A transformer based approach to STEAM integrated english course design in high schools under deep learning Authors: Wenjuan Fan Published: 2025 ABSTRACT This study explores the integration of the educational philosophy of Science, Technology, Engineering, Arts, and Mathematics (STEAM) with deep learning technologies for the development of high school English course resources. By introducing deep learning models, particularly the Transformer architecture, the study enables automatic generation and personalized delivery of instructional content, thereby enhancing the intelligence and relevance of course resources. An innovative Transformer-based framework incorporating STEAM-related semantic features is designed and implemented. The framework is trained on a large-scale, preprocessed English text dataset collected from 300 high school students participating in STEAM-integrated English learning tasks, with data from 150 students forming the core corpus for model training. The model employs an encoder-decoder architecture and is optimized using the Adam optimizer to support the generation of high-quality content tailored to diverse teaching objectives. Through on-site teaching experiments conducted with the 300 high school students, the results demonstrate that, the model’s performance is rigorously evaluated through metrics such as Bilingual Evaluation Understudy score (0.78), innovation index (0.75), and content diversity (0.70). Results reveal that the proposed model outperforms comparative models (Model A and Model B) in language generation accuracy, innovation, and content diversity, highlighting its superior capability in producing high-quality English teaching resources. Although the model demonstrates promising results, its implementation in resource-limited environments presents challenges, highlighting the need for further refinements. Nonetheless, these findings demonstrate the model’s potential to revolutionize English teaching by enhancing resource quality and offering a more efficient, innovative approach to course development. Introduction In recent years, the Science, Technology, Engineering, Arts, and Mathematics (STEAM) education concept has become one of the key directions in global educational reform1–3. However, in high school English teaching, the traditional curriculum resource development model still faces significant challenges. Most existing resources primarily focus on language knowledge and skill training, with relatively monotonous content that lacks deep integration with disciplines such as science, technology, engineering, arts, and mathematics. This makes it difficult for students to develop interdisciplinary thinking and comprehensive application abilities in language learning4–6. Meanwhile, although some schools have attempted to introduce project-based learning and interdisciplinary courses, practical implementation often struggles to be effectively carried out due to the lack of high-quality, dynamically generated curriculum resources7–9. With the advancement of artificial intelligence (AI) technology, deep learning—particularly the Transformer model in the field of Natural Language Processing (NLP)—has demonstrated potential in the automated generation of educational content10–13. However, existing research primarily focuses on vocational education or STEAM courses in engineering, while the development of intelligent resources for high school English courses remains relatively underdeveloped. Most approaches still rely on predefined templates or single media formats, lacking the embedding of semantic-level STEAM features and dynamic optimization mechanisms. Consequently, the generated resources exhibit notable deficiencies in innovation, diversity, and pedagogical adaptability. For example, Rahim et al.14 evaluated the effectiveness of project-based learning integrated with STEAM education in a metal welding course involving 64 vocational students from Negeri Padang. Compared to traditional methods, the STEAM-integrated approach significantly enhanced students’ academic progress. Similarly, Hafis et al.15 improved the quality of shielded metal arc welding instruction by incorporating interactive media videos. The use of these videos led to improved student learning outcomes, with an average score of 77.88 in the first cycle, categorized as moderate. Ernawati et al.16 investigated problem-based learning models to improve performance in fashion design courses. Their findings showed that 55% of students met minimum competency standards after the first cycle, emphasizing the importance of iterative refinement in improving cognitive outcomes. These studies highlight the effectiveness of problem-based learning and iterative improvement in enhancing student learning outcomes. Similar studies include the work of Prasetyo et al.17, who developed and evaluated the effectiveness of short instructional videos for teaching basic lathe operations in mechanical engineering. The results demonstrated that the short video tutorials significantly enhanced students’ understanding and cognitive abilities, although variations in comprehension levels were observed. Setyawan et al.18 adopted a pre-test/post-test experimental design to implement video-based learning in an educational setting. The study involved 33 students majoring in Computer and Network Engineering. Findings revealed a significant improvement in students’ knowledge after engaging in video-based instruction. To fill the research gap mentioned above, this study proposes an intelligent resource generation framework based on Transformer architecture, aiming to achieve a deep integration of STEAM education philosophy and high school English curriculum. This framework supports the automatic generation of personalized and interdisciplinary teaching content by introducing a semantic control mechanism guided by STEAM knowledge graph. The main contributions of this study include three aspects: 1. proposing a Transformer model structure that integrates STEAM semantic features, expanding the application scope of this model in non-STEM disciplines; 2. designing a set of resource generation and evaluation processes for high school English teaching, significantly enhancing the innovation and teaching adaptability of resources; 3. Verifying the superior performance of the model in terms of language quality, interdisciplinary integration, and user satisfaction through multidimensional experiments. To validate the effectiveness of the proposed model, this study focuses on three research questions: (1) How to integrate the STEAM educational philosophy into high school English curriculum design through deep learning methods? (2) How to build an English teaching resource generation model with scalability and educational adaptability? (3) Can this model significantly enhance students’ learning interest, creativity, and interdisciplinary application ability in practical teaching? Based on these considerations, the following research hypotheses are proposed, and on-site teaching experiments are conducted for 300 high school students and multiple teachers. The comprehensive evaluations are conducted from multiple aspects such as learning effectiveness, teacher feedback, and resource quality. Hypothesis 1 (H1) The deep learning model based on the Transformer architecture can generate English teaching resources with stronger STEAM characteristics, outperforming comparative models in terms of content innovation, interdisciplinary integration, and semantic completeness. Hypothesis 2 (H2) Students taught using resources generated by this model will demonstrate higher levels of learning interest, knowledge acquisition, and interdisciplinary application abilities. Hypothesis 3 (H3) The model possesses good teaching adaptability and scalability, reflected by higher scores in teaching convenience and resource diversity in teacher feedback. Introduction to course design The STEAM education concept emphasizes interdisciplinary integration to cultivate students’ comprehensive abilities and innovation skills19. According to Thapaliya20, STEAM education aims to develop students’ problem-solving, critical thinking, and creative skills, preparing them to address future societal and workplace challenges. The use of deep learning in education is increasing, with Li & Kim21 reporting that deep learning technologies can generate personalized learning resources and predict learning outcomes by processing large amounts of educational data, offering new possibilities for educational reform. Al Ka’bi22 proposed a model to enhance students’ cognitive abilities, which outperformed existing algorithms in comparison. Currently, the development of high school English course resources primarily relies on textbooks and teacher expertise. This leads to challenges such as repetitive content, limited innovation, and insufficient interdisciplinary integration23. As Zhang & Ma24 point out, traditional teaching methods often fail to address students’ holistic developmental needs. In contrast, STEAM education offers fresh perspectives and methods for English instruction. For example, Martins & Baptista25 adopted a pre-test and post-test experimental design to assess changes in participants’ content knowledge before and after professional development training. The findings indicate that the training had a positive impact on improving teachers’ content knowledge. Moreover, the study helps to define the specific content knowledge required for interdisciplinary teaching that integrates science, technology, engineering, arts, and mathematics—namely, STEAM pedagogical content knowledge. Additionally, Park & Cho26 explored the intersections between learning objectives and STEAM goals, providing valuable insights for future research and practical applications in integrated curricula. Challenges remain in the development of high school English course resources27. Traditional teaching models often lack interdisciplinary integration, failing to meet the diverse developmental needs of students. Additionally, current methodologies for generating educational resources are limited by data and technological constraints, preventing the creation of personalized and interdisciplinary materials. As a result, there is a critical need to explore new approaches and technologies to address these gaps and drive innovation in high school English course