Conceptual Foundations in Language
Research
Conceptual Foundations in Language
Research, 1, 260104 (2026)
https://doi.org/10.65553/CFLR.260104
Vibe Coding
and Specification-Driven Corpus Linguistics
Wei Lun Wong*
Faculty of Education,
Universiti Kebangsaan Malaysia, Malaysia
Abstract
Corpus linguistics has long been shaped by the tools
available to its users. Vibe coding, the use of natural-language instructions
to develop software with generative AI, loosens that constraint by letting
teachers and researchers build the instrument their question requires. This
position paper argues that the shift is methodological and proposes
specification-driven corpus linguistics as a third stance alongside
corpus-based and corpus-driven inquiry. In this stance, an analyst’s intuition is written as a natural-language
specification, the specification builds the instrument, and the instrument is
treated as a revisable hypothesis. The paper develops the
Intuition-to-Annotation (I2A) Warrant Chain, a model in which intuition, specification,
comparable evidence and contestable annotation lead to a formative decision,
each move with its own warrant and failure mode. Synthesising argument-based
validation, Leech’s maxims of
corpus annotation, learner corpus research and formative assessment, it states
five testable propositions and extends Leech’s maxims to specification transparency. A constructed
case of pronoun reference in a Malaysian Year 6 narrative illustrates the
model. A four-strand agenda addresses specification fidelity, annotation
validity, pedagogical consequence and reporting standards.
Keywords: vibe coding; corpus linguistics; learner corpus; argument-based validation; formative assessment
An ESL teacher reads Aina met Mei at the library after
school. She gave the librarian a notebook. The passage sounds natural.
However, the teacher hesitates over she. The writer may know exactly who gave
the notebook. A reader must infer it. The teacher could circle the pronoun, ask
the writer to clarify who acted or explore whether the earlier paragraph makes
the reference clear. Each response rests on a different judgement about the
same passage. In formative assessment, that difference matters because it
changes what the learner is asked to do next.
Corpus linguistics offers a
long-standing answer to how such judgements might be grounded. It is to compare
the passage with attested language. However, the comparisons a teacher can make
have always depended on the software at hand. Anthony (2013) observes that the functions of corpus tools largely determine which
research methods are available, and successive generations of tools. McEnery
and Hardie (2012) have extended those
functions while leaving their design to a small number of developers. A concordancer retrieves she efficiently. It does not show
which characters were active in the preceding sentence or record why a teacher
withdrew a flag. The question is often reshaped to fit the tool.
Formative assessment theory
already explains why the next decision matters. Black and Wiliam (2009) locate the formative
function in the elicitation, interpretation and use of evidence to improve
decisions about learning. Sadler (1989) adds that learners need access to quality criteria, opportunities to
compare work with those criteria and strategies for improvement. Studies of
data-driven learning (DDL) indicate that corpus consultation can support
language learning (Boulton & Cobb, 2017). Extending these principles to
discourse-level judgement requires an account of how examples bear on a
particular passage, and of how the instrument that retrieves them was built.
Recent research makes this
question consequential for corpus methodology. Wang et al. (2026) find that published
studies of generative AI (GenAI) in L2 writing instruction concentrate on
feedback, revision and polishing. Vibe coding changes who builds
the instrument. Crosthwaite and Özer (2026) describe an
eight-stage approach to instrument development, illustrated by VetLingua and its corpus-supported writing feedback module.
Wong (2026) distinguishes the
evidence needed for computational correctness, linguistic validity and
pedagogical validity. Pan et al. (2026) illustrate how learners can create tools for their own argumentative
writing. These contributions establish how a tool can be built, what claims its
use may support, and who can shape its design. What remains untheorised
is the methodological status of the specification itself. It is the
natural-language account of what a tool should notice, retrieve and display.
Treated as an engineering by-product, the specification disappears from the
method section. Treated as method, it becomes part of the evidence for every
annotation the tool produces.
This paper takes the second view.
