

高质量帖子标准：1.能够引发多种观点；2.需要分析推理，而非简单查找；3.可以持续引发深入讨论；4.结合在线学习情境。

讨论内容需体现以下至少1种行为：解释（对观点进行阐述或补充）、质疑（提出不同观点或挑战）、整合（综合他人观点形成新理解）。