New Ways to Teach and Learn Pragmatics for Advancing Academic Communication Competence: An AI-Mediated Multimodal Platform

2.1. Pragmatic Competence in Academic Communication

Pragmatic competence, understood as the ability to interpret and realize meanings appropriately in context, underpins effective academic communication in multilingual and intercultural higher education, where students must align propositional content with sociocultural expectations, speech-act appropriateness, and interactional norms (Bachman & Palmer, 1996; Sánchez-Hernández & Alcón-Soler, 2019; Taguchi et al., 2016). Across seminars, supervision meetings, email, learning-management forums, and video-conferencing, pragmatic competence entails sensitivity to audience, register, and institutional roles, integrating both pragmalinguistic resources (forms and formulae) and sociopragmatic judgements (what is appropriate for whom, when, and how) (Jakupčević & Portolan, 2021; Mao, 2021; T. T. M. Nguyen, 2011).

Core dimensions include: speech acts (e.g., requests, refusals, critiques, and apologies) and their indirect realizations (Austin, 1962; Searle, 1969); (im)politeness strategies that manage face and rapport in asymmetrical academic relations (Brown & Levinson, 1987; Spencer-Oatey & Jiang, 2003); and conversational principles such as Grice’s maxims and implicature, which organize relevance, precision, and manner in disciplinary talk and writing (Grice, 1975; Lumsden, 2008; Veenstra et al., 2017). Academic discourse also relies on stance-taking and hedging to calibrate epistemic commitment and interpersonal alignment (Hyland, 2005), while multimodal cues, such as prosody, timing, gaze, and gesture, shape how intentions are perceived in presentations and online meetings (Beltrán-Palanques & Querol-Julián, 2018). Conversation-analytic work further shows how turn-taking, repair, and sequencing practices index pragmatic control in classroom and supervision interactions (Youn, 2014).

Within broader models of communicative competence, pragmatics is positioned alongside linguistic, discourse, strategic, and sociocultural components, underscoring that grammatical knowledge alone is insufficient for academic success (Bachman & Palmer, 1996; Canale & Swain, 1980; Celce-Murcia et al., 1995). Intercultural frameworks likewise emphasize the interface between pragmatic competence and intercultural communicative competence, where attitudes, knowledge of social practices, and skills of discovery/interaction inform appropriate language use across cultures and academic registers across languages (Byram, 1997; Chun, 2011; Wang & Li, 2025).

Persistent challenges in multilingual higher education include sociopragmatic transfer (projecting first-language [L1] norms onto second-language [L2]), over- or under-use of politeness markers and hedges, and difficulties managing turn-taking and floor in synchronous/asynchronous environments, often compounded by power distance and institutional hierarchies (Curle et al., 2020; Hu & Chen, 2021; Kamaşak & Şahan, 2023; Y. Lee, 2023). These issues surface in email etiquette (over-formality or undue directness), seminar discussions (late entries, overlap avoidance), and feedback encounters (mitigation of critique), and they can depress participation and perceived professionalism if left unaddressed (Jackson, 2011; Taguchi, 2015).

A growing body of research links pragmatic competence to academically salient outcomes: clearer articulation of ideas, greater willingness to communicate, more credible academic personae, and higher-quality collaboration in intercultural/telecollaborative projects (Aski et al., 2023; Gruber et al., 2023; Lv et al., 2021; Zappa-Hollman & Fox, 2021). Experiential contexts (e.g., study abroad and virtual exchange) are associated with gains in pragmatic routines and sociocultural adaptation, especially when learners actively engage with local practices (Sánchez-Hernández & Alcón-Soler, 2019; Taguchi et al., 2016). Yet without explicit pedagogical support, many learners fail to notice or internalize norms that are diffuse, discipline-specific, and often tacit (Graus & Coppen, 2017; Jenkins et al., 2011; Limberg, 2016).

Consequently, the literature converges on several pedagogical needs. First, explicit instruction and metapragmatic explanation cultivate pragmatic awareness, learners’ conscious attention to the relationships among linguistic forms, contextual variables, and social effects (Taguchi, 2015), and transferable across tasks (Saito & Wu, 2014; Sánchez-Hernández & Barón, 2021). Second, contextualized practice, through scenarios that mirror email to faculty, seminar critique, or supervisor consultations, supports appropriate mapping of form to function in situated interaction (Alsuhaibani, 2020; E. Y. A. Kim & Brown, 2014). Third, iterative feedback cycles (teacher, peer, and automated) enable noticing–reformulation–practice loops that refine strategy use over time; studies with computer-simulated dialogues show improvements in pragmatic appropriateness when feedback is timely and contingent (González-Lloret, 2022; Li et al., 2022; Munshi et al., 2022; Sydorenko et al., 2017). Finally, robust assessment practices, spanning discourse completion tasks (DCTs)/multiple-choice discourse completion tasks (MDCTs), role-plays, performance rubrics, and interaction log, are needed to capture both pragmalinguistic accuracy and sociopragmatic fit with acceptable reliability and validity (Bardovi-Harlig, 2013; Hudson et al., 1995; Ishihara & Cohen, 2010; Roever, 2005).

Taken together, these perspectives motivate instructional designs that couple explicit, context-rich teaching with multimodal, feedback-rich opportunities for practice and reflection. The present study responds to this agenda by evaluating a unified platform that combines automated text analysis, video analysis with multilingual transcription and theory-based annotation, conversation simulations with real-time feedback, AI-supported consultation, and a digital politeness forum: features intended to operationalize the above principles within authentic higher-education workflows.

2.2. Technology-Enhanced Pragmatics Instruction

A substantial body of work indicates that technology-mediated environments, spanning computer-mediated communication (CMC), virtual/immersive simulations, discourse-annotation tools, and automated feedback systems, can support L2 pragmatic development when tasks are explicitly aligned with form–function relations and authentic communicative goals (González-Lloret, 2022; Taguchi, 2015). In these designs, learners rehearse context-sensitive language use, notice mappings between linguistic choices and interactional effects, and attend to multimodal cues characteristic of academic communication.