Its central claim is that vibe coding opens a specification-driven stance in
corpus linguistics, in which the analyst’s intuition, written as a specification, becomes an inspectable and
revisable component of corpus inquiry. Within this stance, an annotation
warrants pedagogical action when its interpretation is grounded in an explicit
criterion and relevant evidence, when the specification that generated it is
recoverable, and when the teacher or learner can revisit the judgement in light of a response. The argument draws on a purposive
synthesis of corpus methodology, formative assessment, argument-based
validation, learner corpus research and feedback literacy. It focuses on
classroom assessment of English as a second language (ESL) writing,
particularly discourse features that admit more than one plausible
interpretation. The unit of analysis is the annotation episode, which starts
with an observation about an identifiable passage and follows it through
interpretation, a response and a decision about the next step.
Conceptual
Framework and Synthesis
Specification-driven corpus
linguistics
Tognini-Bonelli (2001) distinguished
corpus-based work, in which the corpus tests categories the analyst brings to
it, from corpus-driven work, in which categories emerge from the data. Both
stances assume a stable instrument between analyst and data. Vibe coding
removes that assumption. An analyst can now state in ordinary language which
contexts to display, which comparisons count and which decisions to record, and
a generative model turns that statement into a working tool. The instrument
becomes a third object of inquiry, alongside theory and data.
We call this stance
specification-driven corpus linguistics. It rests on three commitments. First,
intuition is not replaced by corpus evidence but written as a specification
that can be tested against it. Second, the specification is versioned and reported
as part of the method, because a change in wording can change what the tool
retrieves. Third, the validity of a finding depends on a chain of warrants
running from the specification to the decision it informs. The stance does not
oppose corpus-based or corpus-driven work. It describes what either
becomes when analysts build their own instruments (Table 1).
Table 1. Three stances in corpus linguistics
|
Dimension |
Corpus-based |
Corpus-driven |
Specification-driven |
|
Starting
point |
A
prior category or theory |
Patterns
in the data |
An
intuition written as a specification |
|
Role
of the instrument |
Given;
selected by the analyst |
Given;
selected by the analyst |
Built
from the specification; a revisable hypothesis |
|
What
is reported |
Corpus
and categories |
Corpus
and procedures |
Corpus,
specification and version history |
|
Locus
of validity |
Fit
between category and data |
Robustness
of emergent patterns |
A
chain of warrants from specification to decision |
The Intuition-to-Annotation
Warrant Chain
The
Intuition-to-Annotation (I2A) Warrant Chain models how a specification-driven
annotation comes to justify action. Intuition, specification, comparable
evidence and contestable annotation converge on a formative decision. Each move
licenses the next only when its warrant holds, and each has a characteristic
failure mode (Table 2). The chain adapts Kane’s (2013) argument-based
approach to unscored classroom decisions and maps onto the computational,
linguistic and pedagogical warrants distinguished by Wong (2026). Five propositions
state the chain in testable form.
Proposition 1: Teacher
intuition as a hypothesis
Experienced teachers
often recognise a pattern of difficulty before they can explain it. A tool can
also direct attention to a passage by flagging a pattern specified by the
teacher. Either route raises a question about where a reader might hesitate.
Diagnosing why requires a further inference. In the opening example, two people
in the preceding sentence could plausibly be the referent. The first-mentioned
subject may be easier to retrieve, while events described earlier could favour
the other person. Research on local discourse coherence helps explain why
pronouns must be interpreted in relation to the prominence of their possible
referents (Grosz
et al., 1995).
The first move turns
that recognition into a hypothesis. The teacher identifies a passage,
interprets a possible difficulty in relation to a criterion and asks what
action the evidence warrants. The task criterion might be that a reader can
follow who performs each action. The teacher can then check the surrounding
paragraph or ask a peer reader which person they understood. A flag can be
withdrawn if the wider text settles the question or if the writer has delayed
naming the actor for an intended effect. The proposition predicts that
requiring a criterion and a check of the surrounding text will reduce
unwarranted flags compared with flagging surface patterns alone.
