What Is Vocabulary | Contextual Predictability — Cloze Probability, Surprisal and Why Some Words Are Easier to Anticipate

CONTEXTUAL PREDICTABILITY · CLOZE PROBABILITY · SURPRISAL · CONTEXTUAL CONSTRAINT · EXPECTATION · WORD PROCESSING · READING TIME

Contextual predictability describes how expected a word is given its preceding linguistic and situational context. A predictable word is not necessarily common everywhere; it is probable here. Researchers have traditionally estimated predictability with cloze probability, while computational work often uses surprisal, the negative logarithm of contextual probability.

This distinction matters for vocabulary because familiar words are processed differently when context strongly predicts them and when context makes them surprising. Predictability therefore helps explain reading time, listening efficiency, lexical inference and why some contexts make an unfamiliar word easier to learn.

This guide explains cloze tasks, contextual constraint, word probability, surprisal, semantic relatedness, syntactic prediction, collocational expectation, discourse context and prediction during reading and listening. Existing eduKateSG Context Clues and Word Frequency articles remain untouched.

Frequency asks, “How common is this word?” Predictability asks, “How likely is this word here?”

2. Contextual predictability

Contextual predictability describes how expected a word is given the words, syntax, situation and discourse that come before it. A highly constrained context narrows the candidate set strongly; a weak context leaves many continuations plausible.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

3. Cloze probability

Cloze probability is traditionally estimated by giving people a sentence fragment and asking them to supply the next word. The proportion producing a particular target estimates how predictable that word is in that context.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

4. Surprisal

Surprisal is a computational measure based on probability. Lower-probability words have higher surprisal and generally require more processing because they violate stronger expectations.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

5. Predictability is not frequency

A word can be globally frequent but unexpected in one sentence, or relatively rare yet highly predictable in a strongly constraining context. Frequency is a property of distribution; predictability is conditional on context.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

6. Contextual constraint

Contextual constraint refers to how strongly the preceding context limits possible continuations. A high-constraint sentence can make one target overwhelmingly likely; a low-constraint sentence allows many reasonable alternatives.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

7. Prediction during reading

Readers use syntax, meaning and world knowledge to generate expectations. Prediction is probabilistic rather than all-or-nothing, so several candidates can remain active at once.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

8. Prediction during listening

Spoken language adds timing, prosody and rapidly unfolding context. Predictable words can be recognised more efficiently because fewer candidates remain plausible.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

9. Vocabulary knowledge and predictability

A learner can only benefit from prediction if likely candidates are represented in the lexicon. Rich vocabulary expands the candidate set and improves fine-grained expectation.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

10. Unknown words in predictable contexts

A strongly constrained sentence can help a learner infer an unfamiliar word or at least its semantic role. This is one mechanism behind incidental word learning from context.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

11. Known words in unpredictable contexts

A familiar word can become difficult when used in an unexpected sense, register or syntactic pattern. Processing cost can come from low contextual fit rather than lexical rarity.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

12. Surprisal and reading time

Research links higher surprisal with longer reading times and greater processing demand. The effect is graded: less expected words usually require more updating.

The practical lesson is to separate lexical knowledge from contextual support. Strong context can rescue a weak lexical representation, while weak context can expose whether the learner actually knows the word independently.

13. Cloze and human expectation

Cloze tasks capture human continuations directly. They reveal the distribution of expectations rather than merely whether a sentence is grammatical.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

14. Language-model prediction

Modern language models estimate contextual probabilities over words or tokens. Their probabilities can be converted into surprisal, but model expectations are not identical to human expectations.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

15. Semantic relatedness

A word can be semantically related to its context without being the most predictable continuation. Relatedness and predictability overlap but are conceptually distinct.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

16. Syntactic prediction

Grammar constrains likely word classes and constructions. A determiner may make a noun more likely; an auxiliary can narrow the expected verb form.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

17. Collocational prediction

Frequent collocations create strong local expectations. After heavy, rain is more predictable than many semantically possible nouns.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

18. Discourse prediction

Earlier sentences create topic and referential expectations. A locally ambiguous word can become highly predictable when discourse has established the relevant domain.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

19. World knowledge

Real-world plausibility influences prediction. Readers expect events, roles and objects that fit familiar situations.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

20. Learning applications

Predictability can be manipulated in teaching. Strong contexts support initial inference; weaker and varied contexts later test whether the word has become independently known.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

21. Assessment

To measure contextual predictability, separate word knowledge from context strength. A learner may fail because the word is unknown, because context is weak or because a dominant competitor wins.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

22. AI-era prediction

AI makes probabilistic next-word behaviour visible, but educational value comes from comparing predictions with human reasoning and understanding why one continuation is more constrained than another.