development. This study aims to investigate how to effectively combine deep learning technology and interdisciplinary content to improve the quality and effectiveness of teaching resources. By analyzing students’ learning behaviors and needs, the study generates personalized teaching materials, enhancing learning efficiency and outcomes. At the same time, it explores integrating STEAM content into English instruction to promote interdisciplinary thinking and competencies, addressing existing research gaps. Based on the literature review, a high school English course is designed using the STEAM concept. A one-week course is presented as an example to illustrate specific teaching activities and tasks. This course, guided by STEAM principles, aims to enhance students’ English proficiency and overall literacy through interdisciplinary methods. It incorporates content from multiple disciplines, fostering innovation and critical thinking through project-based and task-oriented learning approaches. The course design follows the STEAM education concept, integrating knowledge and methodologies from various disciplines to highlight their interconnections and interactions. Project- and task-based teaching methods enable students to apply what they have learned in real-world contexts, enhancing both practicality and engagement. Project-based teaching emphasizes long-term involvement with complex tasks, often requiring interdisciplinary collaboration and culminating in tangible products or presentations. This approach encourages students to apply knowledge to real-world scenarios, fostering critical thinking and problem-solving skills. In contrast, task-based teaching involves shorter, more focused activities aimed at achieving specific learning objectives. These tasks improve language skills through authentic communication, requiring students to immediately apply new vocabulary or grammar. Both methods encourage independent thinking, innovation, and tackling complex challenges, nurturing creativity and critical thinking. Collaborative projects further promote teamwork, enhancing students’ communication and collaboration skills. Table 1 presents a one-week course plan demonstrating the application of STEAM education principles in high school English teaching. The materials used in these courses combine content from existing English textbooks with supplementary materials selected to align with STEAM principles. This approach ensures the course content not only meets course requirements but also embodies the integrative and innovative nature of STEAM education.Table 1Example of a one-week course plan.DateThemeObjectiveActivityTaskMondayEnvironmental ProtectionImprove reading and writing skillsRead environmental protection articles and discussWrite environmental protection reports and create PPTsTuesdayTechnologyEnhance listening and speaking skillsWatch technology videos and participate in discussionsCreate technology-themed videosWednesdayEngineeringCultivate communication and cooperation skillsDesign simple engineering projectsWrite project reports and create modelsThursdayArtImprove creative thinking and expression skillsLearn poetry and create artParticipate in poetry recitation or drama performancesFridayMathematicsCultivate logical thinking skillsLearn mathematical terms and solve problemsComplete mathematical homework and make presentations Table 1Example of a one-week course plan.DateThemeObjectiveActivityTaskMondayEnvironmental ProtectionImprove reading and writing skillsRead environmental protection articles and discussWrite environmental protection reports and create PPTsTuesdayTechnologyEnhance listening and speaking skillsWatch technology videos and participate in discussionsCreate technology-themed videosWednesdayEngineeringCultivate communication and cooperation skillsDesign simple engineering projectsWrite project reports and create modelsThursdayArtImprove creative thinking and expression skillsLearn poetry and create artParticipate in poetry recitation or drama performancesFridayMathematicsCultivate logical thinking skillsLearn mathematical terms and solve problemsComplete mathematical homework and make presentations Example of a one-week course plan. Student learning outcomes are evaluated through weekly logs, project reports, classroom performance, and final tests, ensuring their holistic development within the STEAM framework. Appendix 1 provides a detailed course outline, including course objectives, content, teaching strategies, and evaluation methods, for educator reference. The developed model can also generate English learning materials related to subjects such as physics or biology. For instance, by introducing the basic principles of “photosynthesis,” the model produces English articles that incorporate biological terminology and concepts, while also designing exercises to help students grasp scientific principles and enhance their English proficiency. Similarly, the model aids teachers in generating English teaching materials for engineering topics, such as “mechanical principles” or “bridge design.” These materials include explanations of engineering terminology and demonstrate how these terms apply to real-world problem analysis, thus improving students’ academic English skills. For art courses, the model can generate English materials on topics like art analysis and art history. For example, when discussing “Renaissance art,” the model helps students explore the background, styles, and key works of art in English. To ensure that the STEAM educational philosophy can be systematically integrated into high school English curriculum, this study has constructed a clear instructional design framework. This framework, based on STEAM core principles such as problem oriented learning, interdisciplinary integration, and hands-on practice, is mainly implemented through project-based learning and task driven teaching methods. These principles and methods collectively guide the design of specific course activities and tasks, and their overall structure is shown in Fig. 1.Figure. 1Framework for high school English curriculum design under the STEAM concept. Framework for high school English curriculum design under the STEAM concept. Figure 1 illustrates the core structure of this course design. The core teaching principles of STEAM located at the top level are the guiding ideology for the entire curriculum design. The principle of problem-oriented learning emphasizes guiding students to face and solve complex problems in the real world. The principle of interdisciplinary integration requires breaking down disciplinary barriers and organically integrating knowledge from fields such as science, technology, engineering, art, and mathematics into English learning. The principle of hands-on practice emphasizes students’ personal experience and knowledge application, learning by doing. Guided by these principles, two core teaching strategies are adopted. Project-based learning is a long-term exploratory teaching method that revolves around complex and real-world problems, requiring students to collaborate and ultimately create publicly available outcomes. The task driven teaching principle is centered on completing specific and realistic language tasks, focusing on the meaning and application of language during task execution, to develop students’ language and communication abilities. Ultimately, these strategies are concretized into a series of course activities and tasks. For example, activities such as writing environmental reports and creating presentations are a reflection of project-based learning, problem orientation, and interdisciplinary integration principles. The tasks of producing technology themed videos, designing engineering models, participating in art performances, and completing mathematical demonstrations integrate hands-on practice with interdisciplinary knowledge, forming a complete and rich STEAM English learning experience. Through this clear path, the concept of STEAM education can be integrated from theoretical principles into the teaching practice of every class. Research methodology Integration of deep learning-based STEAM education into high school english course development This study aims to integrate deep learning technology with the STEAM education concept to create personalized, interdisciplinary high school English course resources. The STEAM education model incorporates science, technology, engineering, arts, and mathematics, focusing on fostering innovation, critical thinking, and problem-solving skills through interdisciplinary approaches. It emphasizes the integration of theory and practical application, encouraging students to acquire and apply knowledge in real-world contexts. To incorporate the STEAM concept into the English course, this study developed a series of teaching activities and projects. For example, in project-based learning, students research and discuss environmental protection issues, then write English reports and present their findings using tools such as PowerPoint and video production software. During this process, students engage with scientific literature, analyze data using technological tools, and present their research in English. This project enhances students’ writing and communication skills while also developing their scientific literacy and technological proficiency. Additionally, interdisciplinary inquiry projects combine history, art, and engineering, where students design artworks or models inspired by English literary works. By exploring these literary texts, students gain historical and cultural insights, applying their engineering and artistic knowledge to create their projects. This approach fosters the integration of English learning with other academic subjects. In Task-Based Learning, students select a science and technology topic for an English speech, utilizing multimedia tools for presentations. They engage in researching scientific materials, drafting speeches, and preparing presentation materials. This process enhances both their English speaking skills and their understanding of science and technology concepts. Another task involves integrating programming with literature. As students learn basic programming principles, they design an interactive story based on a literary work, practicing both programming and English writing skills. This activity requires students to