Combining AI and corpus technology in language teaching: Professional development and knowledge co-construction using the CoI framework

Synthesis: As corpus technology and generative AI reshape language education, teachers need evidence-based ways to integrate both. Ma, Mei, Crosthwaite, and Ma report an explanatory sequential mixed-methods study of a 20-week online professional development program in which over 150 pre- and in-service language teachers from 28 countries co-developed AI-empowered TPACK for corpus technology (AI-TPACK-CT) inside a Community of Inquiry. Social network analysis of six discussion threads, CoI coding of the most active workshop, lesson-plan comparisons, and interviews with one central and two peripheral teachers link interaction patterns to design outcomes. The headline result is that posting volume and network centrality did not determine learning: peripheral teachers constructed substantial pedagogical knowledge by reading, reflecting, and revising their designs.

Key Findings

1. Workshop 1 was the most active of the eight discussion threads, with 51 posts, 6881 words, 15 participants, and 36 connections.
2. CoI coding of that thread yielded 137 social-presence units against 80 cognitive-presence units, and only 5 resolution instances against 49 exploration units.
3. Olivia was the central teacher with 23 posts and 20 connections, while Jessica (8 posts, 4 connections) and Tessa (2 posts, 2 connections) were peripheral.
4. Olivia's 53 coded units concentrated in learning and practice (21/53), Tessa's 53 in pedagogical language knowledge (15/53), and Jessica's 58 in learning and practice (21/58).
5. Interviews showed that the peripheral teachers learned substantially by reading and revising, so post counts understated their AI-TPACK-CT growth.

A 20-week design-based community

The program enrolled over 150 pre- and in-service language teachers from 28 countries/regions, who chose their own pathways through synchronous online workshops, self-paced Canvas materials, and asynchronous forum discussion. The 20-week schedule ran eight weekly workshops, draft lesson plans in Weeks 9–10, peer commenting in Week 11, revision in Week 12, assessment of final plans in Weeks 13–14, social network analysis in Weeks 15–16, and interviews of roughly 90 minutes in Weeks 19–20.

Two design features matter most. The first is an ice-breaking activity and an explicitly low-pressure climate intended to make experimentation with unfamiliar corpus and AI tools safe. The second is purposeful cross-tier mixing: pre-service teachers, in-service teachers, and participants with deeper expertise in AI, corpora, or pedagogy were encouraged to work across those boundaries. Tasks were scaffolded in small increments — searching a corpus, testing an AI tool, trying an AI-assisted feedback workflow — and then assembled into a lesson template covering input exploration, guided practice, assessment, and reflection, which the authors say reduces cognitive load and supports classroom transfer.

The knowledge construct is the paper's other contribution. Building on TPACK for corpus technology and an earlier five-component account of corpus-based language pedagogy, the authors add an AI knowledge domain and specify AI- and corpus-specific competencies: model awareness and prompt design, data governance and bias mitigation, workflow integration with concordances and analytics, and human–AI task orchestration and verification.

Engagement: warm social presence, thin resolution

The quantitative phase counted posts and participants across all eight discussion threads and ran social network analysis in NodeXL on six of them, dropping two that generated only two replies each. Workshop 1, "CBLP & AI," was selected for content analysis as the most active and interaction-dense thread: 51 posts, 6881 words, 15 participants, and 36 connections.

Coding adapted from Rourke et al. and Garrison et al. produced 137 social-presence units — 39 affective expression, 51 open communication, 47 group cohesion — against 80 cognitive-presence units. Within cognitive presence, exploration dominated with 49 units, ahead of 14 triggering events, 12 integration, and only 5 resolution. Inter-coder reliability was 0.89 for this pass.

Three excerpts carry the interpretation. An early exchange around Olivia's pet-themed activity contains no cognitive presence but is saturated with affect, appreciation, and inclusive language, which the authors argue lowered the threshold for later risk-taking. Jessica's post reports a concrete struggle — the "scary human hands" problem across six AI image tools — and documents systematic comparison of them. Matthew connects Muse AI to two pedagogical goals and assesses its classroom value, supplying one of the five resolution instances. The thin resolution count is attributed to the short workshop window.