Predictability should be treated probabilistically. Several continuations can be plausible, and an unexpected word can still be perfectly grammatical and meaningful.

23. Predictability laboratory — Cases 1–20

1. The chef sharpened the ___ before slicing the vegetables.

Target: knife. Approximate constraint: high. Main source of prediction: world knowledge + syntax.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing knife is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether knife remains predictable. This shows which cue actually drove expectation.

2. Dark clouds gathered and soon it began to ___.

Target: rain. Approximate constraint: high. Main source of prediction: event expectation.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing rain is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether rain remains predictable. This shows which cue actually drove expectation.

3. She spread butter on a slice of ___.

Target: bread. Approximate constraint: high. Main source of prediction: collocation + scenario.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing bread is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether bread remains predictable. This shows which cue actually drove expectation.

4. The baby was tired, so her father put her to ___.

Target: sleep. Approximate constraint: high. Main source of prediction: formulaic construction.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing sleep is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether sleep remains predictable. This shows which cue actually drove expectation.

5. He unlocked the door with a ___.

Target: key. Approximate constraint: high. Main source of prediction: instrument relation.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing key is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether key remains predictable. This shows which cue actually drove expectation.

6. The scientist recorded the results in a laboratory ___.

Target: notebook. Approximate constraint: medium. Main source of prediction: domain + object role.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing notebook is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether notebook remains predictable. This shows which cue actually drove expectation.

7. The crowd cheered when the striker scored a ___.

Target: goal. Approximate constraint: high. Main source of prediction: sports script.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing goal is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether goal remains predictable. This shows which cue actually drove expectation.

8. The teacher wrote the answer on the ___.

Target: board. Approximate constraint: high. Main source of prediction: classroom script.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing board is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether board remains predictable. This shows which cue actually drove expectation.

9. After the storm, several trees had fallen across the ___.

Target: road. Approximate constraint: medium. Main source of prediction: world knowledge.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing road is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether road remains predictable. This shows which cue actually drove expectation.

10. She checked the train timetable before leaving the ___.

Target: house. Approximate constraint: low. Main source of prediction: weak context.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing house is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether house remains predictable. This shows which cue actually drove expectation.

11. The lawyer presented new ___ to support the claim.

Target: evidence. Approximate constraint: high. Main source of prediction: academic/legal collocation.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing evidence is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether evidence remains predictable. This shows which cue actually drove expectation.

12. The doctor prescribed a course of ___.

Target: antibiotics. Approximate constraint: medium. Main source of prediction: medical script.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing antibiotics is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether antibiotics remains predictable. This shows which cue actually drove expectation.

13. The hikers filled their bottles at the mountain ___.

Target: spring. Approximate constraint: medium. Main source of prediction: domain sense.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing spring is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether spring remains predictable. This shows which cue actually drove expectation.

14. He raised his hand because he wanted to ask a ___.

Target: question. Approximate constraint: high. Main source of prediction: classroom phrase.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing question is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether question remains predictable. This shows which cue actually drove expectation.

15. The audience became silent as the curtain began to ___.

Target: rise. Approximate constraint: medium. Main source of prediction: theatre script.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing rise is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether rise remains predictable. This shows which cue actually drove expectation.

16. She wore gloves because the metal was extremely ___.

Target: cold. Approximate constraint: medium. Main source of prediction: causal expectation.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing cold is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether cold remains predictable. This shows which cue actually drove expectation.

17. The company reduced costs without lowering product ___.

Target: quality. Approximate constraint: high. Main source of prediction: business collocation.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing quality is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether quality remains predictable. This shows which cue actually drove expectation.