translate the plot and characters from literature into programming code to create a basic interactive narrative. It deepens their comprehension of literary texts while honing essential programming skills. To enable the adaptation of this course design by other educators, specific teaching strategies and methods are outlined. First, clear objectives and expected outcomes should be articulated for each project or task, providing direction for teaching activities. Second, preparing relevant resources—such as scientific literature, technical equipment, and multimedia tools—supports students in completing their tasks. Third, detailed teaching steps should be provided, including the implementation process, time allocation, and specific assignments. Teachers should offer support and feedback during instruction to help students navigate challenges and ensure project progression. Lastly, evaluation and feedback mechanisms enable educators to assess students’ learning outcomes and project results, refining the course design based on feedback to improve its effectiveness. In addition, this study reclassifies the difficulty levels of vocabulary and grammar within the learning materials to better align with students’ varying proficiency levels. The contextual expression and situational tasks related to STEAM themes are also enhanced—for example, embedding English writing assignments within scientific experiment descriptions or engineering design projects—to increase the practicality and interdisciplinary integration of the content. From an instructional perspective, the balance between task-based learning and collaborative group activities is adjusted based on student engagement data, aiming to improve classroom participation and the depth of interaction. Furthermore, in response to feedback regarding mismatches in learning pace, modular teaching units with adjustable pacing are introduced. These units allow teachers to flexibly arrange content delivery, thereby better accommodating the diverse learning rhythms and preferences of different student groups. These teaching strategies and methods demonstrate the integration of STEAM education concepts into high school English courses, enhancing the course design’s practicality and adaptability. The methods and case studies serve as references for educators, helping them apply STEAM education principles more effectively in their teaching practice. The model presented in this study consists of four key stages: data processing, feature extraction, model training, and resource generation, as illustrated in Fig. 2. This architecture supports the integration of deep learning techniques with STEAM education principles, enabling the creation of customized resources that enhance the delivery of high school English curricula.Figure. 2Framework of integrating STEAM education into high school English course resource development based on deep learning. Framework of integrating STEAM education into high school English course resource development based on deep learning. Figure 2 outlines the framework for integrating STEAM education into the development of high school English course resources using deep learning methodologies28–30. The foundation of this framework is data processing, a crucial step for constructing and training an effective model. The process begins with data collection, which involves gathering datasets from high school English textbooks, STEAM-related educational materials, and records of student learning behaviors. English textbooks provide textual content, exercises, and supplementary resources related to key chapters. STEAM educational materials include a variety of content and project cases from science, technology, engineering, arts, and mathematics. Records of student learning behaviors capture data such as study durations, assignment completion, and exam performance. After data collection, a cleaning process is applied to remove noise and address missing values, ensuring dataset integrity. For textual data, this step includes eliminating punctuation, special characters, and irrelevant content. During the data preprocessing stage, the text data undergoes detailed annotation, including Part-of-Speech (POS) tagging and syntactic dependency parsing. POS tagging follows the widely recognized Penn Treebank annotation guidelines, ensuring consistency and general applicability within the field of natural language processing. Syntactic dependency annotation is based on the Universal Dependencies (UD) framework, which provides a cross-linguistically consistent method for representing syntactic structures, facilitating unified analysis and processing across diverse text domains. The annotation process combines automated tagging tools with manual proofreading to enhance the accuracy and quality of the annotations. Word embedding techniques, specifically Word2Vec, are used to preprocess the text. Following data processing, the next step is feature extraction, which involves extracting relevant information from the dataset. NLP techniques are employed to extract features from textual data, making it suitable for model training. The primary steps in this process are as follows: Word embedding: A pre-trained word embedding model, such as Word2Vec, is used to convert text into numerical vectors that capture the semantic relationships between words. For a sentence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S = \left\{ {w_{1} ,w_{2} , \ldots ,w_{n} } \right\}$$\end{document}, its word embedding representation is given by Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{S}} = \left\{ {{\text{v}}\left( {w_{1} } \right),{\text{v}}\left( {w_{2} } \right), \ldots ,{\text{v}}\left( {w_{n} } \right)} \right\}$$\end{document} Text feature extraction: The TF-IDF method is employed to extract key features from the text. This approach evaluates the importance of each word in a document by considering two factors: term frequency, which measures how often a word appears in a specific document, and inverse document frequency, which reflects the rarity of the word across the entire corpus. Term frequency indicates the relative occurrence of a word within a single document, while inverse document frequency highlights the extent to which the word is uncommon across all documents—assigning higher weights to rarer terms. By multiplying these two components, the TF-IDF score for each word is obtained, effectively capturing both the representativeness and contextual significance of the term within the document. Learning behavior feature extraction: Key features are extracted from students’ learning behavior records, including study duration, frequency, and grade changes. These features are essential for generating personalized teaching resources. Word embedding: A pre-trained word embedding model, such as Word2Vec, is used to convert text into numerical vectors that capture the semantic relationships between words. For a sentence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S = \left\{ {w_{1} ,w_{2} , \ldots ,w_{n} } \right\}$$\end{document}, its word embedding representation is given by Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{S}} = \left\{ {{\text{v}}\left( {w_{1} } \right),{\text{v}}\left( {w_{2} } \right), \ldots ,{\text{v}}\left( {w_{n} } \right)} \right\}$$\end{document} Word embedding: A pre-trained word embedding model, such as Word2Vec, is used to convert text into numerical vectors that capture the semantic relationships between words. For a sentence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S = \left\{ {w_{1} ,w_{2} , \ldots ,w_{n} } \right\}$$\end{document}, its word embedding representation is given by Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{S}} = \left\{ {{\text{v}}\left( {w_{1} } \right),{\text{v}}\left( {w_{2} } \right), \ldots ,{\text{v}}\left( {w_{n} } \right)} \right\}$$\end{document} Text feature extraction: The TF-IDF method is employed to extract key features from the text. This approach evaluates the importance of each word in a document by considering two factors: term frequency, which measures how often a word appears in a specific document, and inverse document frequency, which reflects the rarity of the word across the entire corpus. Term frequency indicates the relative occurrence of a word within a single document, while inverse document frequency highlights the extent to which the word is uncommon across all documents—assigning higher weights to rarer terms. By multiplying these two components, the TF-IDF score for each word is obtained, effectively capturing both the representativeness and contextual significance of the term within the document. Text feature extraction: The TF-IDF method is employed to extract key features from the text. This approach evaluates the importance of each word in a document by considering two factors: term frequency, which measures how often a word appears in a specific document, and inverse document frequency, which reflects the rarity of the word across the entire corpus. Term frequency indicates the relative occurrence of a word within a single document, while inverse document frequency highlights the extent to which the word is uncommon across all documents—assigning higher weights to rarer terms. By multiplying these two components, the TF-IDF score for each word is obtained, effectively capturing both the representativeness and contextual significance of the term within the document. Learning behavior feature extraction: Key features are extracted from students’ learning behavior records, including study duration, frequency, and grade changes. These features are essential for generating personalized teaching resources. Learning behavior feature extraction: Key features are extracted from students’ learning behavior records, including study duration, frequency, and grade changes. These features are essential for generating personalized teaching resources. To enhance the iterative nature of the study and the model’s capacity for continuous improvement, an iterative optimization mechanism based on learning feedback is integrated into the model’s operational framework. Specifically, at the end of each instructional phase, student performance data—including test scores, questionnaire feedback, and behavioral analytics—are collected and fed into the model’s evaluation module. The data inform adjustments to the content generation