Three trajectories of AI-TPACK-CT

The case teachers were selected from the Workshop 1 network. Olivia, a university translation instructor with 15 years of tertiary experience, was the most central teacher with 23 posts and 20 connections; Jessica, an early-career EFL teacher with 5 years of experience, was peripheral with 8 posts and 4 connections; Tessa, an MA student in English language education with no formal teaching experience, was peripheral with 2 posts and 2 connections. All three submitted lesson plans combining corpus tools and AI, which was a selection condition.

Coding against the six knowledge domains (reliability 0.85) shows different profiles rather than a single ladder. Olivia's 53 coded units concentrated in learning and practice (21/53), AI knowledge (11/53), and corpus knowledge (9/53), with 6/53 contextual, 5/53 pedagogical, and 1/53 subject knowledge — consistent with an experienced teacher who already commanded pedagogy and content and was acquiring tools. Jessica's 58 units concentrated in learning and practice (21/58), corpus knowledge (11/58), AI knowledge (10/58), and pedagogical language knowledge (10/58), with 5/58 contextual and 1/58 subject. Tessa's 53 units concentrated in pedagogical language knowledge (15/53), with 12/53 each in AI knowledge, corpus knowledge, and learning and practice, and 1/53 each in subject and contextual knowledge.

The lesson plans make the trajectories concrete. Olivia built an English-translation lesson for first-year BA students at B1–B2 using COCA and RhymeZone with a customized Mizou chatbot and image generation. Jessica designed a past-simple "be" grammar lesson for Junior 1 students with Versatext and SKELL plus an AI Foreign Teacher app. Tessa designed a word-stress pronunciation lesson for secondary students using a learner spoken corpus, a native spoken corpus, Youglish, ChatGPT, and Murf. Revisions from draft to final were traced to peer and facilitator comments and coded as contextual or pedagogical gains.

Peripheral participation and what post counts miss

The interviews complicate the network picture. Tessa checked discussion topics frequently but seldom commented, contributing only when she believed she could add a distinctive voice; she still rearranged her lesson's activity sequence after peers said it better matched discovery learning. Jessica described reading others' posts as the source of perspectives she would not have generated alone, and recalled a reply praising her rapid tool-testing as a source of confidence. Olivia described a polite atmosphere in which every post drew a response.

The authors' interpretation is that social network analysis and presence coding capture visible posting behavior but not reading, private reflection, integration of peers' ideas, or revision of instructional reasoning without a public post. That gap matters for CoI-based professional development because it can make an equitable program look like a program for the vocal minority. The practical claim is that design-based tasks plus sustained asynchronous discourse gave peripheral participants substantial AI-TPACK growth that participation metrics would have missed.

What this means for practice

• Design online professional development so that a low-stakes ice-breaker and explicit respect norms precede any technical task, because the case data tie early social presence to later willingness to report failures and test tools publicly.
• Mix pre- and in-service teachers deliberately: the pre-service teacher in this study gained most in pedagogical language knowledge, while the in-service teachers gained most in learning and practice and contextual knowledge, so each group supplies what the other lacks.
• Do not use post counts or centrality alone to judge participation or award credit; pair network metrics with interviews, lesson-plan revisions, and reflective tasks so reading-only participants are assessed on what they constructed.

Limitations

• Although over 150 teachers enrolled, only 15 contributed posts in the selected thread, and most limited participation to reviewing materials and reading others' contributions, so the engagement picture rests on a small vocal subgroup.
• The qualitative findings rest on three teachers chosen because they completed lesson plans and agreed to interviews; the authors describe them as illustrative vignettes, not generalizable cases.
• The program was short: only 5 resolution instances appeared in the analyzed thread, and classroom implementation of the redesigned lessons was not tracked.
• Quantitative content analysis drew on a single thread, and the per-teacher counts are descriptive proportions over 53 or 58 units that the authors caution are not statistically significant.
• Participation was voluntary, with no comparison group, so the study cannot isolate the CoI design from prior motivation.

Citation

Ma, Q., Mei, F., Crosthwaite, P., & Ma, X. (2026). Combining AI and corpus technology in language teaching: Professional development and knowledge co-construction using the CoI framework. Computers & Education, 256, 105728.