18. The experiment was repeated to confirm the ___.

Target: result. Approximate constraint: medium. Main source of prediction: research discourse.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing result is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether result remains predictable. This shows which cue actually drove expectation.

19. The witness gave a detailed ___ of what happened.

Target: account. Approximate constraint: medium. Main source of prediction: legal/reporting phrase.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing account is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether account remains predictable. This shows which cue actually drove expectation.

20. The child used a magnifying glass to examine the tiny ___.

Target: insect. Approximate constraint: medium. Main source of prediction: scenario.

Cloze test: hide the target and ask several people for the first continuation that comes to mind. The proportion producing insect is an empirical cloze estimate, not an intuition label.

Competitor analysis: list other plausible completions. High constraint means the distribution is concentrated; low constraint means probability is spread across many alternatives.

Vocabulary test: replace the target with a less familiar but semantically fitting word. If comprehension remains good, context is carrying part of the lexical load.

Transfer: weaken one contextual cue and observe whether insect remains predictable. This shows which cue actually drove expectation.

24. Predictability laboratory — Cases 21–40

21. The river overflowed after days of heavy ___.

Target: rain. Approximate constraint: high. Main source: collocation.

Surprisal thought experiment: imagine the context strongly favours rain. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

22. The pilot checked the weather before the ___.

Target: flight. Approximate constraint: medium. Main source: aviation script.

Surprisal thought experiment: imagine the context strongly favours flight. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

23. The committee reached a final ___.

Target: decision. Approximate constraint: high. Main source: collocation.

Surprisal thought experiment: imagine the context strongly favours decision. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

24. Please attach the file to your ___.

Target: email. Approximate constraint: medium. Main source: digital script.

Surprisal thought experiment: imagine the context strongly favours email. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

25. The students opened their books to page ___.

Target: ten. Approximate constraint: low. Main source: syntactic slot but many values.

Surprisal thought experiment: imagine the context strongly favours ten. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

26. She placed the flowers in a glass ___.

Target: vase. Approximate constraint: high. Main source: object-function.

Surprisal thought experiment: imagine the context strongly favours vase. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

27. The fire alarm sounded and everyone left the ___.

Target: building. Approximate constraint: high. Main source: event script.

Surprisal thought experiment: imagine the context strongly favours building. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

28. The judge listened carefully before announcing the ___.

Target: verdict. Approximate constraint: high. Main source: legal script.

Surprisal thought experiment: imagine the context strongly favours verdict. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

29. The mechanic replaced the worn brake ___.

Target: pads. Approximate constraint: medium. Main source: technical domain.

Surprisal thought experiment: imagine the context strongly favours pads. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

30. The child laughed at the funny ___.

Target: joke. Approximate constraint: high. Main source: semantic expectation.

Surprisal thought experiment: imagine the context strongly favours joke. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

31. The report identified three major ___.

Target: problems. Approximate constraint: medium. Main source: discourse structure.

Surprisal thought experiment: imagine the context strongly favours problems. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

32. The cyclist wore a helmet for ___.

Target: safety. Approximate constraint: high. Main source: purpose relation.

Surprisal thought experiment: imagine the context strongly favours safety. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

33. The team practised every day before the ___.

Target: match. Approximate constraint: medium. Main source: sports script.

Surprisal thought experiment: imagine the context strongly favours match. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

34. She turned down the music because it was too ___.

Target: loud. Approximate constraint: high. Main source: causal adjective.

Surprisal thought experiment: imagine the context strongly favours loud. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

35. He saved the document before closing the ___.

Target: program. Approximate constraint: medium. Main source: computing script.

Surprisal thought experiment: imagine the context strongly favours program. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

36. The gardener watered the plants every ___.

Target: morning. Approximate constraint: low. Main source: many temporal continuations.

Surprisal thought experiment: imagine the context strongly favours morning. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

37. They discussed the proposal during the ___.

Target: meeting. Approximate constraint: medium. Main source: institutional script.

Surprisal thought experiment: imagine the context strongly favours meeting. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

38. The detective searched the room for ___.

Target: clues. Approximate constraint: high. Main source: genre script.