strategy and the logic for instructional recommendation generation. For example, if a particular type of instructional resource is repeatedly flagged as “difficult to understand” or “lacking engagement,” the model will reoptimize the selection and presentation of content based on its structural tags. Additionally, a regular calibration cycle is established—set at biweekly intervals—to allow for rolling updates based on trends in model performance and student learning outcomes. This mechanism not only enhances the model’s adaptability to real-world teaching scenarios but also enables a transition from static content output to dynamic content evolution, thereby empowering the system to continually improve its cognitive effectiveness and instructional efficiency. To further enhance the adaptability of the teaching model and the precision of its feedback mechanisms, this study incorporates a fuzzy logic evaluation system into the existing framework of observational records and qualitative assessments. This addition aims to improve the flexibility and granularity of evaluating students’ interdisciplinary learning outcomes. Traditional scoring systems often struggle with ambiguity and rigid grading scales when assessing subjective dimensions such as creativity, coherence in language expression, and the depth of interdisciplinary integration. By constructing membership functions and fuzzy rules, the fuzzy logic system allows for more progressive and flexible quantification of such indicators. The study adopts a hybrid approach that combines expert knowledge with data-driven refinement to define rules and membership functions. Initially, teaching experts in English and STEAM collaboratively establish a rule framework based on task-oriented assessment standards. Fuzzy levels such as “insufficient,” “basic,” “proficient,” and “excellent” are defined using triangular and trapezoidal membership functions. This framework is then calibrated using empirical data from classroom observations and student project outcomes to ensure both educational relevance and practical interpretability. During implementation, the fuzzy logic system integrates student behavioral data, linguistic output quality, and teacher evaluations to provide real-time feedback for content generation and instructional strategy adjustment. This supplementary mechanism significantly improves the model’s feedback loop capabilities, offering a more feasible technical foundation for the intelligent optimization of interdisciplinary English teaching resources. Selection of deep learning algorithms Following feature extraction, model training involves applying deep learning algorithms to generate personalized and interdisciplinary English course resources. The selection of the deep learning algorithm is crucial. In this study, the Transformer model is chosen as the primary algorithm due to its superior performance in NLP tasks. The Transformer model excels at handling long-range dependencies and detecting intricate patterns in sequence data. A key feature of its architecture is the self-attention mechanism, which processes sequence data in parallel, distinguishing it from traditional recurrent neural network (RNN) and improving both computational efficiency and the ability to capture long-range dependencies. The Transformer model consists of two main components: the encoder and the decoder. In this study, both components are composed of six identical stacked layers. This configuration is selected based on a balanced consideration of model expressiveness and computational efficiency. Each layer includes two core components: a multi-head self-attention mechanism and a feed-forward neural network. For the attention mechanism, the model employs eight parallel attention heads, each with a dimensionality of 64, resulting in a total model dimension of 512 (i.e., 8 × 64). This configuration enables the model to capture semantic information across different subspaces in parallel, thereby enhancing its ability to model complex contextual relationships in interdisciplinary English texts. Regarding the choice of network depth, ablation tests are conducted on architectures with 4, 6, and 8 layers. The results indicate that the 6-layer configuration performs best across key evaluation metrics. Although the 8-layer model shows marginal improvement on certain indicators, it significantly increases computational cost and exhibits a higher risk of overfitting when applied to medium-scale datasets. Therefore, the 6-layer architecture achieves a strong balance between training efficiency, model stability, contextual modeling capability, and semantic representation. The overall architecture of the Transformer model is illustrated in Fig. 3.Figure. 3Structure of the transformer model. Structure of the transformer model. Figure 3 illustrates the structure of the Transformer model, which consists of multiple identical layers in both the encoder and decoder. Each layer contains two key components: the self-attention mechanism and the feed-forward neural network. The self-attention mechanism captures contextual information by evaluating the relevance between each position in the input sequence and all others. The multi-head self-attention mechanism allows the model to gather information from multiple subspaces concurrently. The attention mechanism is defined by the formula in Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{Attention}}\left( {Q,K,V} \right) = {\text{softmax}}\left( {\frac{{QK^{T} }}{{\sqrt {d_{k} } }}} \right)V$$\end{document} In Eq. (2), Q(Queries), K(Keys), and V(Values) are matrices derived from the input sequence through linear transformations, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{k}$$\end{document} represents the dimension of the keys. The feed-forward neural network applies further nonlinear transformations to the self-attention output, comprising two fully connected layers and a ReLU activation function, as shown in Eq. (3):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{FFN}}\left( x \right) = max\left( {0,xW_{1} + b_{1} } \right)W_{2} + b_{2}$$\end{document} The decoder, also composed of multiple identical layers, includes three sub-layers. Its multi-head self-attention mechanism is similar to the encoder’s, but it only computes attention over the current and preceding positions to prevent information leakage from future positions. The encoder-decoder attention mechanism combines the decoder’s current input with the encoder’s output to capture input–output relationships. The decoder’s feed-forward neural network is identical to that in the encoder. In this study, the Transformer model is used to generate personalized and interdisciplinary high school English course resources, as outlined in Fig. 4.Figure. 4Steps for developing applied course resources based on the Transformer model. Steps for developing applied course resources based on the Transformer model. The development of applied course resources follows several key steps. First, high school English textbooks, STEAM-related educational materials, and student learning behavior records are preprocessed and converted into a format compatible with the model. During the training phase, the preprocessed text data is input into the Transformer model. The self-attention mechanism enables the model to capture contextual information and long-range dependencies within the text. For position pos in the sequence and the index i in the embedding dimension, the positional encodings are computed as follows:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PE_{{\left( {pos,2i} \right)}} = {\text{sin}}\left( {\frac{pos}{{10000^{2i/d} }}} \right)$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PE_{{\left( {pos,2i + 1} \right)}} = {\text{cos}}\left( {\frac{pos}{{10000^{2i/d} }}} \right)$$\end{document} Here, pos represents the position of the word in the sentence, i signifies the index of the vector dimension, and d denotes the dimension of the word embedding. During the model optimization phase, a loss function (cross-entropy loss) and an optimization algorithm are used to adjust the model’s parameters, minimizing the loss function to improve performance. Once trained, the Transformer model can generate English course resources that align with the STEAM education concept. By incorporating students’ learning needs and background information, the model creates personalized teaching materials that integrate content from science, technology, engineering, arts, and mathematics, thereby enriching and enhancing the practicality of the resources. The Transformer model offers key advantages in this process. Its self-attention mechanism captures long-distance dependencies within sequence data, which is vital for processing textual information. Additionally, the model can process all positions within a sequence simultaneously, improving computational efficiency. Its versatility also extends to various NLP tasks, such as text generation, translation, and summarization, making it adaptable to diverse applications. By leveraging these capabilities, the Transformer model effectively generates personalized, interdisciplinary high school English course resources, boosting student engagement and supporting the development of comprehensive literacy. Experimental process and data collection To comprehensively evaluate the effectiveness and practical potential of the deep learning-based English instructional model integrated with the STEAM educational philosophy, this study employed a diversified data collection approach and experimental design to ensure multi-dimensional insights into the model’s performance. This study was conducted in public high schools from September to December 2023, involving 300 high school English learners and 12 English teachers to ensure sample representativeness. All participants collectively constituted the overall sample of this experiment and participated in extensive data collection activities. Among them, the classroom activity texts, learning management system data, feedback and evaluation of 150 students were systematically collected and preprocessed, forming a dedicated dataset for training the Transformer model proposed in this study. To evaluate the effectiveness of different teaching interventions, this study employs an inter-group experimental design. Randomly select 150 students from 300 participants and