A student asks an AI tool:
Give me formal alternatives to shows that.
The tool answers: demonstrates that, indicates that, provides evidence that, is suggestive of, lends support to the view that.
All of these can be useful.
But now comes the harder question: Which ones are actually used in the kind of academic writing I am trying to produce?
Then: What words normally come before and after them? Which belong to Science, Humanities, argumentative essays or research writing?
Fluent generation solves one problem. It does not automatically solve usage verification.
A 2026 study in Language Teaching Research examined this distinction directly.
Jun Rong, Yaochen Deng and Dilin Liu studied 210 undergraduate students across six intact writing classes.
The researchers compared three conditions:
- a data-driven learning group using corpus evidence;
- a group using GenAI together with corpus analysis of teacher-generated concordance lines;
- a control group.
The target was not isolated vocabulary. It was multiword constructions in written academic English.
Learning was assessed using a multiword-construction production test, a writing test, overall writing quality, frequency of target constructions and range of construction types.
Both experimental groups improved more than the control group. But the combined GenAI + corpus condition significantly outperformed the corpus-only condition on nearly all measures in both immediate and delayed testing.
That is a strong result.
But the experiment does not show “AI is better than corpora”. The stronger condition still included corpus analysis.
The study tested generation plus evidence against evidence alone.
That gives us a more useful principle: AI can make academic phrase learning more accessible when learners still have a way to inspect how real language behaves.
Quick answer: what is a multiword construction?
A multiword construction is a recurring sequence that behaves as a useful language unit.
Examples include in relation to, as a result of, the extent to which, plays an important role in, it is important to note that and there is evidence that.
Different research traditions may call related units lexical bundles, formulaic sequences, multiword expressions, phrase frames, collocations or multiword constructions.
The categories are not perfectly identical. For students, the practical idea is: English is built from reusable word combinations as well as individual words.
Why this matters for vocabulary
A learner may know evidence but still write make evidence.
The word meaning is present. The phrase knowledge is weak.
Natural English uses combinations such as provide evidence, gather evidence, strong evidence, evidence suggests, evidence for and evidence against.
Vocabulary therefore includes company. A word’s normal neighbours are part of knowing the word.
AI is very good at proposing phrase candidates
Ask: How can I write that a policy had an effect?
AI may suggest had an impact on, contributed to, led to, resulted in, was associated with and played a role in.
This is useful because it turns one idea into several linguistic options. For a learner, that can reduce the cost of finding possibilities.
But the next stage is selection.
A language model predicts plausible language
That is powerful. It also creates a verification problem.
A generated phrase may be grammatical, possible, uncommon, genre-mismatched, semantically too strong, too conversational or subtly awkward.
The phrase can sound convincing before the learner knows whether it fits.
That is where corpus evidence becomes useful.
What is a corpus?
A corpus is a large, searchable collection of real language.
Depending on the corpus, it may contain news, academic articles, spoken conversation, books, learner writing or disciplinary texts.
Instead of asking only Can this phrase exist?, a corpus lets us inspect How has this phrase actually been used?
That changes the evidence.
A concordance line shows language in its neighbourhood
Suppose the learner searches plays an important role in.
A corpus may show lines such as:
X plays an important role in regulating Y.
X plays an important role in the development of Y.
X plays an important role in determining Y.
Now the learner can observe grammatical pattern, typical following forms, subject types and disciplinary contexts.
This is usage evidence.
AI gives breadth; corpus gives constraint
That is the elegant partnership.
AI proposes.
Corpus tests.
AI can quickly generate candidate phrases. Corpus evidence can show which patterns recur.
The learner then makes a linguistic judgement. That judgement is the learning event.
The 2026 study supports the combination
The combined group did better than the data-driven-only group on nearly all measured outcomes.
Why might that happen?
The study argues that the combined approach may lower the difficulty of corpus exploration, provide personalised explanations, make patterns easier to notice, accommodate different learner preferences and connect generated guidance with attested examples.
A raw concordance display can be cognitively demanding. A novice may see twenty lines and think: What am I supposed to notice?
AI can help ask: What pattern repeats?
But AI should not summarise the corpus before the learner looks
There is another failure mode.
Teacher gives corpus lines. Student asks AI: Tell me the answer. AI summarises everything. The learner never inspects the data.
Now corpus evidence has become decorative.
The goal is not to outsource noticing. The goal is to support noticing.
A stronger sequence is propose → inspect → compare → use
Target concept: causal influence.
Step 1 — propose: AI suggests gives rise to, leads to, contributes to, results in.