Proposition 2:
Specification as method
The second move is
distinctive to the specification-driven stance. Teachers using vibe coding can
specify an assessment decision in ordinary language and ask a generative model
to build the interface or retrieval function that supports it. For a lesson on
reference, a teacher could specify that the tool should display the preceding
sentences, preserve names as written, show the context and source of a
retrieved comparison, and let a teacher decline a flag. Each clause is a
methodological decision. It fixes the window of context, the comparison set and
the record kept. Such changes give local educators a means to align the tool
with the writing task, a priority in recent work on corpus pedagogy (Curry & McEnery, 2025).
The role of the model
during development and its role during classroom use are distinct (Wong, 2026). AI can generate
code for a tool that retrieves a stored passage and lets a teacher propose a
label. An optional model used at run time can propose a candidate annotation
whose wording and interpretation are then reviewed. Retrieval can be tested
against known examples and passages with missing context. Model suggestions can
be checked against adjudicated cases, including acceptable variants and
competing interpretations. Versioned criteria and examples let teachers see
which rule produced a result and whether a revision changed earlier behaviour.
The characteristic failure is specification drift, in which a revised
instruction silently alters which passages are flagged. The proposition
predicts that teachers given control over criteria, examples and annotation
decisions will adapt the tool more successfully across tasks, with gains
contingent on their ability to inspect changes and on the effort required.
Proposition 3:
Comparability of corpus evidence
Learner and reference
corpora answer different questions. A corpus of comparable learner drafts can
show how writers at a similar stage handle reference when two characters are
active. A reference collection of successfully interpreted narratives can illustrate
several workable strategies. Writers may repeat a name when the actor changes,
use a pronoun when the actor is clear or supply a contextual clue. The relevant
comparison is between choices made under similar discourse demands. In corpus
design, the target population and the variation it contains determine which
generalisations the sample can support (Biber, 1993).
A high frequency of she in a broad English corpus offers
little guidance about this particular referential
decision. Comparisons should preserve enough surrounding text to show the
candidate referents, and the selected texts should suit the task and its
readership. The example set should represent successful local usage, including
Malaysian English where it fits the writing goal. When the corpus offers few
comparable cases, the annotation can disclose the gap and draw on a peer
reading of the actual draft. Because a vibe-coded tool retrieves only what it
is told counts as comparable, comparability must be written into the specification
rather than assumed. The proposition predicts that comparable passages with
sufficient context will support more defensible feedback decisions than generic
concordance lines, even when the number of examples is held constant.
Proposition 4: Contestable
annotation
Error-tagged learner
corpora offer a basis for analysing recurring difficulties and designing
instruction (Granger,
2003). Leech (1993) cautioned that every
act of corpus annotation is also an act of interpretation, and that schemes
should state how and by whom annotation was carried out. A classroom annotation
extends this maxim. It must make enough of the interpretation visible for a
learner or another reader to challenge it. A useful record contains the marked
passage, the assessment criterion, the proposed interpretation, the evidence
and its source, the specification version that produced the flag and a possible
next step. It also records whether a reader confirmed the concern, asked for
more context or withdrew it. If the available evidence supports two readings
that call for different actions, the annotation can stay provisional and ask
the writer a question.
The learner should be
able to see the reason in ordinary language and recover enough context to
respond. In the narrative case, Who gave the notebook? I could read this as Aina or Mei opens a question about meaning.
Zhang’s (2026) account of
distributed metacognitive agency provides a way to locate this reasoning across
the teacher, the learner and the tool. The proposition predicts that exposing
the reason for a flag and inviting a response will increase justified revisions
and justified rejections compared with showing a tag and an example alone.