evenly distribute them into three experimental groups of 50 each, receiving teaching interventions generated based on Model A, Model B, or the Transformer-based model proposed here. This design facilitates direct comparison between different teaching methods and enhances the external validity of research results, providing robust empirical support for the applicability of the model in different educational environments. All instructional activities are conducted in English to maintain consistency between the language environment and the teaching content. The instructional design integrates materials from STEAM disciplines. For instance, in the “Future City” project, students conceptualize topics such as smart transportation and green architecture, compose creative texts in English, design visual representations, and present and discuss their ideas orally. These activities reflect a strong interdisciplinary integration. Student behavioral data are collected through real-time classroom observations and teacher records, focusing on collaboration skills, task performance, language expression, and learning engagement. Teachers also write reflective teaching notes to document strategies and student responses, serving as feedback for model optimization. In addition, the study gathers student feedback through periodic questionnaires, group discussions, and interviews. These cover students’ understanding of STEAM tasks, changes in learning interest, and development of critical thinking and creativity. Feedback is collected from students of varying proficiency levels to ensure representativeness and diversity. Learning behavior data are also captured via a learning management system, which records online study duration, module completion rates, interaction frequency, and content depth. Particular attention is given to students’ engagement and performance in interdisciplinary projects, supporting the analysis of learning trajectories and refinement of instructional strategies. Regarding teacher feedback, interviews and surveys are conducted regularly to collect evaluations on the model’s adaptability, resource generation efficiency, and instructional effectiveness. Teachers focus on the model’s impact on student engagement, interdisciplinary comprehension, and its integration with existing instructional frameworks. Their feedback informs adjustments to the model’s generation logic and teaching strategy refinement. All data collection strictly adheres to ethical guidelines for educational research. Approval is obtained from school administrators prior to the study, and informed consent is secured from all participants. Students and parents are clearly informed of the data usage and confidentiality measures. Throughout the study, all data are anonymized to protect participants’ privacy and ensure data security. Experimental environment & parameters setting The experimental setup is summarized in Table 2, detailing the hardware, software platforms, and development environments used in the study:Table 2Experimental environment settings.Hardware equipmentSoftware platformDevelopment environmentNvidia GTX 1080 Ti GPUPython 3.8Jupyter NotebookIntel Core i7-8700 CPUTensorFlow 2.6PyCharm16 GB RAMKeras 2.6Google Colab Table 2Experimental environment settings.Hardware equipmentSoftware platformDevelopment environmentNvidia GTX 1080 Ti GPUPython 3.8Jupyter NotebookIntel Core i7-8700 CPUTensorFlow 2.6PyCharm16 GB RAMKeras 2.6Google Colab Experimental environment settings. The parameters were configured based on the experimental requirements and model performance, as outlined in Table 3:Table 3Experimental parameter settings.ParameterValueLearning rate0.001Batch size32Number of network layers6Learning rate decayAdaptively adjusted through the validation setMaximum sequence length256Dropout rate0.2Gradient clipping5.0 Table 3Experimental parameter settings.ParameterValueLearning rate0.001Batch size32Number of network layers6Learning rate decayAdaptively adjusted through the validation setMaximum sequence length256Dropout rate0.2Gradient clipping5.0 Experimental parameter settings. Although the total volume of data processed in this study is substantial, technical strategies such as batch loading, data simplification, and incremental training are employed to effectively reduce system resource consumption. These optimizations enable the entire experiment to be completed on a system with 16 GB of RAM and a GTX 1080 Ti GPU. The model training adopts a mini-batch mechanism with a batch size of 32, and the duration of each training epoch is kept within two hours. The total training time ranges from approximately 36 to 48 h. While the time requirement is moderate, the hardware configuration remains within the capacity of most universities and educational institutions. Performance evaluation To comprehensively assess the performance of the Transformer-based model in generating high school English course resources, several key evaluation metrics are employed:Accuracy. Accuracy. Accuracy. Accuracy measures the consistency between the generated teaching resources and the intended instructional objectives, primarily applied to the classification task of resource difficulty levels31, which is calculated as follows:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Accuracy = \frac{TP + TN}{{TP + FP + FN + TN}}$$\end{document} In Eq. (6), TP (True Positive) represents correctly identified positive examples; TN (True Negative) denotes the number of correctly identified negative examples; FP (False Positive) refers to negative examples incorrectly identified as positive; FN (False Negative) indicates positive examples incorrectly identified as negative. (2)Recall. (2)Recall. Recall. Recall evaluates the proportion of core STEAM elements successfully included in the generated resources, focusing on the coverage of STEAM-related content within instructional materials32, calculated as:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Recall = \frac{TP}{{TP + FN}}$$\end{document} (3)Precision. (3)Precision. Precision. Precision measures the proportion of STEAM-related content in the generated resources that is genuinely relevant to the instructional goals. This metric reflects the relevance and accuracy of the generated content33:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Precision = \frac{TP}{{TP + FP}}$$\end{document} (4)F1 score. (4)F1 score. F1 score. The F1 score is the harmonic mean of precision and recall, providing a balanced evaluation of the model’s accuracy and completeness34, which is computed as:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{F1 - Score}} = 2 \times \frac{{{\text{Precision}} \times {\text{Recall}}}}{{{\text{Precision}} + {\text{Recall}}}}$$\end{document} (5)Bilingual evaluation understudy (BLEU) score. (5)Bilingual evaluation understudy (BLEU) score. Bilingual evaluation understudy (BLEU) score. The BLEU score is a widely used metric for evaluating the quality of machine translation and text generation. It calculates the degree of overlap between the generated text and human-written reference texts by analyzing n-gram matches35. The standard equation is as follows:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{BLEU}} = {\text{BP}} \cdot {\text{exp}}\left( {\mathop \sum \limits_{n = 1}^{N} w_{n} logp_{n} } \right)$$\end{document} In Eq. (10), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{n}$$\end{document} represents the precision of n-gram matches at order n, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{n}$$\end{document} denotes the weight assigned to each n-gram, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{BP}}$$\end{document} refers to the brevity penalty, which penalizes output texts that are excessively short. In educational text generation tasks, a BLEU score between 0.4 and 0.6 generally indicates moderate text quality, while scores above 0.6 suggest strong consistency and fluency relative to human references. In the present study, the proposed model achieves a BLEU score of 0.78 on the test set, significantly outperforming the average scores of comparison models. This result indicates a high degree of similarity to teacher-authored instructional materials in terms of content coherence and linguistic quality. To further evaluate the model’s language diversity and creativity, two complementary metrics are introduced. The first is Self-BLEU, which assesses the similarity between multiple generated texts by treating each as both “candidate” and “reference” in a pairwise comparison. A lower Self-BLEU score indicates greater variation among outputs, reflecting lower redundancy and higher creative diversity. The second metric is Distinct-n, which includes Distinct-1 and Distinct-2, measuring the proportion of unique unigrams and bigrams, respectively, relative to the total number of tokens in the generated texts. This metric is widely used to quantify lexical richness and expressive variability in text generation. The formula is given as Eq. (11):11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{Distinct}} - {\text{n}} = \frac{{{\text{unigram n}} - {\text{grams}}}}{{total {\text{n}} - {\text{gram}}}}$$\end{document} (6)Innovation and uniqueness (6)Innovation and uniqueness Innovation and uniqueness To more effectively evaluate the originality and innovativeness of the generated instructional resources, this study adopts a hybrid metric referred to as the Innovation Index36, which combines expert evaluation with corpus-based comparative analysis. The index is calculated based on two key components: (1) the proportion of generated content that diverges from high-frequency templates in the training set, reflecting the model’s ability to break away from conventional pedagogical expressions; (2) the extent to which cross-disciplinary STEAM-related keywords or semantic features are incorporated into the content, indicating the “interdisciplinary” nature of the resources. In addition, experienced subject-matter experts are invited to score the model outputs using a five-point Likert scale across three dimensions: novelty, creativity, and instructional adaptability. The average scores are then standardized and integrated into the overall Innovation Index, providing a comprehensive assessment of content originality. (7)Baseline models and benchmark description (7)Baseline models and benchmark description Baseline