Using Large Language Models to teach legal english vocabulary

Full Text

Problem statement. The 21st century, marked by the intensive development of new technologies and large-scale digitalization processes affecting all spheres of social practice [1; 2], poses new challenges to contemporary education and determines the emergence of innovative values and demands of the postmodern global information society. A practice-oriented approach to educational goals and learning outcomes [3] emphasizes information as a key strategic resource. Digital society has no territorial or institutional boundaries for information flows, which transforms communication and education. Globalization intensifies intercultural interaction and creates new trends in communication and education. The global response is the development of international frameworks for the digital transition of education systems [4]. International and transcontinental information exchange beyond the boundaries of a particular language community requires a high level of foreign language proficiency, enabling effective participation in global academic, professional, and digital communication. At present, the English language holds the status of the predominant language of international business, academia, law, media, and digital technologies. As a global language [5], English is widely used across diverse economic, cultural, digital, and media landscapes, with approximately 350 million native speakers and more than 1.5 billion learners worldwide [6]. This explains the growing importance of ESP for professional communication. Information technologies are now part of daily life [7], and AI-based tools are increasingly used in education. The implementation of digital technologies creates a new educational space that changes traditional teaching and learning. This includes the fusion of real and virtual contexts, which transforms cognitive and social aspects of learning [8]. Since 2018, Russia has entered a stage of comprehensive digital transformation of education[11]. This stage involves the renewal of the entire educational system and the systematic integration of digital technologies into all its structural components, including educational content, teaching methods, organizational forms, and instructional tools. Recent research demonstrates that contemporary educators increasingly move beyond traditional “chalk and talk” approaches and actively introduce new media technologies and AI-based tools into teaching and learning practices [9]. These tools promote more flexible, interactive, and personalized learning environments. In addition, effective AI-based strategies contribute to positive human-AI interaction, supporting learners’ engagement, autonomy, and critical thinking. Consequently, generative AI tools, particularly LLMs, have become integral to the modern digital educational space. They hold significant potential for creating personalized and interactive learning experiences, particularly in domains requiring mastery of specialized terminology, such as legal English within ESP courses. However, the rapid adoption of these tools often outpaces the development of pedagogically and methodologically grounded frameworks for their effective and critical integration into professionally oriented language instruction. This gap forms the core problem addressed by the present study. Literature Review. Recent systematic reviews indicate a rapid expansion of empirical research on generative artificial intelligence in language learning and teaching, with a particular focus on AI-assisted instructional strategies, learner autonomy, writing support, and vocabulary development. An analysis of peerreviewed studies indexed in Scopus and Web of Science demonstrates that generative AI tools, including large language models, are increasingly examined not as experimental novelties but as emerging components of pedagogically grounded language education [10]. Media technologies have already become an essential part of everyday life, which indicates that modern societies, social practices, and human experience are becoming increasingly digital [11]. New media technologies [12] demonstrated their high educational potential during the COVID-19 pandemic, when online education became the only possible mode of instruction and digital environments ensured the continuity of teaching and learning processes. Even those who had previously been indifferent to information technologies became active users of digital tools - a shift encompassing both learners and educators. Digital technologies serve multiple didactic functions: they visualize content, individualize learning, support self-assessment, help process information, create realistic situations (e.g., via virtual reality), maintain motivation, and develop learner autonomy and creativity [13]. A previous study demonstrated that virtual sticker boards helped law students model legal processes and develop intercultural collaboration skills [14]. The present study moves from such collaborative visual environments towards AIassisted content generation. Virtual whiteboards support collaborative knowledge structuring, whereas large language models generate personalized learning materials. In this context, the teacher’s role is to verify and curate the content. This shift corresponds to Russia’s educational policy, which transitions from digitizing existing practices to integrating AI-augmented teaching methods, as stated in the National Development Goals until 2030[12]. The