Simulations and collaborative tasks have proved especially productive. Technology-supported interaction increases opportunities for “pragmatic episodes” (i.e., observable instances of pragmatic language use during interaction) and elicits strategic behaviour as task complexity rises (Y. Kim & Taguchi, 2016). In CMC, interactional resources such as emojis and orthographic conventions operate as stance and mitigation cues that enrich learners’ repertoires (Maa & Taguchi, 2022). Discourse-annotation and noticing tools further render tacit norms visible, linking speech acts, (im)politeness strategies, and contextual variables to learners’ emergent practice (Cohen, 2008; E. Y. A. Kim & Brown, 2014).

Targeted feedback is a second, consistent leverage point. Web-based instruction with automated, explicit feedback has improved comprehension and production of pragmatic features, particularly where feedback is timely, contingent, and embedded in meaningful tasks (Kerber et al., 2023; Taguchi, 2023). Corrective prompts delivered within online platforms can calibrate learners’ register and mitigation choices (Zhang & Papi, 2021), although instructor mediation remains important for nuance and context sensitivity, as instructors also support students’ problem-solving and communication strategies, helping learners notice pragmatic gaps and self-correct (Eslami et al., 2015).

Mode and modality shape learning affordances. Asynchronous tools (e.g., forums, learning management system [LMS] threads) afford reflection and revision, supporting metapragmatic awareness and deliberate practice (Cohen, 2008; Yang, 2022). Synchronous environments (e.g., video-conferencing, live chat) approximate real-time interaction and enable on-the-spot repair and negotiation that benefit pragmatic fluency (González-Lloret, 2022; Sykes, 2005). Likewise, text-based settings scaffold careful formulation, whereas multimodal environments (audio–video simulations, embodied/virtual reality [VR] tasks) supply prosodic and kinesic cues crucial for appropriateness, albeit with potential cognitive–load trade-offs (Nan, 2022; Qiao & Zhao, 2023). When explicit instruction is coupled with authentic input and guided reflection, short-term interventions can yield measurable pragmatic gains (Hernández & Boero, 2018).

Notwithstanding these advances, recurrent limitations temper impact. Many systems still provide limited personalization, defaulting to one-size-fits-all prompts and feedback that under-attend to proficiency, disciplinary norms, or learners’ interaction histories (Qiao & Zhao, 2023). Insufficient scaffolding is common, with tasks lacking staged progression, modeled language, and faded supports across increasing pragmatic complexity (Cheung, 2021; Satar, 2016). Finally, transfer to real-world academic contexts can be weak when scenarios are decontextualized or when behaviors practiced in structured online tasks do not carry over to spontaneous, high-stakes interaction (Elaish et al., 2022; Xu et al., 2019).

Taken together, the literature supports designs that unite explicit instruction, targeted and timely feedback, and authentic practice across asynchronous/synchronous and text/multimodal modes, while highlighting needs for personalization, principled scaffolding, and improved transfer. The present study responds to this agenda by evaluating a unified platform that integrates automated text analysis, video analysis with multilingual transcription and theory-based annotation, conversation simulations with real-time feedback, AI-supported consultation for just-in-time personalization (i.e., adaptive support tailored to learners’ immediate needs as they arise during interaction), and a digital politeness forum to extend authentic practice within higher-education workflows.

2.3. AI-Mediated Multimodal Scaffolding

Multimodal scaffolding integrates complementary channels of input and interaction, text, audio, video, paralinguistic cues, and interaction traces, to provide contingent, fading support aligned with task demands and learner state (Cosentino et al., 2024; Emerson et al., 2020; A. Nguyen et al., 2022). When coupled with self-regulated learning, it promotes goal setting, strategic monitoring, and reflection through just-in-time prompts that gradually withdraw as competence stabilizes (Azevedo et al., 2022; Järvelä et al., 2023; Munshi et al., 2022).

AI enables this scaffolding at scale by analyzing heterogeneous evidence (e.g., transcripts, prosody, gaze proxies, keystrokes, and clickstreams) to detect inflection points where support is maximally beneficial (Cohn et al., 2024; Järvelä et al., 2023). Beyond individual regulation, AI surfaces group-level patterns and can nudge socially shared regulation, prompting turn-taking, equitable participation, or repair in collaborative tasks (Edwards et al., 2024; Spikol et al., 2018). Advances in modeling underpin dynamic personalization of learning pathways and feedback as needs evolve (Cosentino et al., 2024; S. S. Lee & Azman, 2020; Neshaei et al., 2025), while multimodal analytics guide instructors’ orchestration decisions in situ (Emerson et al., 2020; Noroozi et al., 2019).

Applied to academic pragmatics, an integrated configuration can align five interdependent components to deliver explicit noticing, authentic practice, and adaptive guidance. Automated text analysis flags speech-act realizations, mitigation devices, hedging, and potential breaches of conversational principles in student messages and drafts, enabling immediate metapragmatic feedback and iterative revision (Raković et al., 2022). Video analysis combines multilingual transcription with theory-based annotation to surface timing, prosody, stance, and interactional moves in presentations and discussions, extending noticing beyond lexico-grammar (Mills et al., 2022; Ranker, 2021; Yu et al., 2024). Conversation simulations provide low-risk, high-fidelity rehearsals of academic encounters (e.g., office-hour requests, peer critique) with real-time prompts on appropriateness, turn entry, and repair (Canals, 2021; Huang et al., 2022; H. Lee et al., 2019). AI-supported consultation offers on-demand explanations, exemplars, and targeted suggestions aligned to each learner’s interaction traces and forum activity (Moon et al., 2022). Digital politeness forums sustain asynchronous reflection and peer modeling of strategies, supported by light analytics-informed facilitation to maintain quality and inclusion (Dressen-Hammouda & Wigham, 2022; Poquet et al., 2022).

A balanced orchestration blends asynchronous and synchronous modes to leverage complementary affordances: reflection-friendly drafting and annotation alongside live pragmatics practice with immediate feedback (Cohen, 2008; Sykes, 2005). Likewise, text-centric work supports careful form–function mapping, while multimodal tasks cultivate sensitivity to prosody, timing, gaze, and gesture that is critical for academic face-work (Beltrán-Palanques & Querol-Julián, 2018; Qiao & Zhao, 2023). A typical cycle sequences pretask noticing via text/video analytics, followed by guided simulation, then targeted AI consultation, forum-based reflection and peer exchange, and finally revision; analytics drive personalization and the gradual fading of support across iterations.

