1 Introduction

Intelligent textbooks embed digital textbooks with intelligent tutoring technologies to provide intelligent reading support to students. Intelligent textbooks not only provide interactions that traditional digital textbooks have, such as highlighting, underlining, and note-taking, but also attempt to understand why readers interact with the textbooks and then build scaffoldings to enhance reading experiences. For example, the intelligent textbook Inquire Biology could actively ask the reader a question to promote deep thinking when the reader highlights a sentence. Also, the reader could raise questions to the textbook, which would respond to them using the reasoning technologies (Chaudhri et al. 2013). Over the last 30 years, intelligent textbooks have been used in many schools. Some recent empirical studies about the usage of intelligent textbooks have demonstrated their abilities to improve students’ learning gain (Chaudhri et al. 2014; Ericson 2019; Kim et al. 2020; Koć-Januchta et al. 2020).

This chapter offers a state-of-the-art overview of intelligent textbooks. The overview is divided into three parts. The first part focuses on the history of intelligent textbooks and attempts to answer the question: What are the intelligent textbooks and which authoring tools can be used to create the intelligent textbooks? The second part focuses on the technologies behind the intelligent textbooks and attempts to answer the question: What mechanism makes a textbook intelligent? The third part focuses on the usage of intelligent textbooks and attempts to answer the question: What is the effect of intelligent textbooks on students’ learning? The last section discusses the future and challenges of intelligent textbooks.

2 The Development of Intelligent Textbooks

The emerging of intelligent textbooks was driven by the idea of combining adaptive hypermedia systems and intelligent tutoring systems (ITS). An earlier attempt at intelligent textbook named ELM-ART was proposed by (Brusilovsky et al. 1996a, b) to develop an interactive and adaptive Web-based programming textbook with problem-solving support. The ELM-ART enables students to explore program examples by running them with different parameters, interactively solving problems, and receiving instant feedback. It also provides individual curriculum sequencing based on students’ learning status on the previously visited pages to suggest the next best pages to work on. Although ELM-ART can only offer adaptive multimedia, text presentation, as well as navigation support, it provides a design paradigm of the intelligent textbook that inspired many other studies of this area in the first decade of the twenty-first century.

With the rapid development of artificial intelligence (AI), the recent intelligent textbooks provide more sophisticated learning services, such as automatic resource matching, automatic question answering, personalized learning evaluation, and planning. For example, Interlingua is an intelligent platform where students can study textbooks in a foreign language supported by on-demand access to relevant reading material in their mother language (Alpizar-Chacon and Sosnovsky 2019). FlexBooks is a math & science textbook platform designed to suit learners’ learning styles, regions, languages, or skill levels and allows learners to customize content (Lindshield and Adhikari 2013). OpenDSA is an interactive textbook for data structures and algorithms courses involving the use of many algorithm visualizations and a wide range of automatic exercises assessment (Shaffer et al. 2011). Another tool for studying computer science is Runestone. It incorporates code visualizations and customizes interactive course materials (Miller and Ranum 2014). Reading Mirror is an online reading system that permits students to track their reading progress and compare with peers through a mirrored icicle plot visualization (Barria-Pineda et al. 2019). PASTEL is an online courseware authoring platform that applies embedded skill model and cognitive tutors to divide assessment items into clusters with similar semantic meanings and perform on-demand hints on how to perform the next step (Matsuda and Shimmei 2019). Other intelligent textbook authoring platforms are shown in (Table 1).

Table 1 Some intelligent textbook authoring platforms

3 Intelligent Tutoring Technologies of Intelligent Textbooks

The intelligent tutoring system Brusilovsky et al. (1996b) is formalized by three models: domain, student, and instruction. While it is designed to make use of students’ answering questions or testing data to intervene and regulate students’ learning in real-time, intelligent textbooks combine AI technologies with electronic textbooks; in addition to collecting the result data generated by the exercises and tests in the textbook, it also mines and analyzes the data generated during the process of using textbooks. Developing intelligent textbooks are based on the idea of ITS (Boulanger and Kumar 2019). The domain model is a knowledge base and ontology that stores and codifies a vast amount of knowledge of specific subjects via taxonomies, examples, exercises, and so on. The student model identifies a student’s knowledge state and how it evolves during learning. The instruction model specifies a policy for administering automated instructional actions that are conditioned on the student.