same institutional setting (Vyatka State University law students) and focus on legal competence are retained, ensuring continuity in this line of research on digital ESP instruction. Digital technologies support individualized learning through various methods and tools. They also provide access to global information resources, enable quality monitoring, and facilitate remote interaction between learners and teachers. At present, digital technologies have become an integral component of foreign language education [15], providing an interactive teaching and learning environment both in both classroom and online settings. Innovative digital technologies include multimedia technologies, Internetbased tools, databases in various fields, interactive computer technologies, distance learning tools, as well as online platforms based on neural networks and artificial intelligence [16]. These AI-based platforms can efficiently process and analyze large amounts of information, thus helping learners overcome the technological barrier speech. The broad potential of artificial intelligence in foreign language education is realized through the application of neural networks, which are computer programs simulating human brain activity and capable of automatic learning and selfimprovement based on large datasets [17]. Speech recognition technologies enable neural networks to process, understand, and translate speech into different languages, which significantly expands their educational applicability. Within a practice-oriented approach, neural networks can help create learnercentered materials adapted to students’ proficiency and age, introduce intelligent tutoring systems, and support educational innovation [18]. AI-based tools should be introduced based on institutional curricula and learners’ needs. Despite the significant advantages of neural networks, they cannot replace the teacher as a human mentor [16]. Unlike AI applications, a human educator can consider individual learner characteristics and adjust instruction accordingly. The didactic potential of generative language models for ESP is well recognized. However, specific methodological frameworks for their controlled integration - particularly for teaching specialized vocabulary - are still lacking. This study proposes and tests a three-stage model for using an LLM to master legal terminology in ESP. Methodology. This study uses a LLM - a generative neural network designed to produce coherent text in response to natural language input [16]. Recent academic discussions emphasize that generative AI tools, including large language models, are no longer seen as experimental novelties but as emerging components of everyday educational practice. A recent systematic review by Y. Qian [18] outlines pedagogical trends and methodological frameworks for integrating these tools into teaching and learning, including their use for fostering creativity, critical thinking, learner autonomy and structured teacher - AI collaboration. Contemporary research also highlights the importance of responsible, pedagogically grounded use of generative AI, addressing critical engagement, content reliability and the balance between technological support and the teacher’s guiding role. This study uses a DDR approach. We analyzed the lexical domain “Branches of Law” and related legal terminology, working with an LLM. The methods are theoretical analysis, pedagogical modelling and qualitative analysis of AI-generated content. The theoretical part reviews literature on ESP pedagogy, legal linguistics, vocabulary acquisition (Hatch and Brown’s model) and digitalization of education. Pedagogical modelling produced a three-stage instructional design (initial, working, control). For qualitative analysis, we submitted structured prompts to an LLM and examined the generated word lists, exercises, and assessment tasks. This helped us identify the model’s didactic potential, limitations, and output patterns. The study proceeded in four steps: problem formulation, model development, content generation, and output analysis. The working hypothesis is that the methodologically structured use of an LLM can support individualized learning and autonomous acquisition of legal vocabulary. Vocabulary is central to mastering English for Legal Purposes [19]. The proposed model builds on Hatch and Brown’s vocabulary acquisition theory, which describes stages from encountering a word to using it in communication. This framework shaped the instructional design [20]. The study also takes into account the limitations of large language models, especially the risk of factual errors and the need for critical evaluation of outputs [21]. Therefore, the model includes the teacher’s guiding and curating role to ensure pedagogical control and foster students’ critical reflection. Pilot study. A quasi-experimental study was conducted at Vyatka State University in spring 2025 to test the proposed model. The sample consisted of 24 third-year law students (specialisation 40.03.01 Jurisprudence) enrolled in the ESP course “Legal English”. They were assigned to an experimental group (n = 12) and a control group (n = 12). The two groups had equivalent prior English proficiency (B1 CEFR, as measured by the university placement test) and comparable baseline legal vocabulary knowledge (pre-test scores did not differ significantly between groups, p > 0.05). The intervention lasted four weeks (16 contact hours). The experimental group followed the three-stage model (initial, working, control) using an LLM accessible via a free web interface, while the control group used traditional textbook-based instruction for the same lexical unit (“Branches of Law”). Assessment included a pre-test and a post-test with 30 items covering recognition, form recall and productive use of legal terminology, as well as semi-structured interviews with six randomly selected students from the experimental group. Statistical analysis was performed using SPSS 26.0. Due to the small sample size and non-normal data distribution, the non-parametric Mann-Whitney U test was used to compare posttest results between the two groups. Ethical considerations. The study was embedded in the regular “Legal English” course at Vyatka State University. All participants were informed of the research purpose and gave voluntary consent. They were assured that their results would remain anonymous in publications and that they could withdraw from the study at any stage without academic consequences. Personal data were processed in accordance with Federal Law No. 152-FZ “On Personal Data”. Results and discussion. Teaching legal lexis with a large language model: three stages. The proposed approach follows the classical triad of lexical skill formation: introduction, training and consolidation [22]. This aligns with Hatch and Brown’s theory, which describes stages from encountering a word to its active use. These two frameworks are combined into a three-stage instructional model. An LLM supports learning at each stage. In the Initial Stage, students learn the form and meaning of new words. They formulate precise queries to the LLM to generate thematic lexical databases (e.g., lists of terms with definitions). The Working Stage aims to build strong memory associations through practice. Here, the LLM produces varied exercises (multiple-choice, fill-in-the-blank, and error correction exercises) for vocabulary automatization. In the Control Stage, the focus shifts to active use. The LLM helps create tests that assess understanding and move words into students’ active vocabulary. This model was applied and illustrated using legal English units (“Branches of Law”, “Contract Law”, “Intellectual Property Law”). The following description shows its use for the first unit, “Branches of Law”, as a representative example. Stage 1: Generating a Lexical Database. Students formulate a query to obtain a structured list of legal terms with definitions. For example, the query “Branches of law” prompts the LLM to produce a list of key branches and concise definitions (e.g., Criminal Law: governs crimes and punishments; Civil Law: deals with disputes between individuals or organizations). Table 1 summarizes the full list of 20 branches generated for this study. Table 1. LLM-generated fragment of the lexical database for the topic “Branches of Law” Serial number Branch of Law Core Definition 1 Criminal Law Governs crimes and punishments, addressing actions harmful to society 2 Civil Law Deals with disputes between individuals or organizations (contracts, property) 3 Constitutional Law Focuses on the interpretation and application of the constitution ... ... ... 10 Intellectual Property Law Governs the protection of creations of the mind (inventions, trademarks) Source: compiled by Ekaterina G. Nikulina, Olga S. Rubleva, Natalia A. Usova. Stage 2: Generating Practice Exercises. At this stage, students or the teacher use an LLM to create various exercises for targeted vocabulary practice. The LLM produces different task types, which helps with differentiation and individualization (Table 2). Table 2. LLM-generated exercise types for the “Branches of Law” vocabulary Exercise type and prompt Example task Pedagogical focus Multiple Choice “Create a multiple-choice exercise…” Which branch of law deals with disputes between individuals? a) Criminal b) Civil c) Constitutional Recognizing and matching concepts with their definitions Fill-in-the-blank “An exercise Insert an appropriate word…” _ law governs crimes and determines punishments. (Word bank: Criminal, Civil, Administrative) Recalling and applying terms in a minimal context Error correction “An exercise correct the mistakes…” Intellectual property law regulates marriage and divorce. (Error: This is Family law) Deep conceptual understanding and differentiation between similar terms Note. The LLM also provided a complete answer key for each exercise type. Source: compiled by Ekaterina G. Nikulina, Olga S. Rubleva, Natalia A. Usova. Stage 3: Generating Assessment Materials. In the final stage, students create a test to check vocabulary acquisition and readiness to use. They prompt the LLM for “a test for the same words”. The LLM returns a multi-format test with three parts: multiple-choice questions (e.g., “Which branch regulates crimes?”), true/false statements (e.g., “Family law covers contracts” - false), and short-answer questions (e.g., “Define civil law and provide an example”). Table 3 summarizes the test structure. Table 3. LLM-generated test structure for the topic “Branches of Law” Part Task type Example item Testing focus 1 Multiple Choice (5 items) Which branch of law governs the formation of businesses? a) Tax b) Corporate c) Criminal Recognition and selection of The correct term 2 True/False (5 items) Administrative law ensures government agencies follow regulations. - True Understanding of the core function of a legal branch 3 Short Answer (5 items) Explain the difference between Felonies and Misdemeanors Ability to