Empirical work indicates that technology-mediated tasks, automated feedback, and virtual exchanges can advance pragmatic development by providing explicit instruction, timely feedback, and richer practice than conventional classrooms (González-Lloret, 2022; Kerber et al., 2023; Maa & Taguchi, 2022; Taguchi, 2015). Nonetheless, limitations persist: many systems personalize weakly, offer insufficiently staged scaffolding, or underprepare learners for transfer to spontaneous, high-stakes contexts (Cheung, 2021; Elaish et al., 2022; Nan, 2022; Xu et al., 2019). Crucially, most studies evaluate single modalities in isolation rather than integrated ecosystems that align analytics, practice, consultation, and community features around pragmatics-specific outcomes (Järvelä et al., 2023; Qiao & Zhao, 2023; Sharma et al., 2024; Wei, 2023).

As integration deepens, ethical and practical guardrails are essential. Transparent data practices, opt-in analytics, bias audits, and designs that keep human judgement central are required to sustain trust and protect identity, particularly when feedback addresses face and politeness in intercultural cohorts (Blackmon & Major, 2023; Mutimukwe et al., 2022).

Building on these foundations, the present study implements a five-part, AI-mediated multimodal configuration, automated text and video analytics, conversation simulations, AI consultation, and a digital politeness forum, and evaluates its classroom use in multilingual higher education. The design premise is that analytics amplify noticing and metapragmatic awareness, simulations build procedural fluency under communicative pressure, consultation accelerates uptake through contingent explanation, and forum-mediated reflection consolidates confidence and transfer; the empirical sections that follow examine overall effectiveness and feature-level contributions within this integrated pedagogy.

3.1. Research Design

This study used a quasi-experimental, mixed-methods design with a pretest–posttest, nonequivalent control group. Two intact class sections of a semester-long pragmatics course at two Indonesian public universities were allocated at the class level to an experimental condition (AI-mediated multimodal platform) or a control condition (conventional pragmatics instruction). The 5-week intervention was embedded in regular coursework (Weeks 6–10). The quantitative strand provided the primary test of effectiveness; qualitative evidence (brief weekly reflections and end-unit discussions) was embedded to explain observed effects and illuminate learner experience. All instruction, tasks, and assessments were delivered in Indonesian; “multilingual” refers to the range of first languages across the cohort rather than the medium of teaching.

To mitigate selection bias from nonrandom assignment, baseline equivalence was checked on the pragmatic pretest, Indonesian academic proficiency measured by the UKBI (Uji Kemahiran Berbahasa Indonesia) overall score, a nationally standardized test developed by Badan Bahasa (Ministry of Education) that covers listening, responding to rules, reading, writing, and speaking, with an overall score scale of roughly 251–800 and 7 proficiency predicates from VII “Terbatas” [limited] to I “Istimewa” [exceptional], as well as grade point average (GPA; 0–4), prior overseas/virtual-exchange experience, and year level. Any residual imbalance was addressed analytically via analysis of covariance (ANCOVA; pretest and covariates) with propensity-score sensitivity checks. Instructor effects were managed through aligned syllabi and weekly objectives, shared assessment rubrics, and standardized contact hours; potential contamination was reduced by scheduling sections at different times and using separate LMS shells. Ethical safeguards included informed consent, opt-in analytics, deidentification of logs, and secure storage.

3.2. Participants

Participants were undergraduates enrolled in pragmatics (N = 240; experimental group = 120, control group = 120). Cohorts reflected typical Indonesian language-and-literature enrolments: predominantly domestic students with a small subset of international students who met institutional requirements for Indonesian-medium study. Domestic students were native or near-native speakers of Indonesian whose first languages included regional languages such as Minangkabau, Javanese, Sundanese, and Batak. International students came from Malaysia, Timor-Leste, Thailand, South Korea, and Japan; all had achieved a minimum UKBI score of 500 (Predicate IV, “Madya” [intermediate]) and were studying through full immersion. Competence levels thus ranged from advanced non-native (Predicate IV) among international students to near-native proficiency (Predicates II–III) among domestic students. Baseline comparability across conditions was verified prior to outcome analyses. Table 1 presents the baseline characteristics.

OPEN IN VIEWER

Table 1.

Baseline Characteristics of Participants by Condition.

Characteristic	Experimental group(n = 120)	Control group(n = 120)	Total (N = 240)
Age (years), M (SD)	20.8 (1.0)	20.7 (1.1)	20.7 (1.1)
Gender, n (%) — female / male	81 (67.5) / 39 (32.5)	79 (65.8) / 41 (34.2)	160 (66.7) / 80 (33.3)
GPA (0–4), M (SD)	3.48 (0.24)	3.46 (0.26)	3.47 (0.25)
UKBI overall score (≈251–800), M (SD)	629.1 (45.2)	627.6 (47.8)	628.3 (46.5)
Year level, n (%) — Year 2 / Year 3	58 (48.3) / 62 (51.7)	60 (50.0) / 60 (50.0)	118 (49.2) / 122 (50.8)
First language — domestic, n (%)
Minangkabau	43 (35.8)	42 (35.0)	85 (35.4)
Javanese	37 (30.8)	38 (31.7)	75 (31.2)
Batak	12 (10.0)	13 (10.8)	25 (10.4)
Sundanese	10 (8.3)	11 (9.2)	21 (8.8)
Other regional	10 (8.3)	10 (8.3)	20 (8.3)
First language — international, n (%)
Malay	2 (1.7)	2 (1.7)	4 (1.7)
Tetun	2 (1.7)	1 (0.8)	3 (1.2)
Thai	2 (1.7)	1 (0.8)	3 (1.2)
Korean	1 (0.8)	1 (0.8)	2 (0.8)
Japanese	1 (0.8)	1 (0.8)	2 (0.8)

3.3. Description of the AI-Mediated Multimodal Platform

This study developed and implemented Tutura (https://tutura.my.id), a browser-based platform that scaffolds academic pragmatic competence through an integrated cycle of noticing → practice → consultation → reflection. The name blends tutor (guide) and tutur (utterance), signaling guided support for context-appropriate language use. Five modules (automated text analysis, video analysis, conversation simulations, an AI consultation assistant, and a digital politeness forum) let learners move from theory-informed analysis to practice and reflective discussion in one workspace. Interfaces and feedback are presented in Indonesian to match course delivery; analytic lenses implement established frameworks for speech acts, (im)politeness, and conversational principles. Because all on-screen labels are in Indonesian, English translations of key interface elements accompany each figure below to orient the reader.