3.1 Domain Modeling Technologies in the Intelligent Textbook

The domain model provides the knowledge base of an intelligent textbook. Usually, an authoring tool or platform is required for instructors to manually create learning content, build scaffoldings, and link resources. This process is incredibly time-consuming and expensive, and some recent efforts are invested to develop automated modeling technologies to save expert effort. Domain knowledge is complicated and currently, we cannot expect technologies to generate delicate domain knowledge, but they can replace or assist humans in knowledge annotation. Knowledge annotation is a fundamental but critical component of intelligent textbooks as automated algorithms like machine learning algorithms need well-labeled data as the training samples. Without high-quality annotated data, intelligent linking, matching, and recommendation services could not be implemented. Current efforts in automatic knowledge annotation can be simply categorized into the following three categories.

The first approach is the automatic concept extraction that extracts concepts and knowledge from text automatically. Although a wide range of concept extraction methods has been developed, few have been applied in intelligent textbooks context. According to what features are used, three popular approaches for concept extraction are the pure word-based method (bag-of-words), chapter-based method (coarse-grained semantic-based), and latent topic-based method (fine-grained semantic-based). Huang et al. (2016) compared the three approaches and found that the latent topic-based method outperformed the others on predicting students’ knowledge acquisition state after reading textbooks. To extract concepts from text automatically, Chau et al. (2021) proposed a supervised feature-based machine learning method that uses multi-view features, including linguistic-based, statistics-based, title-based, and external resources-based features. The proposed method outperformed several state-of-the-art concept extraction approaches. Furthermore, some concept extraction technologies focus on using formatting rules and internal structures of textbooks (Alpizar-Chacon and Sosnovsky 2020) or discourse and text layout features of textbooks (Sachan et al. 2019). Although several new features and technologies can be used for concept extraction, their performances are still very low, which makes them not effective enough to use in real-world tasks. Human extraction is still the most reliable approach. Most recently, Wang et al. (2021) proposed a team-based systematic knowledge engineering approach for fine-grained concept annotation of textbooks.

The second approach is the automatic concept relationship extraction, including internal relationships (hierarchy concepts or prerequisite concepts) as well as external relationships. (Guerra et al. 2013) proposed a latent Dirichlet allocation (LDA)-based method to generate intelligent links among textbooks sections that presented a similar topic based on the LDA model. Wang et al. (2015) argued the concept hierarchy in textbooks is not only decided by the relatedness between the concept and the subchapter but also by the coherence between this concept and the concepts in the same/different subchapter(s). They furtherly formalized the concept extraction from the textbook as an optimization problem and combined local features and global features to train a support vector machine to extract concept hierarchies. Labutov et al. (2017) proposed two probabilistic graphical models to identify outcome and prerequisite concepts on six textbooks and demonstrated improvements over several baselines of automatic concept linking. Meng et al. (2017) explored multiple knowledge-based contents linking algorithms for connecting online resources with textbooks, and this algorithm reported its value for improving textbook subsection linking performance. Alpizar-Chacon and Sosnovsky (2021) presented an extensible linking model to enrich textbook contents connected with internal or external resources with the help of DBpedia.

A third strategy is to extract concepts and relationships among concepts simultaneously. For example, Lu et al. (2019) created a learning graph by classifying semantically similar chapters via an unsupervised clustering method, then extracted the structural relationship, and built the metro map by applying an integer linear programming-based technique. Wang et al. (2016) proposed a concepts extraction and concept relationship-building framework using the knowledge maps of textbooks. Sastry et al. (2017) extracted concept relationships through an elegant algorithm of the idea of transitive closure and visualized the concept relationship as a network graph. The Interlingua is an intelligent tool that links textbooks in different languages covering the same topic (Alpizar-Chacon and Sosnovsky 2019). The Interlingua first extracts index terms and pages referenced by the terms from the textbook and then uses them as semantic anchors to link pages and sections of the textbook to the concepts and through them to other textbooks available in the repository.