Proposition
5: Formative decision
An annotation becomes
formative when it informs a decision about what to do next (Black & Wiliam, 2009). In the example, the
writer might replace she with Mei, keep the pronoun and strengthen
an earlier cue, or explain why the passage already works. Teachers may use
evidence from several comparable annotations to plan a short lesson on
maintaining and shifting reference. A count of flags can reflect how many opportunities
the task creates or how readily the tool flags them.
Feedback literacy
explains why learners need to practise making and acting on judgements (Carless & Boud, 2018). Learners may engage
differently with the same kind of comment (Han & Hyland, 2015). A learner who
repairs the highlighted sentence shows successful use of help on that draft. A
new narrative with an unmarked reference problem can reveal whether the learner
independently notices and handles a similar difficulty. The proposition predicts
that guided comparison and a learner decision will produce stronger independent
monitoring on a later task than an immediate rewrite for the learner to adopt.
Table 2. The Intuition-to-Annotation (I2A) Warrant Chain
|
Move |
Warrant |
Failure mode |
Observable consequence |
|
1. Intuition |
A flag is linked to a task criterion |
Flagging surface patterns |
Checking the criterion and passage reduces
unwarranted flags. |
|
2. Specification |
The tool does what its versioned specification
states (computational) |
Specification drift |
Teachers adapt tools across tasks without losing
earlier behaviour. |
|
3. Comparable evidence |
Examples match the context, readership and
function (linguistic) |
Frequency substituted for comparability |
Matched passages improve the quality of feedback
judgements. |
|
4. Contestable annotation |
Reasoning is visible and open to challenge |
Opaque tagging |
Learners can justify both changes and decisions to
retain wording. |
|
5. Formative decision |
Evidence fits the action taken (pedagogical) |
Counting flags as learning |
Uptake is visible now; independent monitoring is
tested on a new task. |
An illustrative case
The constructed
scenario begins with the draft excerpt about Aina and Mei. The lesson goal is
for another pupil to follow who does what in a short narrative. The teacher’s specification is brief: show the two preceding
sentences, list every named character, retrieve passages in which two
characters are active, and record whether a reader confirmed the concern. The
teacher marks she for review, reads the
surrounding paragraph and asks a peer to identify who handed over the notebook.
A comparison passage shows that a pronoun can still work with two characters in
the story.
Mei brought a notebook to show Aina. Aina waited by the library door. When Mei
reached the desk, she handed the notebook to the librarian. Here, Mei is named
again just before she, and the sequence of actions supports the same reading.
Both passages are constructed for illustration.
The annotation shown to the
learner
She could refer to Aina or Mei in
this paragraph. A classmate read it as Aina. Who did you mean? If you meant
Mei, try one way to make that clear and ask a classmate to read the sentence
again.
Suppose the writer
meant Mei. The writer could revise the second sentence to Mei gave the librarian a notebook and explain that the
name makes clear who carried out the action. A peer rereading can test whether
the revision helped. If the writer shows that earlier context already
establishes Mei as the actor, the teacher can reconsider the original flag.
When the teacher later adapts the tool to opinion essays, it is the
specification, not the code, that is revised, versioned and reported.
Theoretical
and Practical
Implications
Theoretical implications
The framework contributes to
corpus linguistic theory on three fronts: what counts as a methodological
object, how analyst and instrument are related, and where validation takes
place.
The specification as a methodological
object.
Methodological debate in corpus linguistics has addressed corpus design (Biber, 1993), annotation (Leech, 1993), the relation
between theory and data (Tognini-Bonelli, 2001)
and the balance between statistical reporting and linguistic description (Larsson et al., 2022). Each debate made a
previously tacit decision reportable. Anthony (2013) extended this logic to
software, arguing that tool functionality shapes the methods available. While
tools were shared and stable, their limits could be documented once. When each
study generates its own instrument from natural language, the specification
becomes study-specific and must be reported with the same care as the corpus.
Hence, Leech’s (1993) maxims can be
extended. A vibe-coded annotation should be recoverable to the specification
version that produced it, the specification should be published in a form
another analyst can rerun, and revisions should be logged with their effect on
earlier cases.