Surprisal thought experiment: imagine the context strongly favours clues. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

39. The student revised carefully for the final ___.

Target: exam. Approximate constraint: high. Main source: education script.

Surprisal thought experiment: imagine the context strongly favours exam. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

40. The boat returned to the harbour before the ___.

Target: storm. Approximate constraint: medium. Main source: weather scenario.

Surprisal thought experiment: imagine the context strongly favours storm. A very different continuation would have lower probability and therefore higher surprisal, even if that alternative word is globally frequent.

Syntax check: identify what word classes and grammatical forms are licensed at the blank. Syntax narrows the candidate set before semantics chooses among them.

World-knowledge check: state the event script or real-world relation that makes the target plausible. Prediction relies on knowledge beyond vocabulary definitions.

Learning use: for a new target, begin with a strongly constraining version, then later practise in less predictive contexts so the word becomes less cue-dependent.

25. Predictability laboratory — Cases 41–60

41. The nurse measured the patient’s blood ___.

Target: pressure. Approximate constraint: high. Main source: medical collocation.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

42. He apologised for the misunderstanding and accepted ___.

Target: responsibility. Approximate constraint: high. Main source: collocation.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

43. The speaker paused to allow the audience to ___.

Target: reflect. Approximate constraint: low. Main source: discourse but many verbs.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

44. She compared the two graphs before drawing a ___.

Target: conclusion. Approximate constraint: high. Main source: academic collocation.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

45. The child hid behind the sofa during the loud ___.

Target: thunder. Approximate constraint: medium. Main source: event expectation.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

46. The engineer tested the bridge under heavy ___.

Target: load. Approximate constraint: high. Main source: technical collocation.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

47. The journalist verified the information before publishing the ___.

Target: story. Approximate constraint: medium. Main source: media script.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

48. The customer asked for a full ___.

Target: refund. Approximate constraint: high. Main source: commercial script.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

49. The hikers followed the marked ___.

Target: trail. Approximate constraint: high. Main source: outdoor collocation.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

50. The phone battery was almost ___.

Target: empty. Approximate constraint: medium. Main source: device state.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

51. The referee blew the whistle to stop the ___.

Target: game. Approximate constraint: medium. Main source: sports script.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

52. She reread the paragraph because the meaning was ___.

Target: unclear. Approximate constraint: medium. Main source: reading context.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

53. The researcher calculated the average from the collected ___.

Target: data. Approximate constraint: high. Main source: research collocation.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

54. The manager scheduled another meeting to discuss the ___.

Target: issue. Approximate constraint: medium. Main source: business discourse.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

55. The children built a sandcastle near the ___.

Target: sea. Approximate constraint: medium. Main source: beach scenario.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

56. The musician tuned the guitar before the ___.

Target: performance. Approximate constraint: medium. Main source: music script.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

57. The patient felt better after taking the ___.

Target: medicine. Approximate constraint: high. Main source: health script.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

58. The road was closed because of a serious ___.

Target: accident. Approximate constraint: medium. Main source: traffic script.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

59. The librarian asked everyone to speak ___.

Target: quietly. Approximate constraint: high. Main source: institutional expectation.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

60. The farmer repaired the fence around the ___.

Target: field. Approximate constraint: medium. Main source: rural scenario.

Relatedness versus predictability: generate a word that is semantically related to the sentence but would not be a natural continuation. This demonstrates that semantic fit is broader than next-word probability.

Frequency versus predictability: compare a globally frequent alternative with the target. A common word can still be surprising if the local context disfavors it.

Discourse extension: add one preceding sentence that increases or decreases target predictability. Predictability can come from discourse, not just the immediately preceding words.

Processing prediction: predict which version should be easier to read or listen to and explain why in probabilistic terms.