3.2 Student Modeling Technologies

An important feature of distinguishing an intelligent textbook from a normal digital textbook is whether it provides personalized learning services. Student modeling aims to understand students’ learning using their interaction data as they work on problems in the text. The student model drives the learning system to adapt to the needs and knowledge of students. Generally, a completed student model contains students’ knowledge state, behavior patterns, learning emotional state, as well as some domain-independent traits such as cognitive ability, learning style, motivation, and attitude.

One of the most popular student modeling approaches in ITS is “knowledge tracing,” which aims to predict students’ knowledge acquisition state using their performance data. Three popular knowledge tracing methods are Bayesian Knowledge Tracing (Corbett and Anderson 1995), logistic model (Pelánek 2017), and deep knowledge tracing (Piech et al. 2015). The Bayesian Knowledge Tracing uses a hidden Markov chain to estimate knowledge mastery probabilities, and the logistic model combines multiple factors that affect learning into a logistic regression model to make predictions; the deep knowledge tracing applies a long short-term memory neural network to model student learning. However, these well-explored approaches could not be directly used in intelligent textbooks, as these methods require students’ response data that is generated in solving problems, yet the most frequent learning activity in textbook-based learning is reading.

Recently, Mouri et al. (2016) analyzed the relationship between students’ e-book reading time and their final grade using the Bayesian network based on association analysis with social network analysis. They found that more time devoted to reading the e-book before the class was associated with a higher final grade. Meanwhile, Huang et al. (2016) incorporated the reading time variable into a Bayesian Knowledge Tracing model and two logistic models to predict students’ acquisition state on the concepts covered by a textbook. This study serves as the first step to construct a dynamic knowledge tracing model in intelligent textbooks. However, only considering reading time is not robust as students’ reading logs are noisy. For example, we cannot identify whether a student read a specific page even if he or she opened the page and kept it open there for a long time. Thaker et al. (2018) incorporated both the reading data and the performance data in an improved Bayesian Knowledge Tracing model. The comparison results show that the model using two-view data significantly outperformed the model that only uses reading data and the model that only considers quiz performance data. Furthermore, Thaker et al. (2019) presented a logistic model that also takes into account students’ previous performances and reading behaviors to predict their success rate for a given question. Okubo et al. (2018) also used students’ reading time in an e-book system and previous quiz scores to predict their final grades. Besides the reading time, other reading behaviors such as underlining and highlighting can also be used to predict students’ performance Okubo et al. (2017). Kim et al. (2020) investigated whether students’ comprehension and knowledge retention could be predicted by their highlighting behavior. The data analysis suggests that when students choose to highlight, the specific pattern of highlights can explain about 13% of the variance in observed quiz grades.

Students’ reading behavior also helps us to understand students’ preferences and cognitive features. For example, recent studies used clustering algorithms and lag sequence analysis to explore students’ reading behavior patterns in using an e-book. They found a very interesting phenomenon that students always use the memos and bookmarkers function rather than underlines and highlights (Yin et al. 2019; Yin and Hwang 2018). With students’ reading behavior data, Gu et al. (2020) applied multiple classification models, including logistic regression, support vector machine, and decision tree to predict students’ learning styles. The results show that the decision tree achieves promising performance in the prediction of learning style.

The domain-independent traits describe student profiles of cognitive ability, learning style, motivation, attitudes, working memory capacity, and emotions when using cognitive processing skills and strategies, such as induction and reasoning in the process of selecting and acquiring knowledge. A variety of technologies in cognitive science and psychometrics are being used to measure learners’ traits. For example, ELM-ART intelligent textbook platform can diagnose learners’ cognitive abilities changes of programming process based on example-based and constraint-based model (Weber and Brusilovsky 2016). A new didactical model for modern online textbooks was applied for developing student self-regulated competence (Railean 2010). A personalized recommendation mechanism was presented through some information about the individual cognitive levels and learning styles (Sun et al. 2013). Besides, some recent studies also used wearable smart devices like eye tracking (Ishimaru et al. 2016) and Kinect (Lin et al. 2017) to track students’ attention and emotional state.