The automated text analysis workspace targets sentences and short academic texts (emails to faculty, forum posts, chat excerpts). Learners paste or type a passage, select one analysis focus (speech-act analysis or politeness-strategy analysis), then choose a relevant theoretical lens. The selector adapts to the chosen focus: Austin and Searle are offered for speech acts, whereas Brown and Levinson, Leech, Grice, Gumperz, and other pragmatics models available in the interface are offered for politeness/maxims. After submission, Tutura highlights candidate acts or strategies in context, provides concise, theory-grounded explanations, and visualizes distributions with selectable pie, bar, or radar charts. Results can be copied or printed to support revision and assessment, as shown in Figures 1 and 2.

OPEN IN VIEWER

Figure 1.

Automated text analysis input interface with focus tabs and theory selector.

OPEN IN VIEWER

Figure 2.

Automated text analysis report with highlighted excerpts, explanations, and chart controls.

In this interface, the header “Analisis Tindak Tutur” (speech-act analysis) names the active mode; the subtitle “Deteksi jenis tindak tutur dan strategi bertutur dari input teks” (detect speech acts and strategies from text) is the task instruction displayed to the student. The second tab reads “Analisis Strategi Bertutur” (politeness-strategy analysis). The theory dropdown is set to J.L. Austin (Lokusi, Ilokusi, Perlokusi), and the sample input shown is a student group-chat excerpt. The “Analisis” (analyze) button initiates the analysis.

In this view, “Hasil Analisis” (analysis results) heads the output. Categories are color-coded; here, Lokusi (locution, 30), Ilokusi (illocution, 50), and Perlokusi (perlocution, 20), with each excerpt highlighted and explained in Indonesian. A “Jenis Chart” (chart type) selector lets the student switch among visualization formats. “Salin Hasil” (copy results) and “Cetak Hasil” (print results) buttons at the bottom allow the report to be archived.

The video analysis workspace extends the same noticing routine to multimodal interaction. Students click a YouTube link, the transcript is aligned to the timeline, and the same two analysis foci and theory menus are applied. Output mirrors the text module: Tutura highlights candidate acts or strategies in context, adds concise, theory-grounded explanations, and visualizes distributions with the same chart options, anchored to time stamps so learners can jump from a chart segment to illustrative moments in the clip to examine how prosody, timing, and stance realize pragmatic intent, as shown in Figure 3.

OPEN IN VIEWER

Figure 3.

Video analysis with synchronized player, transcript, theory-based summary, and distribution chart.

In this interface, the example analyses an Indonesian presidential inaugural address. The results page pairs the embedded video player with a timestamped transcript panel labeled “Hasil Transkrip” (transcript results). An analysis panel presents theory-based annotations, and a bar chart under “Visualisasi Data” (data visualization) plots the category distribution. The button “Analisis Seluruh Transkrip” (analyze full transcript) runs a complete pass over the full video.

Conversation simulations provide low-risk simulated practice of recurrent university encounters (requesting supervisor feedback, negotiating group work, discussing thesis progress, preparing for a presentation). After choosing a scenario and interlocutor role, learners proceed through short scenes whose response options vary in directness, mitigation, and stance; subtle prompts nudge attention to pragmatic trade-offs without interrupting flow, as shown in Figure 4. On completion, a feedback view reports an overall performance score and scene-level diagnostics mapped to the selected framework (e.g., face-threat mitigation; adherence to quantity, quality, relation, and manner), followed by targeted suggestions for an immediate retry. Learners can then open advanced pragmatics analysis, which reanalyzes the full simulation transcript under the chosen lens (speech act or politeness strategy), providing per-turn labels, concise explanations, and distribution charts consistent with the text/video modules, along with export options for documentation, as illustrated in Figure 5.

OPEN IN VIEWER

Figure 4.

Conversation simulations: scenario/role selection and first interaction screen.

OPEN IN VIEWER

Figure 5.

Simulation feedback with per-scene diagnostics, improvement pointers, and access to advanced pragmatics analysis.

In this interface, the landing page is titled “Petualangan Komunikasi Akademik” (academic communication adventure). Four scenario cards read “Meminta Bantuan Dosen” (requesting lecturer’s help), “Presentasi Kelas” (class presentation), “Diskusi Kelompok Akademik” (academic group discussion), and “Konsultasi Skripsi” (thesis consultation). The role menu lists “Dosen Pembimbing” (academic supervisor), “Dosen Mata Kuliah” (course lecturer), and “Koordinator Program” (program coordinator). “Mulai Petualangan” (start adventure) launches the simulation, and a scene indicator (“Adegan 1 dari 5”, Scene 1 of 5) tracks progress.

In this view, “Petualangan Selesai!” (adventure completed!) heads the page, followed by the “Skor Komunikasi” (communication score) and a rating label “Sangat Baik” (very good). “Penilaian Keseluruhan” (overall evaluation) summarizes performance across two tabs: “Evaluasi Umum” (general evaluation) and “Analisis Pragmatik” (pragmatic analysis). “Analisis Lanjutan dengan Teori Pragmatik” (advanced analysis with pragmatic theory) provides the framework-specific reanalysis. “Analisis Per Adegan” (per-scene analysis) lists each scene with a strategy label, e.g., “Strategi: Tidak Langsung (Indirect) & Positif (Positive Politeness),” and a score. “Kekuatan Anda” (your strengths) and “Area Perbaikan” (areas for improvement) close the feedback view.

The AI consultation assistant delivers on-demand support across any pragmatics-related query in Indonesian academic registers. Students can request definitions, contrast frameworks (e.g., Brown and Levinson versus Leech), ask for strategy guidance tied to an ongoing scenario, run rubric-aligned appropriateness checks, or request alternative phrasings tailored to power/distance/degree-of-imposition. Responses blend brief theory refreshers with adaptable example language and persist in a searchable history linked to the learner’s artefacts, as shown in Figure 6.

OPEN IN VIEWER

Figure 6.

AI consultation assistant with concise explanations, cross-framework contrasts, and register-appropriate examples.