26. Teaching with contextual predictability

1. knife

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit world knowledge + syntax so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal knife, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where knife is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

2. rain

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit event expectation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal rain, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where rain is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

3. bread

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit collocation + scenario so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal bread, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where bread is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

4. sleep

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit formulaic construction so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal sleep, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where sleep is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

5. key

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit instrument relation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal key, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where key is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

6. notebook

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit domain + object role so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal notebook, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where notebook is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

7. goal

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit sports script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal goal, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where goal is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

8. board

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit classroom script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal board, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where board is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

9. road

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit world knowledge so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal road, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where road is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

10. house

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit weak context so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal house, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where house is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

11. evidence

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit academic/legal collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal evidence, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where evidence is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

12. antibiotics

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit medical script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal antibiotics, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where antibiotics is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

13. spring

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit domain sense so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal spring, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where spring is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

14. question

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit classroom phrase so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal question, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where question is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

15. rise

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit theatre script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal rise, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where rise is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

16. cold

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit causal expectation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal cold, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where cold is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

17. quality

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit business collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal quality, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where quality is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

18. result

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit research discourse so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal result, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where result is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

19. account

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit legal/reporting phrase so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal account, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where account is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

20. insect

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit scenario so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal insect, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where insect is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

21. rain

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal rain, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where rain is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

22. flight

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit aviation script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal flight, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where flight is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

23. decision

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal decision, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where decision is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

24. email

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit digital script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal email, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where email is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

25. ten

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit syntactic slot but many values so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal ten, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where ten is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

26. vase

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit object-function so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal vase, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where vase is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

27. building

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit event script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal building, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where building is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

28. verdict

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit legal script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal verdict, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where verdict is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

29. pads

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit technical domain so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal pads, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where pads is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

30. joke

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit semantic expectation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal joke, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where joke is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

31. problems

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit discourse structure so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal problems, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where problems is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

32. safety

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit purpose relation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal safety, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where safety is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

33. match

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit sports script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal match, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where match is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

34. loud

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit causal adjective so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal loud, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where loud is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

35. program

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit computing script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal program, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where program is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

36. morning

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit many temporal continuations so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal morning, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where morning is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

37. meeting

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit institutional script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal meeting, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where meeting is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

38. clues

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit genre script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal clues, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where clues is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

39. exam

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit education script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal exam, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where exam is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

40. storm

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit weather scenario so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal storm, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where storm is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

41. pressure

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit medical collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal pressure, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where pressure is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

42. responsibility

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal responsibility, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where responsibility is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

43. reflect

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit discourse but many verbs so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal reflect, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where reflect is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

44. conclusion

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit academic collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal conclusion, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where conclusion is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

45. thunder

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit event expectation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal thunder, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where thunder is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

46. load

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit technical collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal load, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where load is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

47. story

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit media script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal story, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where story is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

48. refund

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit commercial script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal refund, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where refund is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

49. trail

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit outdoor collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal trail, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where trail is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

50. empty

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit device state so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal empty, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where empty is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

51. game

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit sports script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal game, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where game is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

52. unclear

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit reading context so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal unclear, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where unclear is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

53. data

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit research collocation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal data, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where data is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

54. issue

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit business discourse so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal issue, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where issue is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

55. sea

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit beach scenario so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal sea, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where sea is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

56. performance

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit music script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal performance, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where performance is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

57. medicine

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit health script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal medicine, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where medicine is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

58. accident

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit traffic script so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal accident, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where accident is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

59. quietly

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit institutional expectation so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal quietly, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where quietly is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

60. field

Stage 1 — support: present the word inside a context that strongly constrains its meaning. For this item, exploit rural scenario so the learner can make an informed prediction before seeing the target.

Stage 2 — verify: reveal field, check the actual sense, pronunciation, spelling and collocation, and correct any inference that was merely plausible rather than accurate.

Stage 3 — reduce predictability: move the word into a weaker context. The learner should now rely more on stored lexical knowledge than on sentence constraint.

Stage 4 — produce: ask for an original sentence where field is appropriate but not trivially predictable. Productive control should survive when context no longer hands over the answer.

27. Frequently asked questions

What is contextual predictability?

How likely a word is given its preceding context.

What is cloze probability?

The proportion of people who supply a particular continuation for a sentence fragment.

What is surprisal?

A probability-based measure where lower-probability words have higher surprisal.

Is predictability the same as frequency?

No. Frequency is global or corpus-based; predictability is conditional on context.

What is contextual constraint?

How strongly a context narrows possible continuations.

Can a rare word be predictable?

Yes, if the context strongly points to it.