3.3 Instructional Technologies

An instructional model takes the domain and student model as input and determines what next information to present to the student. This section summarizes several instructional technologies utilized in intelligent textbooks, including hyperlink annotation and direct navigation support, error-sensitive feedback, tutoring dialog instruction, and content presentation orders.

Hyperlink annotation and direct navigation support are the most frequently used instructional techniques in intelligent textbooks. Online textbooks contain several types of instructional resources, such as graphics, audio, videos, and plain texts. Hyperlink annotation instruction is used to create a nonlinear medium among these multimedia. The navigation support instruction is to guide learners through hyperspace by making direct next-link suggestions. Nowadays, these instruction techniques extend to intelligent links, semantic relationships, concept mapping, knowledge graphs, and so on. For example, KBS-HyperBook created intelligent links to external Web learning resources to satisfy learners’ knowledge, goals, and preferences on Java programming (Henze and Nejdl 2001). Wikibooks provided intelligent links instruction to the course concepts in the collaborative textbook (O’Shea 2011). Interlingua connected automated semantic relationships of sections and subsections across textbooks with on-demand access to relevant reading material in their mother tongue (Alpizar-Chacon and Sosnovsky 2019). MM4Books automatically build metro knowledge graphs among massive electronic textbooks (Lu et al. 2019). Another study proposed a concept mapping instruction method that allows students to link words in the textbook (Wang et al. 2017).

Error-sensitive feedback is an instruction technique to be given when learners answer a question incorrectly, are unsure of a correct answer, or repeatedly request help. This technology can not only judge whether an answer is correct or not but also mainly aim to fix students’ misunderstandings. For example, CS Circle tracked their programming progress and gave instant feedback on code exercises (Pritchard and Vasiga 2013). IntDynGeo Book offered hints and automatic corrections about geometry knowledge (Billingsley and Robinson 2005). Intextbooks developed interactive assessment question components to fix students’ knowledge concepts (Alpizar-Chacon and Sosnovsky 2020).

Tutoring dialog is an instructional technique that uses natural language processing to engage students in interactive dialogs. These tutoring dialogs often supply guidance for during problem-solving and motivational supports. For example, the intelligent textbook, Inquiry used inquiry-based instruction through a question-asking dialog to ask the student a question if they highlight a word or sentence (Chaudhri et al. 2014). Another intelligent textbook, MoFaCTS, provided a dialog system to correct student conceptual misunderstandings of cloze sentence practice contents (Pavlik et al. 2020). LiveHint is a dialog-driven textbook via a chatterbot with access to thousands of context-sensitive hints (Fisher et al. 2020).

Personalized content sequencing is another instruction technique that has the function of organizing sequential KCs and then presenting students with learning paths. One example is SmartBook, which implemented a tailor-made courseware solution for learners (Koychev et al. 2009). Another textbook is iRead that provided personalized learning content and activities by analyzing their profiles and reading history logs (Deligiannis et al. 2019).

Furthermore, there are other instruction techniques rarely used in intelligent textbooks. For example, in the intelligent textbook, Runstone applied the learning-by-doing strategy that encourages students to experiment with examples as they are reading (Ericson 2019). Runstone also provides a visualization tool to demonstrate and control the step-by-step execution of a program. Like Runstone, FlexBooks also provides an interactive simulation tool that supports learning by playing (Lindshield and Adhikari 2013).

4 Evaluation of Intelligent Textbooks

Reviewing the development in the past 10 years, researchers have carried out many empirical studies in schools, demonstrating the effectiveness of intelligent textbooks. According to these findings, intelligent textbooks were exceptional in facilitating students’ reading and learning. Meanwhile, combined with the users’ reflections of intelligent textbooks, the promising prospects of this new form of the digital textbook could be expected.