formulate definitions and explain concepts independently Source: compiled by Ekaterina G. Nikulina, Olga S. Rubleva, Natalia A. Usova. A complete answer key was generated by the LLM for all test parts. Choice of lexical unit. The unit “Branches of Law” was chosen for the pilot study for three reasons. It is central to legal English curricula for non-linguistic majors at Russian universities and serves as a foundation for later topics (contract law, criminal procedure, etc.). The unit contains conceptually close terms (e.g., administrative vs. constitutional law, civil vs. commercial law) that require precise differentiation, which makes it possible to test how well a large language model generates nuanced definitions. Large language models are also prone to factual errors - for example, confusing intellectual property law with commercial law. This makes the unit suitable for examining how the teacher verifies domain-specific accuracy. According to Hatch and Brown, vocabulary acquisition works best when lexical items are both frequent and easy to distinguish in a professional domain. Testing the model: A pilot study. To test the proposed three-stage model, a pilot study was conducted at Vyatka State University in spring 2025. The participants were 24 third-year law students (specialization 40.03.01 Jurisprudence) enrolled in the ESP course “Legal English”. They were divided into two groups using a quasiexperimental design. The experimental group (n = 12) studied the lexical unit “Branches of Law” using the three-stage model with an LLM under teacher supervision. The control group (n = 12) followed traditional instruction (textbook exercises, teacher-prepared materials, frontal teaching) for the same lexical unit. Both groups had similar prior English proficiency (B1 CEFR, based on the university placement test) and comparable baseline legal vocabulary knowledge (pre-test scores did not differ significantly, p > 0.05). Procedure. The intervention lasted four weeks (16 contact hours). In the experimental group, students worked individually with an LLM following the three-stage sequence. In Stage 1, they formulated queries to generate a lexical database of 20 legal terms with definitions (Table 1). In Stage 2, they created and completed personalized exercise sets (multiple-choice, fill-in-the-blank, and error correction exercises) using prompts from Table 2. In Stage 3, they generated and completed a self-assessment test (Table 3), after which the teacher verified answers and discussed inaccuracies. The teacher also provided prompt templates in Stage 1, checked the factual accuracy of definitions, facilitated group discussion of model errors, and gave corrective feedback in Stage 3. Assessment instruments. Students took a pre-test and a post-test with 30 vocabulary items covering recognition (matching terms with definitions), form recall (fill-in-the-blank), and productive use (short definitions in their own words). The maximum score was 30 points. Semi-structured interviews were conducted with six randomly selected students from the experimental group to explore perceived usefulness, challenges, and skills related to formulating queries. Statistical analysis. The data were processed using SPSS 26.0. Due to the small sample size, the non-parametric Mann-Whitney U test was applied to compare the post-test results between the two groups. Post-test results showed statistically significant differences between the groups (Table 4). Table 4. Comparison of vocabulary acquisition outcomes based on post-test scores Group n Mean score Standard deviation Mann-Whitney U test value p Experimental, EG 12 24.3 2.1 38.5 0.012 Control, CG 12 20.7 3.4 - - Source: created by Ekaterina G. Nikulina, Olga S. Rubleva, Natalia A. Usova. Descriptive analysis indicated that 81% of students in the experimental group scored above the threshold, compared to 64% in the control group. The Mann-Whitney U test confirmed a significant difference between the two groups (U = 38.5, p = 0.012). A similar advantage was observed in the recognition subtask (EG mean = 9.1, CG mean = 8.3, p = 0.041). Qualitative analysis of the interviews identified three main findings. Students developed greater awareness of factual accuracy, for example by double-checking model definitions against their textbook. They also improved their ability to formulate precise queries; one student noted progress from typing “law branches” to specifying “list 20 branches of law with concise definitions for law students”. Finally, students recognized that model outputs needed teacher verification, especially for nuanced legal concepts - for instance, when the model confused administrative and constitutional law. These qualitative findings align with the observations of G. Hambardzumyan, who noted that while AI tools are highly effective for providing personalized, scalable practice opportunities, human instructors remain essential for delivering deep contextual knowledge and emotional support [23]. Limitations. The pilot study has several limitations: a small sample, a short duration, and a focus on a single lexical unit. These should be addressed in future large-scale validation. The implementation of the three-stage model provides several practical observations. First, the quality of the LLM output depends on how precisely the user formulates the query. Learning to write effective prompts is a useful digital skill for students. Second, the teacher remains essential. Although the LLM generates structured content