In this interface, the header reads “Konsultan Pragmatik AI” (AI pragmatics consultant). A welcome panel lists topics such as “Teori Tindak Tutur” (speech act theory), “Strategi Kesopanan” (politeness strategy), and “Prinsip Kerjasama” (cooperative principle), with suggested starter questions. In the example, the student types “Jelaskan secara singkat empat maksim kerja sama menurut Grice” (briefly explain the four maxims of the cooperative principle according to Grice); the AI tutor replies with numbered definitions, simplified paraphrases (“Sederhananya”), and illustrative Indonesian-language scenarios. The “Riwayat Obrolan” (conversation history) sidebar preserves previous queries.

The digital politeness forum provides a class-bounded discussion space with a familiar board design. Students share dilemmas and insights, attach artefacts (analysis reports, simulation summaries), embed media, and discuss strategy choices in threaded conversations. Lightweight voting (up/down) surfaces useful exemplars, while tags and filters (e.g., requests, refusals, hedging, and power distance) support discovery; instructor moderation sustains quality and inclusion across the intervention weeks, as shown in Figure 7.

OPEN IN VIEWER

Figure 7.

Digital politeness forum showing a class-scoped thread with attached artefacts and voting.

In this interface, the header reads “Forum Kesantunan Digital” (digital politeness forum). Threads are sortable by “Populer” (popular) or “Terbaru” (newest); “Buat Diskusi” (create discussion) opens a new post. In the example thread, “Menulis Email kepada Dosen” (writing email to a lecturer), the author has embedded a screenshot of a draft Gmail message: this is an “attached artefact,” a concrete piece of student work shared for peer critique. The upvote/downvote buttons beneath each post constitute the “voting” mechanism: classmates endorse or flag contributions, and highly endorsed posts rise to the top under the “Populer” sort, helping the class surface useful exemplars. A comment count (“422 komentar”) and threaded replies show peer exchange on register choices.

From an implementation standpoint, Tutura is built with TypeScript/Next.js and a relational datastore for users, classes, artefacts, and interaction logs. Pragmatic analyses are produced via schema-conditioned prompts to the Gemini family (Google AI Studio): for each request, the system fixes category sets, operational definitions, decision rules, and a typed JSON schema, then inserts a small number of in-context exemplars (including edge cases). To stabilize decisions, Tutura uses self-consistency (three independent generations with majority voting) and resolves ties with lightweight rule checks. Rule-based indices, hedge index H = #hedges/#tokens, imperative index I = #imperatives/#clauses, and interrogative ratio Q = #interrogatives/#sentences, flag improbable labels and enrich feedback wording. For a document with N utterances, category distributions are computed as p ( c ) = ( 1 / N ) ∑ u = 1 N y ^ u , c , where y ^ u , c , is the one-hot label vector after voting. Simulation scoring aggregates scene-level metrics with fixed weights: appropriateness 0.40, politeness 0.35, clarity 0.15, and turn-management 0.10. Video metadata and transcripts are retrieved via the YouTube Data API prior to analysis. All large language model (LLM) calls execute server-side; no student artefacts are used for model training. Data are stored in encrypted, deidentified form with role-based access control, rate limiting, and instructor-visible logs, aligning with the ethical procedures described in Section 3.1. Integration across modules enables a repeatable weekly cycle, text/video analysis → conversation simulations → AI consultation → forum reflection, with exportable artefacts to support feedback and evaluation.

3.4. Intervention Procedure

The intervention was embedded in the existing pragmatics course over 5 consecutive teaching weeks (Weeks 6–10), with pretesting and orientation in Week 5. Both conditions received two 100-minute class meetings per week, yielding equivalent total contact time over 10 meetings. In the experimental sections, each week followed a stable analysis → practice → consultation → reflection cycle using Tutura. Class meetings opened with guided noticing tasks: students analyzed short Indonesian academic texts (e.g., emails to faculty, forum posts, and chat excerpts) in automated text analysis and brief clips in video analysis, selecting one focus (speech-act analysis or politeness-strategy analysis) and an appropriate theoretical lens. The same meeting then shifted to conversation simulations to practice comparable situations (e.g., requesting supervisor feedback and negotiating group work). A brief closure activity required students to consult the AI consultation assistant to clarify a concept or request register-appropriate formulations aligned to the simulation, and to post a concise takeaway to the digital politeness forum to support peer modeling and reflective explanation.

The second weekly meeting consolidated learning through applied tasks that mirrored course objectives (e.g., drafting a request message and planning a supervision conversation). Students iterated once more through analysis → simulation → consultation, and continued forum discussion asynchronously between meetings. In Weeks 9 and 10, a short capstone sequence asked students to complete two linked simulations and then run advanced pragmatics analysis on their interaction logs to surface strengths and areas for improvement before posttesting. Timestamped platform logs recorded the following mean engagement indicators over the 5-week period: weekly time on platform, 3.4 hours (SD = 0.9); conversation simulations completed, 8.6 (SD = 2.3); text and video analyses submitted, 4.2 (SD = 1.5); AI consultation queries initiated, 6.1 (SD = 2.8); and forum posts contributed, 3.8 (SD = 1.7). These indicators served as predictor variables in the regression analyses reported in Section 4.1.

The control sections covered the same weekly topics, learning objectives, and contact hours but without access to Tutura: noticing tasks used printed transcripts analyzed under instructor guidance, practice took the form of live role-plays using scenarios from the same pool of academic situations, feedback was delivered through whole-class debriefs and individual instructor comments, and reflection was supported by plenary discussions and short written summaries on the LMS. Materials and activity sequences were codeveloped with the experimental-section instructor to ensure alignment. Implementation fidelity was monitored through a brief observer checklist (sampled in approximately 20% of meetings) and platform analytics dashboards; adherence averaged 92% in the experimental sections and 89% in the control sections, with no substantive deviations. To minimize contamination, sections were scheduled at different times and used separate LMS spaces; instructors were asked not to share materials across conditions during the study window. Ethical procedures followed the safeguards described in Section 3.1.

3.5. Instruments

This study employed four instruments, a pragmatic competence test, a DCT, a self-report questionnaire, and qualitative reflections and discussions, to capture both measurable outcomes and learner perspectives. Each is described in the following.