models and benchmark description To validate the performance advantages of the proposed model, two baseline models and two literature-based comparative models are introduced.Baseline model 1: A traditional RNN-based Seq2Seq model using a standard Long and Short-Term Memory (LSTM) architecture without an attention mechanism. This model serves as a reference representing early-stage deep learning frameworks.Baseline model 2: A standard Transformer model that excludes the integration of STEAM-related concepts and semantic guidance. This configuration is used to assess the performance gain attributable to interdisciplinary embedding. Baseline model 1: A traditional RNN-based Seq2Seq model using a standard Long and Short-Term Memory (LSTM) architecture without an attention mechanism. This model serves as a reference representing early-stage deep learning frameworks. Baseline model 1: A traditional RNN-based Seq2Seq model using a standard Long and Short-Term Memory (LSTM) architecture without an attention mechanism. This model serves as a reference representing early-stage deep learning frameworks. Baseline model 2: A standard Transformer model that excludes the integration of STEAM-related concepts and semantic guidance. This configuration is used to assess the performance gain attributable to interdisciplinary embedding. Baseline model 2: A standard Transformer model that excludes the integration of STEAM-related concepts and semantic guidance. This configuration is used to assess the performance gain attributable to interdisciplinary embedding. These comparisons offer a controlled evaluation framework to isolate and measure the contributions of both the model architecture and the STEAM-oriented enhancements. Additionally, Comparison Model A, proposed by Qi et al.37, was included for further comparison. Model A is an English course resource development model that utilizes traditional machine learning techniques, such as Gaussian process regression, to assess teaching quality. This model analyzes instructional behaviors and student feedback data to predict teaching outcomes and offer improvement recommendations. Performance data for Model A was derived from this study’s experimental results. In addition, Comparison Model B, proposed by Zhang et al.38, is considered. Model B is based on predefined rules and templates for generating English course resources. It emphasizes interdisciplinary integration but relies on static, rule-based methods to create content. Performance data for Model B also originated from the experimental results of this study. (8)User satisfaction survey metrics. (8)User satisfaction survey metrics. User satisfaction survey metrics. This study collects user satisfaction data through a structured questionnaire survey. The questionnaire is designed based on a five-point Likert scale and covers multiple dimensions, including satisfaction, practicality, innovativeness, and ease of use. Respondents include both students and teachers who participated in the course, encompassing individuals with varying academic performance and teaching experience to ensure diversity and representativeness of the data. The questionnaire is administered after the course implementation, and both students and teachers complete it anonymously via an online platform to ensure the authenticity and voluntariness of the feedback. Results and discussion Experimental results The evaluation of the comprehensive accuracy index is presented in Fig. 5:Figure. 5Comparison of comprehensive accuracy indicators. Comparison of comprehensive accuracy indicators. The Transformer-based model achieved an overall accuracy score of 0.87, outperforming Comparison Model A (0.82) and Comparison Model B (0.80). In specific categories such as grammatical accuracy, vocabulary relevance, and cultural sensitivity, the Transformer-based model also excelled, achieving scores of 0.90, 0.85, and 0.82, respectively. These results highlight the model’s superior accuracy and practicality in generating English teaching resources that effectively address students’ learning needs. This performance stems from the Transformer model’s advanced deep learning architecture and its self-attention mechanism, which captures long-term dependencies in language. The self-attention mechanism ensures high consistency and fluency in text generation, reducing errors commonly found in traditional approaches. The model demonstrates particular strength in achieving grammatical precision and lexical coherence. Additionally, the training process included multiple iterations of hyperparameter optimization, such as fine-tuning learning rates and batch sizes. These optimizations significantly enhanced the model’s performance across various evaluation metrics, further solidifying its advantages in grammar, vocabulary, and cultural sensitivity. The Transformer model’s adaptability is another key factor contributing to its effectiveness. It processes complex sentence structures and diverse linguistic expressions with remarkable precision. This flexibility ensures that the generated English teaching resources maintain grammatical accuracy, consistent expression, and innovative content. The combination of these strengths underscores the model’s potential for application in diverse educational contexts. Figure 6 illustrates the recall and precision metrics achieved by the Transformer-based model across various STEAM domains. In the science field, for example, the model attains a recall of 0.90 and a precision of 0.88, both significantly surpassing the scores of Comparison Model A and Comparison Model B. These results demonstrate the model’s ability to integrate diverse STEAM knowledge effectively, thereby producing accurate and comprehensive resources that enhance students’ learning experiences.Figure. 6Comparison of recall and precision in various STEAM fields (a. recall; b. precision). Comparison of recall and precision in various STEAM fields (a. recall; b. precision). This superior performance is attributable to the Transformer model’s multi-layered architecture and self-attention mechanism, which facilitate the effective processing of interdisciplinary knowledge. The model generates highly precise and well-rounded teaching materials, particularly in scientific domains. Unlike traditional rule-based or template-driven methods, the deep learning approach employed by the Transformer model enables a better understanding of relationships between disciplines, leading to the creation of more accurate and contextually relevant content. The training process further contributes to this success. The Transformer model was trained on a rich, interdisciplinary dataset, which allowed it to incorporate subject-specific knowledge into its generated content across STEAM fields. This data-driven training enhanced the model’s ability to achieve high recall and precision in all domains. Moreover, the model’s use of multi-task learning enables it to simultaneously process data from multiple disciplines while maintaining efficient performance across tasks. This capability ensures the generation of accurate and comprehensive STEAM resources, addressing the fragmentation of knowledge often associated with traditional methods. The results emphasize the model’s potential to provide integrated and precise educational materials in diverse STEAM contexts. The comprehensive evaluation results for BLEU Score and innovation indicators are presented in Fig. 7. The Innovation Index measures a generation system’s capacity to offer creative solutions beyond traditional methods. It assesses the model’s ability to generate innovative and diverse teaching resources, especially in language generation tasks. The index evaluates not only the accuracy and fluency of the generated content but also its diversity, uniqueness, and level of creativity. The Innovation Index is computed through weighted scoring across multiple dimensions. Each dimension is assigned a predefined weight, contributing to the overall score. For example, technical innovation may contribute 40%, functional innovation 30%, and pedagogical innovation the remaining 30%. This approach ensures a balanced evaluation of the model’s innovative capabilities in creating educational materials, highlighting its potential to improve both teaching methodologies and resource variety.Figure. 7Comprehensive evaluation results of BLEU Score and innovation index. Comprehensive evaluation results of BLEU Score and innovation index. Figure 7 presents a comprehensive comparison of the proposed model with Model A, Model B, and two common baseline models (RNN Seq2Seq and standard Transformer) across multiple language generation quality metrics. In terms of BLEU score, the proposed model achieves 0.78, significantly outperforming Model A (0.72), Model B (0.70), as well as the baseline models, RNN Seq2Seq (0.66) and standard Transformer (0.69). This indicates that the proposed model exhibits higher generation accuracy and language matching. For the Innovation Index, the proposed model scored 0.75, surpassing Model A (0.68), Model B (0.65), and both baseline models (RNN Seq2Seq: 0.60, standard Transformer: 0.63). This metric, which evaluates the model across multiple dimensions such as technical structure, functional implementation, and adaptation to the teaching context, reflects the model’s innovative performance in generating educational content. In terms of content diversity, the proposed model outperformed all other models, scoring the highest in Content Diversity (0.70), Distinct-1 (0.52), and Distinct-2 (0.37). This indicates that its output contains a richer vocabulary and more varied sentence structures. Additionally, the Self-BLEU score of 0.41, the lowest among all models, means that the generated content has the least similarity and redundancy, ensuring maximum diversity in output. Overall, the proposed model excels across the dimensions of generation quality (BLEU), innovation (Innovation Index), and diversity (Content Diversity), outperforming both comparison and baseline models. This validates its comprehensive capabilities and practical utility in generating educational content. These improvements directly contribute to a positive impact on English learning outcomes. Higher generation accuracy ensures that students receive more accurate and authentic language input, which enhances grammar mastery and expression precision. More innovative content sparks students’ cognitive interest and critical thinking, boosting learning motivation. Moreover, the richer and more diverse language materials help avoid content repetition, broadening students’ vocabulary and language structure mastery, and enhancing their language transfer and expressive abilities. Therefore, the model’s superior performance not only demonstrates algorithmic advancement but also effectively supports the personalization, diversification, and efficiency of the English learning process. As shown in Fig. 8, the proposed model demonstrates significant improvements in student learning interest, knowledge mastery, and application ability. Specifically, learning interest increased by 15%, knowledge mastery achieved a score of 0.86, and application ability improved by 20%. These metrics surpass those of Comparison Models A and B, indicating that the proposed model more effectively engages students, improves learning outcomes, and enhances overall literacy. Figure 9 presents teacher feedback and implementation convenience.Figure. 