Instrumental genesis made explicit. Vérillon
and Rabardel (1995) treat an instrument
as an artefact joined to its user’s schemes of use,
formed through instrumentalisation, in which users
adapt the artefact, and instrumentation, in which the artefact shapes how users
act. With off-the-shelf concordancers, instrumentalisation was confined to options a developer
anticipated, and instrumentation went largely unexamined. Vibe coding enlarges instrumentalisation and makes it linguistic, since the
teacher’s adaptation exists as text. The
specification is therefore an analysable trace of instrumental genesis. The
same account identifies the risk. A generated tool carries defaults nobody
requested, and those defaults can shape what teachers notice. Thus, specification
drift is a theoretical concern about how instruments reshape judgement.
Validation located in the annotation
episode. Kane’s (2013) interpretation
builds on Toulmin’s (2003) layout of claim,
data, warrant and rebuttal, and Chapelle et al. (2008) showed how such
arguments can organise validation in language assessment. The I2A chain moves
this logic from the test to the single annotation, which places three
inferences within one episode. A reader may have trouble identifying a
referent. That difficulty may matter for the task criterion. A particular
intervention may help the writer respond. Evidence for one inference can leave
the next undecided, so the warrant is specific to the action. Rebuttals come
from learners themselves when they contest a flag, which makes consequences
part of the evidence, in line with Messick’s (1989) inclusion of
consequences within validity. The chain also gives DDL a defined role in
formative assessment: corpus examples support comparison, while the decision
rests on how the comparison illuminates a local writing problem (cf. Dong & Wang, 2025).
Practical implications
Teachers can begin by specifying
a criterion and collecting a small set of comparable passages that show both
clear and difficult cases. They can write the tool specification in the plain
language they would use with a colleague and save each version with the
examples it was tested on. For pronoun reference, the context usually needs to
include at least the preceding sentence and sometimes more of the paragraph.
During moderation, teachers can compare why they flagged or retained a form and
agree on what further evidence would resolve a disagreement.
An interface can show the learner’s text, a short comparison and the question the
teacher wants the writer to consider. It can keep proposed labels editable and
retain the reason for a changed decision. Source records can preserve the task
and year group while omitting details that identify pupils.
A research agenda for
specification-driven corpus linguistics
The agenda is ordered as a
trajectory from instrument to discipline. Strands 1 and 2 ask whether
vibe-coded tools and their annotations can be trusted; Strand 3 asks whether
they change learning; Strand 4 asks how findings can accumulate. Each strand
states a guiding question, a design and the direction it opens.
Strand 1. Specification fidelity
(computational warrant). Does
the tool do what its specification says, across revisions? Crosthwaite and Özer
(2026) outline a staged
workflow for building instruments, and Wong (2026) separates computational
correctness from linguistic and pedagogical validity. Fidelity now needs direct
testing. Regression sets of passages with known answers, including missing
context and local varieties, can be rerun after every revision to detect drift.
Ask teachers to adapt an annotation design across two genres, comparing a
conventional editor with a vibe-coding environment that allows equivalent
changes. Record what teachers adjust, whether the changes still implement
agreed criteria, and how much time they spend checking earlier cases. The
longer-term direction is a shared benchmark of specification–behaviour pairs
against which new tools and models can be compared.
Strand 2. Annotation validity
(linguistic warrant). When
are vibe-coded annotations defensible readings of learner text? Build a held-out set
of full paragraphs that includes readily interpretable references,
consequential ambiguities and cases that remain uncertain. Ask independent
teachers to record the criterion, their reading and which classroom response,
if any, they would use. Adjudication should preserve plausible alternative
readings rather than force a single gold label, consistent with Leech’s (1993) view of annotation as interpretation. Comparison sets should be
sampled with the attention to representativeness expected of research corpora (Biber, 1993), and analyses should
examine whether local varieties such as Malaysian English are flagged when they
suit the task. Report missed concerns, unwarranted flags and deferred decisions
alongside teacher checking time.