efficiently, the teacher must select and verify materials (especially in law), design the learning sequence, and help students develop critical awareness of model-generated information. Third, the LLM allows rapid creation of personalized exercises and tests, supporting learner autonomy. Fourth, the need to check the model’s output for errors or simplifications is not a weakness but a built-in feature that fosters critical thinking. This turns a technical limitation into a pedagogical opportunity. The LLM thus functions not as an autonomous instructor but as an assistant for content generation and routine tasks. This frees the teacher to focus on creative and interactive work with students. Success depends on embedding the LLM in a teacher-guided framework. At the initial stage, the LLM provides thematic word lists with definitions. At the working stage, it generates varied exercises (multiple-choice, fill-in-the-blank, and error correction exercises). At the control stage, it produces multi-format tests with answer keys. The model is modular and can be transferred to other ESP areas or general English instruction. In summary, the LLM can help both students and teachers by automating basic material generation and practice tasks, which saves time and supports individualised learning. However, this assistance is most effective when the teacher remains the central curator of content and critical engagement, a point also emphasized by L. Kohnke, B.L. Moorhouse, and D. Zou [24]. Conclusion. This study addressed the methodological gap in the pedagogically grounded use of generative artificial intelligence in ESP instruction. Focusing on specialized vocabulary as a core component of professional competence, the research developed and tested a three-stage instructional model for integrating an LLM into teaching, using Legal English as a case study. The model demonstrates how an LLM can function as an assistant in mastering legal lexis. This aligns with international priorities for using digital technologies in education[13]. The proposed three-stage structure - introduction, training, and consolidation - offers a viable way to move from experimental to methodologically sound AI application in language education. The LLM serves three functions: generating thematic lexical databases (initial stage), creating varied practice exercises (working stage), and producing multi-format tests (control stage). The main theoretical contribution is the alignment of generative AI capabilities with established pedagogical frameworks, namely the lexical skill formation triad and Hatch and Brown’s vocabulary acquisition theory, within the paradigm of digitalization of education. Several conditions are necessary for the model to work. The teacher retains an irreducible role as curator of AI-generated content, designer of the learning path, and facilitator of critical thinking. Students need to develop critical digital literacy, so that verifying AI outputs becomes a planned learning outcome rather than a limitation. Pedagogical design must take precedence over technological functionality, ensuring that LLM use serves specific didactic goals in ESP. The study argues for a balanced, complementary approach. The LLM acts not as an autonomous tutor but as a digital assistant that automates routine content generation. This frees the instructor for higher-order teaching activities - direct interaction, mentoring, complex feedback - while fostering learner autonomy through personalized tasks. Although the model was developed and illustrated with legal terminology, its modular, stage-based structure makes it transferable to other ESP domains (business, medicine, engineering). The main limitation is the methodological, illustrative nature of the work; further experimental validation is needed to measure effects on vocabulary acquisition and long-term retention. Nevertheless, the pilot study with 24 law students provided initial empirical evidence of effectiveness, with a 17% advantage in post-test scores for the LLM-assisted group (81% vs 64% success rate). This supports the model’s practical applicability while confirming the need for larger-scale validation. The research offers a concrete methodological framework for the digitalization of ESP instruction. Effective integration of generative AI depends not on prohibition or uncritical adoption, but on thoughtful embedding into a human-centric pedagogical system where expert mentorship and critical engagement remain paramount.

About the authors

Ekaterina G. Nikulina

Vyatka State University

Author for correspondence.
Email: eknikulina986@gmail.com
ORCID iD: 0000-0002-8151-6863
SPIN-code: 3825-1924

Candidate of Philological Sciences, Associate Professor at the Department of Foreign Languages for Non-Linguistic Specialties

36 Moskovskaya St, Kirov, 610000, Russian Federation

Olga S. Rubleva

Vyatka State University

Email: olgarue@mail.ru
ORCID iD: 0000-0001-7346-025X
SPIN-code: 3366-6380

Candidate of Philological Sciences, Associate Professor at the Department of Foreign Languages for Non-Linguistic Specialties

36 Moskovskaya St, Kirov, 610000, Russian Federation

Natalia A. Usova

Russian Presidential Academy of National Economy and Public Administration

Email: usova-na@rudn.ru
ORCID iD: 0000-0002-1728-7736
SPIN-code: 8658-2032

Candidate of Pedagogical Sciences, Associate Professor, Associate Professor at the Institute of Management

82 Vernadsky Ave, bldg 1, Moscow, 119571, Russian Federation

References

Supplementary files