8Evaluation of student participation and learning effectiveness.Figure. 9Teacher feedback. Evaluation of student participation and learning effectiveness. Teacher feedback. The success of the Transformer model can be attributed to its capacity to generate personalized teaching resources tailored to diverse student needs and abilities. By offering flexible learning paths, the model increases student engagement, motivating active participation and improving both knowledge mastery and application skills. The task-driven approach emphasizes solving real-world problems, which enhances application abilities. In contrast to traditional models, the Transformer model generates more challenging and practical tasks, significantly boosting student motivation and practical skills. Additionally, the model creates interdisciplinary teaching resources across various STEAM fields, allowing students to explore a wider array of learning content. This diverse learning experience enriches student outcomes, particularly in applying interdisciplinary knowledge. As shown in Fig. 9, the proposed model received high ratings for its ease of integration into existing courses, support for interdisciplinary teaching, and teaching flexibility. The model’s ease of integration was rated 4.2, support for interdisciplinary teaching 4.5, teaching flexibility 4.0, and overall feedback satisfaction 4.3. These ratings exceed those of Comparison Models A and B, indicating the model is more convenient for implementation, better integrates into existing teaching environments, and is highly appreciated by teachers. This performance can be attributed to the Transformer model’s flexibility, which allows seamless integration into current curricula. High teacher evaluations reflect the model’s ability to adapt easily to the existing teaching context, reducing adaptation costs and enhancing the practical utility of teaching resources. The model is particularly effective in supporting interdisciplinary teaching, integrating knowledge from various disciplines to generate resources that meet the needs of cross-disciplinary education, a key feature in STEAM education. Teachers particularly value this capability. Additionally, the model’s ability to generate diverse resources enables customization to meet specific student needs, enhancing teacher autonomy and convenience. This flexibility contributes to increased teacher satisfaction with the model. This study also collects qualitative feedback from students and teachers through open-ended interviews, followed by thematic analysis. Three core themes emerge from student responses. First, students generally perceive the generated learning content as closely aligned with their interests and proficiency levels, which enhances the relevance and engagement of their learning experience. One student noted, “The examples in these materials feel tailor-made for me. I can connect them to what we have learned in our science and art projects.” Second, many students express appreciation for the seamless integration of interdisciplinary knowledge within the course content, stating that such connections make learning more coherent and practical. As one student remarked, “Reading about climate change or electric circuits in English made me feel that learning across subjects is interconnected and more meaningful.” Third, several students highlight their preference for the autonomy offered by the content selection mechanism, indicating that this flexibility improves their motivation and sense of initiative. One student shared, “I can choose the topics I want to read about, which makes me more eager to keep learning.” Teacher feedback also reveals multiple advantages of the model in instructional practice. Teachers commonly report that the generated content demonstrates high coherence and practicality, significantly reducing their lesson preparation time. One teacher commented, “These texts are more coherent and relevant than I expected. They can be used directly or easily adapted to suit students’ levels.” Another teacher emphasized, “These resources connect English with science and the arts, sparking students’ curiosity and increasing their engagement.” To further validate these metrics, Multivariate Analysis of Variance (MANOVA) is used to assess the significance of differences among the models. This method is ideal for evaluating multiple dependent variables simultaneously. The specific results are summarized in Table 4.Table 4Overview of MANOVA results.Model comparisonMetricProposed model mean ± SDComparison model mean ± SDF-valuep-valueη2 (Effect size)Proposed vs. model AAccuracy (%)93.5 ± 1.292.3 ± 1.58.530.001 **0.12 (medium)Recall (%)91.8 ± 1.690.1 ± 1.87.120.008 **0.10 (medium)Precision (%)92.0 ± 1.390.6 ± 1.76.870.010 **0.09 (medium)F1 score (%)91.9 ± 1.490.4 ± 1.68.210.002 **0.11 (medium)BLEU score0.78 ± 0.050.75 ± 0.065.940.017 *0.08 (small)Originality & Uniqueness0.82 ± 0.040.78 ± 0.054.650.032 *0.07 (small)Proposed vs. model BAccuracy (%)93.5 ± 1.291.0 ± 1.79.86 < 0.001 ***0.15 (large)Recall (%)91.8 ± 1.689.2 ± 1.98.730.004 **0.13 (medium)Precision (%)92.0 ± 1.389.5 ± 1.87.450.006 **0.11 (medium)F1 score (%)91.9 ± 1.489.4 ± 1.79.03 < 0.001 ***0.14 (medium)BLEU score0.78 ± 0.050.73 ± 0.076.590.011 *0.09 (small)Originality & Uniqueness0.82 ± 0.040.77 ± 0.065.320.026 *0.08 (small)* indicates p < 0.05, ** indicates p < 0.01, *** indicates p < 0.001; η2 refers to partial eta squared effect size, which measures the magnitude of differences between groups. Table 4Overview of MANOVA results.Model comparisonMetricProposed model mean ± SDComparison model mean ± SDF-valuep-valueη2 (Effect size)Proposed vs. model AAccuracy (%)93.5 ± 1.292.3 ± 1.58.530.001 **0.12 (medium)Recall (%)91.8 ± 1.690.1 ± 1.87.120.008 **0.10 (medium)Precision (%)92.0 ± 1.390.6 ± 1.76.870.010 **0.09 (medium)F1 score (%)91.9 ± 1.490.4 ± 1.68.210.002 **0.11 (medium)BLEU score0.78 ± 0.050.75 ± 0.065.940.017 *0.08 (small)Originality & Uniqueness0.82 ± 0.040.78 ± 0.054.650.032 *0.07 (small)Proposed vs. model BAccuracy (%)93.5 ± 1.291.0 ± 1.79.86 < 0.001 ***0.15 (large)Recall (%)91.8 ± 1.689.2 ± 1.98.730.004 **0.13 (medium)Precision (%)92.0 ± 1.389.5 ± 1.87.450.006 **0.11 (medium)F1 score (%)91.9 ± 1.489.4 ± 1.79.03 < 0.001 ***0.14 (medium)BLEU score0.78 ± 0.050.73 ± 0.076.590.011 *0.09 (small)Originality & Uniqueness0.82 ± 0.040.77 ± 0.065.320.026 *0.08 (small)* indicates p < 0.05, ** indicates p < 0.01, *** indicates p < 0.001; η2 refers to partial eta squared effect size, which measures the magnitude of differences between groups. Overview of MANOVA results. * indicates p < 0.05, ** indicates p < 0.01, *** indicates p < 0.001; η2 refers to partial eta squared effect size, which measures the magnitude of differences between groups. Based on the multivariate analysis of variance results shown in Table 4, the proposed Transformer-based model significantly outperforms both comparison Model A and Model B across all evaluated metrics. Specifically, compared to Model A, the proposed model exhibits statistically significant advantages in accuracy (F = 8.53, p = 0.001), recall (F = 7.12, p = 0.008), precision (F = 6.87, p = 0.010), F1 score (F = 8.21, p = 0.002), and BLEU score (F = 5.94, p = 0.017), along with notable improvements in originality and uniqueness (F = 4.65, p = 0.032). The effect sizes (η2) range from 0.07 to 0.12, indicating small to medium effects, which demonstrate that these differences are both statistically significant and practically meaningful. Compared to Model B, the proposed model’s superiority is even more pronounced across all metrics, particularly in accuracy (F = 9.86, p < 0.001, η2 = 0.15) and F1 score (F = 9.03, p < 0.001, η2 = 0.14), showing large effect sizes and further validating the model’s robustness and multidimensional performance advantages. Improvements in originality and uniqueness also reach statistical significance (F = 5.32, p = 0.026), indicating that the model clearly outperforms the rule-based traditional approach in terms of content originality and diversity. Overall, although the mean differences in some metrics are relatively modest, statistical analyses confirm their significance and meaningful effect sizes. These findings suggest that the proposed deep learning model not only enhances the quality and innovativeness of teaching resources but also possesses strong practical value and application potential, making it suitable for intelligent generation of interdisciplinary English course resources. Fuzzy classification and corpus processing: illustrative analysis To provide a clearer understanding of the model’s operational process and effectiveness in real-world applications, this study supplements the quantitative evaluation with comparative examples of fuzzy classification, annotated corpus samples, and disambiguation handling. These qualitative analyses address the limitations of purely statistical metrics and illustrate how the model responds to complex interdisciplinary instructional content. The fuzzy logic classification framework is designed to more accurately reflect the multi-disciplinary and multi-attribute nature of instructional resources. In conventional single-label classification, it is often difficult to fully capture the intersecting characteristics of