3.5.1. Pragmatic Competence Test

The first instrument was a validated pragmatic competence test, administered in pretest and posttest sessions to both experimental and control groups. It comprised 40 MDCTs representing recurrent academic communication contexts, including emailing faculty, participating in seminar discussions, consulting supervisors, negotiating group projects, and responding in academic presentations. Each item presented a situational prompt and four response options varying in speech-act type, politeness strategy, and degree of appropriateness. Learners selected the option judged most suitable, enabling the assessment of their pragmatic awareness and decision-making in context.

The test specification was anchored in core pragmatic theories widely recognized in the field, including speech-act theory (Austin, 1962; Searle, 1976), cooperative principles (Grice, 1975), politeness principles (Leech, 2014), politeness theory (Brown & Levinson, 1987), and interactional sociolinguistics (Gumperz, 1982). These frameworks provide clear analytic categories for classroom learning and assessment while remaining accessible to students. To ensure theoretical robustness and contextual validity, the design also incorporated broader perspectives such as facework (Goffman, 1967), rapport management (Spencer-Oatey, 2008), intercultural pragmatics (Blum-Kulka et al., 1989; House & Kasper, 1981; Scollon & Scollon, 1995; Kecskes, 2014), instructed pragmatics research (Ishihara & Cohen, 2010; Kasper & Rose, 2002; Taguchi, 2011, 2019), and relational work and (im)politeness frameworks (Culpeper, 2011; Locher & Watts, 2005). This combination reflects both the practical usability of the test and the analytical flexibility needed to scaffold pragmatic noticing, practice, and reflection through multiple lenses rather than a single prescriptive model.

Content validity was established through expert review by three applied linguists, while a pilot administration with 30 students yielded satisfactory internal consistency (Cronbach’s α = .82). Table 2 presents an overview of the test scenarios and target pragmatic functions.

OPEN IN VIEWER

Table 2.

Overview of Test Scenarios and Target Pragmatic Functions.

Item	Scenario type	Example prompt/context	Target speech act/strategy	Supporting theory	Measurement focus
1	Email to lecturer	Requesting deadline extension	Request / indirectness	Brown and Levinson (1987)	Mitigation and appropriateness in formal email
2	Seminar discussion	Responding to peer’s disagreement	Disagreement / politeness	Leech (2014)	Balancing critique with solidarity
3	Thesis consultation	Reacting to supervisor’s suggestion	Acceptance / refusal	Searle (1976)	Managing power and authority relations
4	Group project meeting	Proposing division of tasks	Suggestion / persuasion	Grice (1975)	Cooperative negotiation strategies
5	Presentation Q&A	Answering a challenging audience question	Explanation / facework	Goffman (1967)	Maintaining authority while saving face
6	Email to faculty	Requesting recommendation letter	Request / formal politeness	Brown and Levinson (1987)	Formality and indirect request strategies
7	Academic workshop	Introducing oneself formally	Self-introduction / illocution	Austin (1962)	Locution–illocution alignment
8	Group discussion	Disagreeing politely with peer	Disagreement / rapport management	Spencer-Oatey (2008)	Preserving peer solidarity
9	Thesis defense	Responding to examiner’s critique	Acceptance / justification	Leech (2014)	Strategic disagreement with authority
10	Academic email	Clarifying assignment instruction	Information-seeking / request	Grice (1975)	Maxim of relevance and quantity
. . .	. . .	. . .	. . .	. . .	. . .
40	Academic ceremony speech	Expressing gratitude to audience	Thanking / positive politeness	Brown and Levinson (1987)	Building solidarity in public discourse

3.5.2. Discourse Completion Task

The second instrument was an open-ended DCT designed to assess learners’ productive ability to generate pragmatically appropriate responses in authentic academic situations. Whereas the pragmatic competence test measured receptive awareness through multiple-choice recognition, this instrument required participants to compose their own responses, thereby capturing their capacity to construct context-sensitive language in real time. Ten situational prompts were included, covering both oral and written domains of academic interaction, such as seminar discussions, supervisor consultations, group work, email communication, and ceremonial speech. This design ensured that the DCT tapped a wider range of performance skills, from managing turn-taking to mitigating face threats and demonstrating register alignment. For example, one DCT item presented the following scenario: “Your thesis supervisor has suggested a major change to your research design that you believe is unnecessary. Write a response in which you acknowledge the suggestion, explain your reservations, and propose an alternative approach.” This prompt was designed to elicit face-threatening speech acts (disagreement, refusal) requiring mitigation strategies such as hedging, justification, and positive politeness, thereby assessing learners’ ability to negotiate hierarchical academic relationships.

Content validity was established through expert review by three applied linguists, and a pilot administration with 30 students confirmed the clarity of prompts and their potential to elicit pragmatically rich responses. Inter-rater reliability for coding was satisfactory (Cohen’s κ = .81), demonstrating consistent application of analytic categories across raters. Responses were evaluated using the same pragmatic frameworks applied in the pragmatic competence test to maintain consistency in theoretical grounding. Table 3 presents an overview of the open-ended DCT scenarios and expected pragmatic features.

OPEN IN VIEWER

Table 3.

Overview of Open-Ended DCT Scenarios and Expected Pragmatic Features.

Item	Scenario type	Example prompt/context	Expected pragmatic feature	Evaluation dimension
1	Seminar discussion	Politely interrupting a peer to add a point	Turn-taking / mitigation	Managing floor and solidarity
2	Supervisor office hours	Asking clarification on confusing feedback	Clarification / request	Clarity and appropriateness in hierarchical talk
3	Group chat (online)	Coordinating a meeting time with peers	Suggestion / negotiation	Cooperative digital interaction
4	Academic email	Apologizing for late submission	Apology / facework	Accountability and face-saving
5	Presentation Q&A	Deflecting a critical question	Evasion / justification	Authority maintenance and audience management
6	Peer feedback session	Offering constructive critique on a draft	Critique / politeness	Balancing critique and encouragement
7	Workshop introduction	Greeting and introducing oneself to strangers	Self-introduction / stance	Formality and register alignment
8	Group project meeting	Persuading peers to adopt your proposal	Persuasion / justification	Negotiation strength with mitigation
9	Thesis defense	Responding to examiner’s suggestion	Acceptance / conditional refusal	Authority negotiation and respect
10	Academic ceremony closing	Delivering closing thanks to organizers	Thanking / positive politeness	Solidarity building in formal public discourse

3.5.3. Self-Report Questionnaire

The third instrument was a self-report questionnaire designed to capture learners’ perceptions of their pragmatic awareness, confidence, and experience with the AI-mediated multimodal platform. While the pragmatic competence test and DCT assessed performance outcomes, this questionnaire provided complementary evidence on learners’ metapragmatic reflection and affective engagement. The instrument comprised 20 items on a 5-point Likert scale (1 = strongly disagree to 5 = strongly agree), organized into three subscales: pragmatic awareness (8 items, e.g., noticing speech acts, politeness strategies, and maxims), confidence in academic communication (6 items, e.g., interacting with lecturers and peers appropriately), and perceived platform support (6 items, e.g., feedback usefulness, authenticity of simulations).