Strand 3. Pedagogical consequence
(pedagogical warrant). Do
annotation episodes build learners’ independent judgement? Meta-analytic
evidence shows that DDL can support learning (Boulton & Cobb, 2017), and a thirty-year
review of DDL research traces which recommended directions the field has and
has not taken up (Boulton & Vyatkina, 2021). Delayed and transfer effects
are among those this strand would address. Compare teacher-approved comments
that give a label and corpus example with comments that also explain why the
example is relevant and invite learners to justify a change or retention
decision, followed by a delayed task with a new prompt and no annotation
assistance. Learners’ accounts of their
decisions can show whether feedback literacy develops (Carless & Boud, 2018). The evidence can
distinguish improved revision of the marked passage from later monitoring of an
unmarked passage.
Strand 4. Reporting standards
(disciplinary warrant). How
can vibe-coded studies accumulate? Replication is rare in second language
research, and Marsden et al. (2018) recommend changes to design, infrastructure and incentives to support
it. Vibe coding adds an obstacle. Two studies naming the same tool may run
different code. Thus, shared conventions should require the specification text,
version log, regression set, the run-time role of any model and the human
checks applied. Because reporting priorities in corpus linguistics shift over
time (Larsson
et al., 2022), this strand would also track whether specification reporting becomes routine
and whether it improves replication.
Together, the strands move the
field from isolated proof-of-concept tools towards cumulative evidence: shared
repositories of specifications, benchmarks and adjudicated cases that allow
vibe-coded corpus research to be compared, replicated and taught.
Conclusion
This paper has argued that vibe
coding marks a methodological turn in corpus linguistics. When analysts build
instruments by describing them, the specification joins theory and data as an
object of inquiry, and specification-driven corpus linguistics names the stance
that follows. The I2A Warrant Chain shows how an intuition becomes a classroom
decision through specification, comparable evidence and contestable annotation,
and where that chain can break. The framework’s immediate claims concern low-stakes writing decisions about particular features. Extension to classroom speaking or
summative scoring would require a different account of the evidence and its
use. Future work should pursue the four strands of the agenda, extend the chain
beyond referential cohesion to other discourse features and learner
populations, and establish reporting conventions for vibe-coded corpus
instruments. The main empirical question is whether learners can draw on these
annotation episodes to make sound judgements in new writing when the
highlighted passage and suggested response are absent.
Anthony, L. (2013). A critical look at
software tools in corpus linguistics. Linguistic
Research, 30(2), 141–161. https://doi.org/10.17250/khisli.30.2.201308.001
Biber, D. (1993). Representativeness in
corpus design. Literary and Linguistic
Computing, 8(4), 243–257. https://doi.org/10.1093/llc/8.4.243
Black, P., & Wiliam, D. (2009).
Developing the theory of formative assessment. Educational Assessment, Evaluation and Accountability, 21(1), 5–31.
https://doi.org/10.1007/s11092-008-9068-5
Boulton, A., & Vyatkina, N. (2021). Thirty years of data-driven learning: Taking stock and
charting new directions over time. Language
Learning & Technology, 25(3), 66–89. https://doi.org/10.64152/10125/73450
Boulton, A., & Cobb, T. (2017). Corpus use in language learning: A meta-analysis. Language Learning, 67(2), 348–393. https://doi.org/10.1111/lang.12224
Carless, D., & Boud, D. (2018). The
development of student feedback literacy: Enabling uptake of feedback. Assessment & Evaluation in Higher
Education, 43(8), 1315–1325. https://doi.org/10.1080/02602938.2018.1463354
Crosthwaite, P., & Özer, M. (2026). Vibecoding as method: A structured workflow for AI-assisted research instrument
development in applied linguistics. Research
Methods in Applied Linguistics, 5(3), Article 100356. https://doi.org/10.1016/j.rmal.2026.100356
Curry, N., & McEnery, T. (2025). Corpus
linguistics for language teaching and learning: A research agenda. Language Teaching, 58(2), 232–251. https://doi.org/10.1017/S0261444824000430
Dong, J., & Wang, H. (2025). Exploring
the effectiveness of a DDL-GenAI teaching mode in students’ EFL writing
classroom. Computer Assisted Language
Learning. Advance online publication. https://doi.org/10.1080/09588221.2025.2577369
Granger, S. (2003). Error-tagged learner
corpora and CALL: A promising synergy. CALICO Journal, 20(3), 465–480. https://doi.org/10.1558/cj.v20i3.465-480
Grosz, B. J., Joshi, A.