Step 2 — inspect: look at real examples.
Step 3 — compare: Does contributes to imply total cause? No. It usually presents partial causal contribution. Does results in sound stronger? Often yes.
Step 4 — use: Reduced access to public transport may contribute to social isolation among older residents.
Now the learner has meaning + phrase + strength + context.
Academic phrases encode reasoning
To some extent controls degree.
In contrast to controls comparison.
On the basis of controls evidential foundation.
Is associated with controls relationship without necessarily claiming causation.
Academic vocabulary is partly a library of reasoning structures.
The wrong phrase can change the claim
Compare:
X causes Y.
X is associated with Y.
X may contribute to Y.
These are not stylistic alternatives. They carry different epistemic strength.
If an AI tool gives all three under “better academic phrases”, the learner still needs reasoning control.
A corpus can show usage. Subject knowledge must decide truth.
This article is not the ChatGPT-versus-dictionary article
eduKateSG already has a published article comparing ChatGPT and dictionary lookup. That page owns word-meaning lookup and retention.
This page owns multiword academic phrase learning through generation plus attested corpus evidence.
This article is not the general collocation article
Collocation asks which words naturally occur together.
This article asks: how can a learner use GenAI and corpus evidence together to learn those recurring combinations for writing?
Mechanism: generation → evidence → pattern extraction → production.
Singapore Secondary English
A student wants to improve argumentative writing. AI suggests raises the question of, has implications for and should be considered in light of.
Good. Now inspect how these phrases are used.
Then write: The rapid adoption of AI-generated text raises the question of how schools should assess independent student thinking.
The phrase now carries an argumentative function.
General Paper
GP students need phrase precision more than phrase decoration.
Useful families include qualification (to some extent, in certain contexts, subject to, insofar as), evidence (the evidence suggests, on the basis of, is consistent with), cause (contributes to, gives rise to, results in) and contrast (in contrast to, at odds with, at variance with).
A corpus can show how real writers deploy them. AI can help organise the observations.
Science
Target phrase: is associated with.
Science students must not automatically convert association into causation.
A corpus can show the phrase repeatedly appearing in cautious research reporting. AI can explain why. But experimental-design knowledge still decides whether causation is justified.
Mathematics
Multiword constructions appear in mathematical exposition too: with respect to, is equivalent to, can be expressed as, it follows that.
These are reasoning signals. A student may know every individual word and still fail to recognise the discourse function.
Humanities
History and Social Studies rely on causal and comparative constructions such as in response to, in the wake of, at the expense of, was shaped by and contributed to.
These phrases encode relationships among events. The meaning lives partly in the combination.
Corpus frequency is not a command
A phrase appears often. Does that mean use it? No.
Frequency tells us what happens. It does not tell us what this sentence needs.
A phrase may be common in biomedical articles, economics or conversation and still be poor in a Secondary 2 composition.
Corpus evidence needs genre interpretation.
AI output is not corpus evidence
A generated sentence is synthetic output. A corpus line is recorded usage from a language collection.
Both can be useful. They answer different questions.
AI: What could I say?
Corpus: What do people in this dataset actually say?
The learner: What should I say here?
A corpus can also mislead if the corpus is wrong
Suppose the student wants Singapore General Paper but inspects casual social-media English. The phrase patterns may be authentic but genre-mismatched.
Evidence quality depends on source fit.
Ask: Who produced this language? For what purpose? In what genre?
Data-driven learning can feel difficult at first
A concordance is not a textbook explanation. It gives examples. The learner must infer the pattern.
That can be powerful. It can also be overwhelming.
The combined 2026 approach is interesting because GenAI can potentially reduce that entry cost without removing the underlying evidence.
That is augmentation.
Diagnosis before prescription
Student uses AI phrases that sound polished but slightly wrong
Diagnosis: plausible generation has not been checked against usage.
Repair: inspect corpus examples and compare phrase environments.
Student copies frequent phrases without understanding them
Diagnosis: frequency has replaced semantic control.
Repair: explain what reasoning relation the phrase encodes.
Student cannot read concordance lines
Diagnosis: raw data volume exceeds pattern-noticing skill.
Repair: reduce to 8–12 lines and ask one narrow observation question.
Student asks AI to summarise every corpus pattern
Diagnosis: evidence inspection has been outsourced.
Repair: learner identifies the pattern first; AI checks the observation second.
Student uses an academic phrase in the wrong subject
Diagnosis: genre evidence is mismatched.
Repair: inspect discipline-relevant examples.