4.1 Students’ Comments on Intelligent Textbooks

It was gratifying that most students made positive evaluations of intelligent textbooks. Users’ evaluation of the popular intelligent textbook ELM-ART proved that students had high satisfaction with intelligent textbooks and expressed a strong willingness to continue to use them (Weber and Brusilovsky 2001). Another investigation shows that when students were faced with static PDF textbooks and interactive intelligent textbooks (their content was the same), students were more inclined to use intelligent textbooks (Pollari-Malmi et al. 2017). Most students believed that intelligent textbooks altered their learning patterns (Barria-Pineda et al. 2019).

Pursel et al. (2019) present an intelligent textbook authoring tool that can retrieve open educational resources from Wikipedia for users to create their books. The responses from the student survey indicated generally favorable reactions when asked questions about this intelligent textbook compared to a traditional textbook. Most recently, Feng and Li (2019) developed an offline-to-online intelligent textbook that grade and correct students’ calculation in a paper-based workbook automatically by cell phone’s camera and then use it to provide adaptive tutor service to students. An investigation showed that more than 30% have become active users and more than 20% of active users have recommended it to others.

4.2 The Effectiveness of Intelligent Textbooks

Inspired by the positive influence of social learning, the intelligent textbook Reading Mirror extended social navigation with social comparison. It enabled students to visually track their reading and test progress through icicle plots and compared them with their peers. Researchers have performed a series of classroom studies in three different courses. They proved that the Reading Mirror could help students (N = 200) focus on the most important pages and increase their reading engagement. The social comparison would encourage students to work harder and achieve higher achievement in quizzes (Barria-Pineda et al. 2019). Researchers have used Runestone to create several free intelligent textbooks for introductory computing courses. By analyzing the log files, they reported that owe to various interactive components, the intelligent textbooks created by Runestone improved students’ learning gains and motivation in programming (Ericson 2019). The results of a large-scale study (N > 600) showed that in programming courses, interactive intelligent textbooks were more conducive to enhancing students’ learning motivation, gains, and feedback on learning resources than static PDF format textbooks (Pollari-Malmi et al. 2017).

Intelligent textbooks also exerted unexpected benefits for teachers. In a small simple study of high school teachers (N = 10), they used the intelligent textbooks developed by Runestone, which helped them improve their professional knowledge and teaching confidence (Ericson et al. 2015). Some studies also showed the positive effect of intelligent textbooks in improving students’ academic performance. For example, Inquire Biology significantly improved students’ homework quiz scores (p = 0.02) and quiz scores (p = 0.05) (Chaudhri et al. 2013). The intelligent textbook created based on ELM-ART significantly improved the test scores of those students with weak programming skills (p = 0.011) (Weber and Brusilovsky 2001).

It was worth noting that not all intelligent textbooks could help students achieve expected learning gains. Just like the Math CyberBook (Matsuda and Shimmei 2019), it did not achieve a significant impact on students’ academic performance (p = 0.63). The reason for this phenomenon needs further analysis. Moreover, students did not achieve the expected learning progress in the first 3 weeks of using the intelligent textbook created by Reading Mirror (Barria-Pineda et al. 2019). After comparative analysis, researchers believed that one of the reasons that could explain this issue was that students needed time to adapt to the social comparison feature. Maybe it proved that some external conditions should be satisfied for its desired functions to work.

5 Discussions and Conclusions

Intelligent textbooks have attracted much attention in the past decade, with increasing evidence demonstrating their positive influences on improving students’ reading and learning. A short review of tools, adaptation technologies, and evaluations provided in this chapter could serve as a collection of useful information for the researchers and developers of the next generation of intelligent textbooks. Although intelligent textbook research has made big progress in the past decade, many crucial technical and usage problems remain unsolved. For example, current technologies cannot understand the mathematical language within the textbooks very well, which seriously hinders the development of mathematical intelligent textbooks. Also, authoring a new intelligent textbook is expensive, so while making the huge quantity of existing PDF-based digital textbooks intelligent is very necessary, it is challenging (Alpizar-Chacon et al. 2021). Another area of future work is interconnecting intelligent textbooks, learning management systems, practices, and exams to construct a closed intelligent learning loop.