content, resulting in overly simplistic and information-deficient outcomes. In contrast, the fuzzy logic system enables the assignment of multiple degrees of membership to a single resource. For example, a project that integrates both physical principles and mathematical computation is assigned corresponding subject affiliation scores, reflecting its interdisciplinary nature. Table 5 compares traditional classification labels with fuzzy logic-based membership scores for typical instructional content, demonstrating how this method enhances the precision and granularity of content categorization. In tasks involving the intersection of arts and engineering, the model also accurately identifies and assigns relevant tags for artistic design, thereby increasing the diversity and creativity of the generated content.Table 5Comparison of instructional resource labels before and after fuzzy classification.Resource titleTraditional LabelFuzzy logic membership scores (Example)Smart transportation design projectEngineeringEngineering (0.8), Technology (0.7), Mathematics (0.5)Sustainable architecture proposalEngineeringEngineering (0.9), Arts (0.4), Science (0.6)Robotics programming instructionTechnologyTechnology (0.85), Mathematics (0.7), Science (0.65)Environmental protection reportScienceScience (0.9), Language Arts (0.5), Technology (0.3)Arts and literature integration projectArtsArts (0.95), Literature (0.8), History (0.4) Table 5Comparison of instructional resource labels before and after fuzzy classification.Resource titleTraditional LabelFuzzy logic membership scores (Example)Smart transportation design projectEngineeringEngineering (0.8), Technology (0.7), Mathematics (0.5)Sustainable architecture proposalEngineeringEngineering (0.9), Arts (0.4), Science (0.6)Robotics programming instructionTechnologyTechnology (0.85), Mathematics (0.7), Science (0.65)Environmental protection reportScienceScience (0.9), Language Arts (0.5), Technology (0.3)Arts and literature integration projectArtsArts (0.95), Literature (0.8), History (0.4) Comparison of instructional resource labels before and after fuzzy classification. Furthermore, the construction of the corpus is based on standardized POS tagging and syntactic parsing to ensure consistency and quality in data preprocessing. For instance, in the sentence “Students design smart cars,” each token is accurately labeled with its POS and syntactic role, providing structured linguistic input for deep semantic modeling. This multi-level annotation allows the model to extract features more effectively, enhancing the fluency and logical coherence of the generated teaching content. Table 6 presents sample annotations to demonstrate the structure and their contribution to model training.Table 6Sample corpus annotations.SentenceTokenPOS tagSyntactic roleStudents design smart carsStudentsNoun (NNS)SubjectdesignVerb (VB)PredicatesmartAdjective (JJ)ModifiercarsNoun (NNS)Object Table 6Sample corpus annotations.SentenceTokenPOS tagSyntactic roleStudents design smart carsStudentsNoun (NNS)SubjectdesignVerb (VB)PredicatesmartAdjective (JJ)ModifiercarsNoun (NNS)Object Sample corpus annotations. In addition, the model incorporates a semantic disambiguation mechanism that leverages both contextual inference and fuzzy logic rules to accurately interpret polysemous words. For example, the term “light” can refer to “illumination,” “lightweight,” or “pale color,” depending on the context. The model effectively distinguishes among these meanings based on surrounding linguistic cues, thereby preventing misinterpretation in generated content. This capability enhances the accuracy and relevance of teaching materials, significantly improving their practical instructional value. Table 7 provides representative examples of disambiguation processing.Table 7Semantic disambiguation examples for the word “light”.Sentence textWordContextual cluesModel interpretationThe light in the room is brightlightAccompanied by “room,” “bright”Light (illumination)She carried a light bag todaylightAccompanied by “carried,” “bag”Light (weight)We need to study the light spectrumlightAccompanied by “study,” “spectrum”Light (physics context)The room was decorated with light colorslightAccompanied by “colors,” “decorated”Light (color tone) Table 7Semantic disambiguation examples for the word “light”.Sentence textWordContextual cluesModel interpretationThe light in the room is brightlightAccompanied by “room,” “bright”Light (illumination)She carried a light bag todaylightAccompanied by “carried,” “bag”Light (weight)We need to study the light spectrumlightAccompanied by “study,” “spectrum”Light (physics context)The room was decorated with light colorslightAccompanied by “colors,” “decorated”Light (color tone) Semantic disambiguation examples for the word “light”. These illustrative examples collectively demonstrate the model’s capacity to handle interdisciplinary, ambiguous, and richly contextual language data, highlighting its practicality and robustness in real-world educational scenarios. Discussion The experimental results of this study confirm the comprehensive advantages of the proposed Transformer-based model in generating high school English STEAM teaching resources. The achievement of these results is mainly attributed to the deep integration of model architecture and teaching philosophy. The self-attention mechanism of the Transformer model enables it to effectively capture and correlate interdisciplinary semantic information, thereby generating more coherent and accurate text in both language and content. This coincides with Samala et al.39 assertion that emerging AI and deep learning technologies are becoming key forces driving educational innovation and equity. The proposed model is a concrete practice under this trend, which transforms STEAM educational concepts into dynamically optimized personalized resources through intelligent content generation technology. The significant improvement in student engagement and learning outcomes observed in this study can be explained from the perspective of integrating teaching methods and technology. Yu et al.40 found that tools such as mobile learning technology could enhance learning engagement and outcomes by improving interactivity. On this basis, this study further explores the personalized and interdisciplinary nature of the resources generated by the model, which stimulates students’ cognitive interest, and deepens Yu et al.’s advocacy of “enhancing student participation through technological tools” by placing English learning in meaningful real-life problem contexts. Moreover, the high satisfaction level of teacher feedback, especially the recognition of the model’s ease of integration and support for interdisciplinary teaching, strongly responds to Nicolaou et al.41 viewpoint that educational technology innovation must be seamlessly integrated into existing courses and easy for teachers to use in order to maximize its effectiveness. As a content generation tool, this model has successfully lowered the threshold for teachers to implement high-quality STEAM English teaching. In terms of interdisciplinary integration, the successful practice of this model provides new insights beyond traditional STEM integration paths. Maass et al.42 delved into the core role of mathematics in STEM education, with a focus on the connections within science and engineering disciplines. While this study expands the interdisciplinary scope to include the complete STEAM spectrum of arts and humanities, and successfully combines it with the non-traditional STEM field of language education. This indicates that semantic modeling based on deep learning can build semantic bridges between seemingly discrete disciplines, providing technical possibilities for achieving deep integration in a wider range of disciplinary fields. However, this study still has several limitations. Firstly, the teaching experiment period is relatively short, mainly evaluating the short-term teaching effectiveness of the model. Its long-term impact on students’ interdisciplinary thinking and language abilities still needs to be verified through longer longitudinal studies. Secondly, the training and inference process of the model still requires certain computing resources, which may become an obstacle to promotion in educational environments with limited resources. In addition, current research mainly focuses on the quality of teaching resource generation and initial teaching effectiveness. There has been no in-depth exploration of the specific interactive modes of generated content in the classroom and how teachers can conduct secondary development based on generated content in deep teaching methods. Based on the above limitations, future research can be conducted in the following directions: 1. conduct long-term studies lasting one semester or academic year to evaluate the sustained impact of the model; 2. commit to developing a lightweight version of the model to enhance its applicability in low resource environments; 3. explore in depth the teaching mode of human–machine collaboration and study the best way for teachers to collaborate with AI generated content. Conclusion This study successfully combines the STEAM educational philosophy with Transformer-based deep learning technology, providing an innovative solution for the intelligent development of high school English curriculum resources. The empirical research results show that this scheme can generate high-quality, innovative, and diverse teaching materials at the technical level, and effectively enhance students’ learning interest, knowledge mastery, and application ability at the teaching practice level, while gaining high recognition from teachers. The main contribution of this study lies in surpassing the limitations of traditional resource development models and achieving the organic integration of interdisciplinary content and language learning through a data-driven and semantically profound approach, providing a valuable example for empowering educational innovation with AI. Although further exploration is needed in terms of long-term effectiveness verification and universality, this study undoubtedly provides strong empirical evidence and forward-looking guidance for promoting the development of intelligent education technology and building a more inclusive and adaptive future education ecosystem.