Student uses more multiword constructions but writing does not improve
Diagnosis: phrase frequency is increasing without argument quality.
Repair: test whether each phrase performs a necessary reasoning function.
A practical GenAI + corpus routine
Target: contributes to.
- Ask AI for three nearby alternatives: causes, contributes to, is associated with.
- State provisional difference: contributes to probably means partial causal role.
- Inspect corpus examples: subject types, object types and disciplinary use.
- Refine: contributes to = one factor helping produce an outcome.
- Contrast: causes = stronger claim.
- Produce: “Long commuting times may contribute to student fatigue.”
- Transfer: Science: “Deforestation contributes to habitat loss.” Humanities: “Food shortages contributed to unrest.”
Now the phrase has reusable structure.
Parents: ask for the evidence behind the phrase
If a child says “AI told me this is formal English,” ask: Can you show me how the phrase is used?
That is a strong literacy habit.
The point is not distrust AI. It is separate suggestion from evidence.
Teachers: choose a small phrase set with a real writing job
Do not give fifty academic phrases. Choose six.
Example set: contributes to, in contrast to, on the basis of, to some extent, is consistent with, raises the question of.
Then connect each to one reasoning function. Vocabulary becomes usable.
AI-assisted practice
A useful public prompt is: “Suggest five academic multiword constructions for expressing partial causation. For each, tell me what corpus pattern I should look for rather than claiming the phrase is correct by authority. After I inspect real examples, quiz me on the semantic strength and ask me to use each phrase in a Singapore GP, Science or Humanities context.”
A quiet literary lens
A high-level Hilary Mantel lens is useful because recurring language can carry institutional habits. A court says things one way. A scientist another. A government report another.
The phrase is not merely elegant. It belongs to a system of use.
Close observation means noticing who uses the phrase, around which words, to make what kind of claim.
Internal-link opportunities
- Lexical Chunks and Phrase Frames
- Light and Delexical Verbs
- How Context Improves Vocabulary
- Lexical Quality
- How Language Works
Connections eduKateAI can learn
GenAI ↔ candidate generation: generative systems can rapidly propose alternative phrases and examples, reducing the search cost of exploring language.
Corpus ↔ attested usage: corpus evidence provides observable instances of how expressions occur in real datasets rather than merely plausible generated output.
Multiword construction ↔ reasoning function: recurring academic phrases often encode causation, qualification, contrast, evidence or stance.
Generation ↔ verification: fluent output becomes educationally stronger when the learner can compare it with independent usage evidence.
Corpus ↔ genre: authentic language is not automatically relevant language; source, discipline and audience determine whether a pattern transfers.
Frequency ↔ judgement: common phrases are candidates, not commands. The writer still decides whether the expression matches the intended claim.
AI ↔ noticing: AI can reduce the difficulty of interpreting concordance evidence without replacing the learner’s own pattern detection.
Subjects ↔ phrase ecology: Science, Mathematics, Humanities and General Paper use overlapping phrase structures but distribute them differently according to disciplinary reasoning.
AI language learning ↔ evidence hierarchy: systems should distinguish generated possibility, observed corpus usage, subject truth and writer choice.
Final checkpoint
What did the 2026 study show?
Not: AI replaces corpus learning.
It showed that students using GenAI together with corpus analysis improved more than students using data-driven corpus learning alone on nearly all measured multiword-construction outcomes.
The useful educational sequence is: generate → inspect → compare → verify → use → revisit.
AI supplies possibility. Corpus evidence supplies constraint. The learner supplies judgement.
Research basis
- Rong, J., Deng, Y., & Liu, D. (2026). Effects of using data-driven and generative AI-assisted instructions on learning multiword constructions in written academic English. Language Teaching Research. First published online 19 January 2026. https://doi.org/10.1177/13621688251412525
- The study included 210 undergraduate students across six writing classes and compared data-driven corpus learning, GenAI combined with teacher-generated corpus concordance analysis, and a control condition.
- Qin, J. (2026 issue; first published 30 December 2025). Enhancing Verb-Noun Collocation Learning among Chinese EFL Learners: A Comparison of Cognitive Linguistics-Inspired, GenAI-Assisted, and Traditional Instruction. International Journal of Applied Linguistics. https://doi.org/10.1111/ijal.70088
- 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.
This article deliberately owns GenAI plus corpus evidence for learning academic multiword constructions. It does not replace eduKateSG’s general collocation, lexical-chunk, ChatGPT-versus-dictionary or AI-vocabulary pages.