Content validity was reviewed by three applied linguists to ensure alignment with theoretical constructs, and a pilot administration with 30 students indicated acceptable clarity and satisfactory reliability (Cronbach’s α = .86 overall). Table 4 presents the structure of the questionnaire and one illustrative item from each subscale, demonstrating how the construct was operationalized.

OPEN IN VIEWER

Table 4.

Structure of the Self-Report Questionnaire.

Subscale	Item (example)	Sample item	Measurement focus
Pragmatic awareness	Item 4	“I can identify whether a request is phrased directly or indirectly.”	Sensitivity to speech act strategies
Confidence in communication	Item 11	“I feel confident asking my supervisor for clarification politely.”	Self-efficacy in hierarchical relations
Platform support	Item 17	“The platform’s feedback helps me improve my pragmatic choices.”	Perceived usefulness of AI-based support

3.6. Data Analysis

Data analysis followed a mixed-methods approach consistent with the quasi-experimental design of the study. Quantitative data from the pragmatic competence test and DCT were first analyzed descriptively to establish means, standard deviations, and distributional properties; normality was assessed through Shapiro–Wilk tests and Q–Q plots, and reliability coefficients (Cronbach’s α, Cohen’s κ) were calculated for the relevant instruments. Pretest–posttest comparisons between experimental and control groups were then conducted using paired-sample and independent-sample t tests when parametric assumptions were met, and Wilcoxon signed-rank or Mann–Whitney U tests when those assumptions were not satisfied. In addition, hierarchical regression analyses were performed, with engagement indicators (e.g., simulation completions and consultation frequency) entered as predictors, and effect sizes reported using Cohen’s d and standardized beta coefficients. For the self-report questionnaire, exploratory factor analysis confirmed the dimensionality of the subscales (pragmatic awareness, confidence, perceived platform support), internal consistency was established through Cronbach’s α, and pre–post differences were tested with paired-sample t tests, supplemented by correlation analyses linking self-reported gains to objective performance.

Qualitative data from reflections and discussions were subjected to inductive and deductive thematic analysis, guided by pragmatic categories (speech acts, politeness strategies, maxims, relational work) and emergent themes. Two independent raters achieved intercoder reliability above κ = .80, with discrepancies resolved through discussion; representative excerpts were then selected to illustrate key findings.

Finally, triangulation across test scores, DCT performance, questionnaire responses, and qualitative accounts enabled the integration of convergent and divergent patterns, providing a comprehensive understanding of how the AI-mediated platform supported learners’ pragmatic competence.

4.1. Quantitative Results

The quantitative strand examined whether the AI-mediated multimodal platform yielded measurable improvements in students’ pragmatic competence, reported in the same sequence as the instruments so as to mirror the progression from receptive recognition to productive performance and metapragmatic reflection. Independent-samples t tests and χ² tests confirmed baseline equivalence across all pretest measures (pragmatic competence test, DCT, questionnaire subscales), UKBI overall score, GPA, first language distribution, year level, and prior exchange experience; no comparison reached statistical significance (all p > .25). Assumption checks (normality and variance homogeneity) were satisfactory; thus, parametric tests are reported, with ANCOVA used to adjust for baseline where appropriate.

Across the pragmatic competence test, both groups improved over 5 weeks, but the magnitude of change differed markedly: the experimental group’s gain was nearly three times that of the control group, with a large within-group effect (d = 0.93) indicating practically meaningful improvement, whereas the control cohort showed only a small effect (d = 0.31), as presented in Table 5. An ANCOVA adjusting for baseline pretest scores confirmed a robust between-group difference, F(1, 236) = 52.30, p < .001, partial η² = .18, indicating that the platform accounted for approximately 18% of variance in posttest performance beyond what baseline differences would predict.

OPEN IN VIEWER

Table 5.

Pragmatic Competence Test Results (Pre–Post Comparisons; N = 240).

Group	Pretest M (SD)	Posttest M (SD)	Gain	t (df)	p	Cohen’s d
Experimental	71.0 (6.8)	86.0 (6.9)	+15.0	14.98 (119)	<.001	0.93
Control	70.9 (7.1)	76.1 (7.0)	+5.2	6.45 (119)	<.001	0.31

Performance on the open-ended DCT, which targeted productive, context-sensitive language, followed the same directional pattern, as detailed in Table 6. Notably, the productive-task effect size was the largest observed in the study (d = 0.97), suggesting that the platform’s iterative simulation-and-feedback cycle supported not only declarative recognition but also the real-time construction of pragmatically appropriate responses. ANCOVA confirmed a significant treatment advantage, F(1, 236) = 45.11, p < .001, partial η² = .16, reinforcing that productive pragmatic skills, often the most resistant to instruction, benefited substantially from the integrated practice environment.

OPEN IN VIEWER

Table 6.

DCT Results (Pre–Post Comparisons; N = 240).

Group	Pretest M (SD)	Posttest M (SD)	Gain	t (df)	p	Cohen’s d
Experimental	70.2 (7.7)	84.3 (7.8)	+14.1	13.82 (119)	<.001	0.97
Control	69.8 (7.9)	74.6 (7.7)	+4.8	5.41 (119)	<.001	0.25

Convergent evidence was obtained from the self-report questionnaire. Exploratory factor analysis supported the expected three-factor solution (pragmatic awareness, confidence in academic communication, and perceived platform support) with high internal consistency (Cronbach’s α = .83–.88). In the experimental group, all three subscales yielded medium-to-large effects, whereas the control group showed only negligible-to-small shifts, as summarized in Table 7. The perceived-platform-support subscale registered the largest experimental gain, suggesting that learners recognized the adaptive, feedback-rich environment as directly beneficial to their pragmatic development.