K., & Weinstein, S. (1995). Centering: A framework for modeling the local
coherence of discourse. Computational Linguistics,
21(2), 203–225. https://aclanthology.org/J95-2003/
Han, Y., & Hyland,
F. (2015). Exploring learner engagement with written
corrective feedback in a Chinese tertiary EFL classroom. Journal of Second Language Writing, 30, 31–44. https://doi.org/10.1016/j.jslw.2015.08.002
Kane, M. T. (2013). Validating the
interpretations and uses of test scores. Journal
of Educational Measurement, 50(1), 1–73. https://doi.org/10.1111/jedm.12000
Larsson, T., Egbert, J., & Biber, D.
(2022). On the status of statistical reporting versus linguistic description in
corpus linguistics: A ten-year perspective. Corpora,
17(1), 137–157.
https://doi.org/10.3366/cor.2022.0238
Leech, G. (1993). Corpus annotation schemes. Literary and Linguistic Computing, 8(4),
275–281. https://doi.org/10.1093/llc/8.4.275
Marsden, E., Morgan-Short, K., Thompson, S.,
& Abugaber, D. (2018). Replication in second language research: Narrative and
systematic reviews and recommendations for the field. Language Learning, 68(2), 321–391. https://doi.org/10.1111/lang.12286
Pan, M., Li, K., & Lai, C. (2026).
AI-assisted vibe coding with Base44: Creating personalized tools for
argumentative writing. Assessing Writing,
69, Article 101076. https://doi.org/10.1016/j.asw.2026.101076
Sadler, D. R. (1989). Formative assessment
and the design of instructional systems. Instructional
Science, 18(2), 119–144. https://doi.org/10.1007/BF00117714
Tognini-Bonelli, E. (2001). Corpus
linguistics at work. John Benjamins. https://doi.org/10.1075/scl.6
Toulmin, S. E. (2003). The uses of argument (Updated ed.). Cambridge University Press.
Vérillon, P., & Rabardel, P. (1995). Cognition and
artifacts: A contribution to the study of thought in relation to instrumented
activity. European Journal of Psychology
of Education, 10(1), 77–101. https://doi.org/10.1007/BF03172796
Wang, Z., Lei, H.,
& Chen, K. (2026). The influence of GenAI on second language
writing instruction: A systematic review. System,
140, Article 104081. https://doi.org/10.1016/j.system.2026.104081
Wong, W. L. (2026). Vibe coding for
corpus-based language research and learning: From working tools to warranted
claims. Emerging Educational Technologies.
https://doi.org/10.65553/EET.260105
Zhang, L. J. (2026). Metacognitive agency in
language learning: A dynamic, distributed, and language-anchored framework. Conceptual Foundations in Language Research,
1, Article 260102. https://doi.org/10.65553/CFLR.260102
Author
bio
Wei Lun Wong is a senior lecturer
at Faculty of Education, Universiti Kebangsaan
Malaysia. His research interests include GenAI learner corpora, teacher
education and applied linguistics. His works appear at International Journal of Applied
Linguistics, The Asia-Pacific Education Research, Research in Science and
Technological Education, SAGE Open, Applied Linguistics, Innovation in Language
Teaching and Learning and Scientific Reports among others.
Email: colinw@ukm.edu.my
ORCID: https://orcid.org/0000-0002-4367-0157