OPEN IN VIEWER

Table 7.

Self-Report Questionnaire Results (Pre–Post Comparisons; N = 240).

Subscale	Group	PretestM (SD)	PosttestM (SD)	Gain	t (df)	p	Cohen’s d
Pragmatic awareness	Experimental	3.10 (0.57)	3.92 (0.60)	+0.82	11.60 (119)	<.001	0.77
	Control	3.13 (0.56)	3.36 (0.60)	+0.23	3.05 (119)	.003	0.18
Confidence in communication	Experimental	3.07 (0.61)	3.86 (0.63)	+0.79	11.98 (119)	<.001	0.79
	Control	3.09 (0.59)	3.41 (0.64)	+0.32	3.52 (119)	.001	0.21
Platform support	Experimental	3.06 (0.62)	4.01 (0.58)	+0.95	14.05 (119)	<.001	0.89
	Control	3.08 (0.61)	3.34 (0.63)	+0.26	2.80 (119)	.006	0.16

To illuminate mechanisms of impact, hierarchical regression models entered engagement indicators as predictors of posttest pragmatic competence (with pretest scores in the first block). Consistent participation in conversation simulations emerged as the strongest predictor (β = 0.58, p < .001), followed by consultation-assistant use (β = 0.29, p = .012). The final model accounted for a substantial proportion of variance (adjusted R² = .42) with acceptable collinearity diagnostics, highlighting the role of repeated practice and targeted feedback.

In sum, the experimental group outperformed the control group across all three instruments, with effect sizes consistently in the medium-to-large range. The parallel gains on two methodologically complementary measures, a receptive multiple-choice test and a productive open-ended DCT, strengthen the inference that learners improved in both recognizing and generating contextually appropriate pragmatic strategies. Questionnaire data corroborated these performance gains with parallel increases in self-reported awareness, confidence, and perceived support, while regression analyses identified simulation-based practice and consultation use as the strongest behavioral predictors of outcome. These converging lines of evidence, obtained under a controlled design with covariate adjustment, support the claim that the AI-mediated platform fostered educationally meaningful growth in academic pragmatic competence. The qualitative strand that follows examines the experiential mechanisms behind these gains.

4.2. Qualitative Results

The qualitative strand of the study provided experiential depth to the quantitative outcomes by revealing how students made sense of pragmatic learning through the AI-mediated multimodal platform. Analysis of 480 written reflections and 22 group discussions yielded three recurrent themes that were consistently identified by both independent raters, with intercoder reliability reaching κ = .84. These themes, enhanced pragmatic noticing, increased confidence in hierarchical interactions, and appreciation of a safe practice space, explain the mechanisms through which measurable gains in pragmatic competence were achieved. Two cross-cutting insights on task authenticity and personalized consultation further enriched the findings.

The first theme, enhanced pragmatic noticing, highlighted students’ shift from intuitive judgments to theory-informed awareness. Many learners described how analytic highlights and feedback helped them identify strategies they had overlooked previously. One participant remarked:

“Before, I just relied on feeling if my message sounded polite. Now I can actually see which strategy I used—like indirectness or hedging—and understand why it works in that situation.” (P37)

Another reflection illustrated how noticing translated into revision practices:

“When the system highlighted that my email was too direct, I added a reason first. It immediately changed the tone, and I realized how small adjustments can make a big difference.” (P12)

Some further emphasized how the theoretical scaffolding deepened awareness:

“Seeing my response coded under Grice’s maxims finally explained why it was off-topic.” (P63)

Although a few students noted that explanations were sometimes “too technical,” the overall trend pointed to heightened sensitivity, complementing the significant increases in pragmatic awareness reported in the questionnaire.

The second theme, increased confidence in hierarchical interactions, reflected the way simulations reduced fear of negative evaluation and encouraged learners to engage assertively yet appropriately with authority figures. As one student explained:

“I used to avoid disagreeing with lecturers, but after rehearsing alternative phrasings in the simulation, I could express disagreement respectfully without feeling anxious.” (P84)

Others reported that this rehearsal carried over into real-life encounters:

“I feel more confident asking my supervisor for clarification now, because I already practiced similar situations on the platform.” (P21)

These accounts substantiate the medium-to-large effect sizes observed in the confidence subscale, showing how experiential practice underpinned self-efficacy gains.

The third theme, appreciation of a safe practice space, was voiced repeatedly in both individual reflections and group discussions. Learners valued the opportunity to experiment with pragmatic strategies without social embarrassment, a quality that made the platform distinct from classroom interaction. One participant described:

“If I made a mistake in the simulation, I learned from the feedback without feeling ashamed. It was like practicing privately before trying it in real discussions.” (P95)

Another echoed this emphasis on low-stakes exploration:

“In class I hesitate to try new language because I don’t want to be laughed at, but here I could test different options freely and learn from them.” (P48)

The perception of safety, combined with targeted feedback, explains why engagement indicators such as simulation completions and consultation frequency predicted stronger test performance in the regression analysis. Two additional insights emerged across these themes. First, students praised the authenticity of tasks, emphasizing how scenarios mirrored their daily academic experiences and facilitated immediate transfer:

“The situations were the same as what we face—emails to lecturers, group disagreements, presentation Q&A—so I reused the phrasing right away in my real courses.” (P7)

Second, the AI consultation assistant was recognized as a form of personalized scaffolding. For many, it felt like having an individual tutor providing just-in-time guidance:

“It felt like having a personal tutor. When I needed alternative wording, the explanation was short and practical.” (P59)

At the same time, a minority offered critical nuance:

“Sometimes the feedback was too general and repeated what I already knew, so I had to refine the prompts to get more useful advice.” (P103)

Taken together, the qualitative findings reveal that the platform’s effect was not merely statistical but experientially meaningful: students perceived the tasks, feedback, and interactions as directly relevant to their real academic lives. The three themes, enhanced noticing, increased confidence, and the value of safe practice, together with the cross-cutting appreciation of task authenticity and personalized consultation, illuminate the mechanisms through which the quantitative gains reported in Section 4.1 were achieved and sustained.