Can a frequent word be surprising?

Yes, if it is unlikely in that specific context.

Does predictability help reading?

Predictable words are generally easier to process because context narrows alternatives.

Can predictability help word learning?

Yes, strongly constraining contexts can support inference, but learners later need less supportive contexts for transfer.

What is the simplest rule?

Ask not only whether the word is known, but how much the context is doing for the learner.

28. Research grounding

Cambridge research on L2 reading notes that predictability has traditionally been estimated with human cloze responses and that surprisal provides a computational alternative: higher surprisal means a word is less probable in context and generally more cognitively demanding to process. This literature treats predictability alongside other lexical properties such as frequency, age of acquisition and semantic relatedness rather than collapsing them into one factor.

• Cambridge — Predictability, surprisal and L2 word processing

29. eduKateSG routes

• Lexical Inferencing and Context Clues
• Word Frequency
• Retrieval Cues
• Vocabulary Master Router

30. Final model

Contextual predictability is conditional vocabulary knowledge in action. The sentence, discourse and world model narrow what is likely before the word arrives.

Strong readers use prediction without becoming trapped by it: they generate expectations, update when the word surprises them, and use the unexpected word to revise their model of the sentence.

Prediction prepares the lexicon. Surprise updates it.

31. Final predictability calibration lab

1. knife

Constraint manipulation: write one version of the sentence that makes knife almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

2. rain

Constraint manipulation: write one version of the sentence that makes rain almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

3. bread

Constraint manipulation: write one version of the sentence that makes bread almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

4. sleep

Constraint manipulation: write one version of the sentence that makes sleep almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

5. key

Constraint manipulation: write one version of the sentence that makes key almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

6. notebook

Constraint manipulation: write one version of the sentence that makes notebook almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

7. goal

Constraint manipulation: write one version of the sentence that makes goal almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

8. board

Constraint manipulation: write one version of the sentence that makes board almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

9. road

Constraint manipulation: write one version of the sentence that makes road almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

10. house

Constraint manipulation: write one version of the sentence that makes house almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

11. evidence

Constraint manipulation: write one version of the sentence that makes evidence almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

12. antibiotics

Constraint manipulation: write one version of the sentence that makes antibiotics almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

13. spring

Constraint manipulation: write one version of the sentence that makes spring almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

14. question

Constraint manipulation: write one version of the sentence that makes question almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

15. rise

Constraint manipulation: write one version of the sentence that makes rise almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

16. cold

Constraint manipulation: write one version of the sentence that makes cold almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

17. quality

Constraint manipulation: write one version of the sentence that makes quality almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

18. result

Constraint manipulation: write one version of the sentence that makes result almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

19. account

Constraint manipulation: write one version of the sentence that makes account almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

20. insect

Constraint manipulation: write one version of the sentence that makes insect almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

21. rain

Constraint manipulation: write one version of the sentence that makes rain almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

22. flight

Constraint manipulation: write one version of the sentence that makes flight almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

23. decision

Constraint manipulation: write one version of the sentence that makes decision almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

24. email

Constraint manipulation: write one version of the sentence that makes email almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

25. ten

Constraint manipulation: write one version of the sentence that makes ten almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

26. vase

Constraint manipulation: write one version of the sentence that makes vase almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

27. building

Constraint manipulation: write one version of the sentence that makes building almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

28. verdict

Constraint manipulation: write one version of the sentence that makes verdict almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

29. pads

Constraint manipulation: write one version of the sentence that makes pads almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

30. joke

Constraint manipulation: write one version of the sentence that makes joke almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

31. problems

Constraint manipulation: write one version of the sentence that makes problems almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

32. safety

Constraint manipulation: write one version of the sentence that makes safety almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

33. match

Constraint manipulation: write one version of the sentence that makes match almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

34. loud

Constraint manipulation: write one version of the sentence that makes loud almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

35. program

Constraint manipulation: write one version of the sentence that makes program almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

36. morning

Constraint manipulation: write one version of the sentence that makes morning almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

37. meeting

Constraint manipulation: write one version of the sentence that makes meeting almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

38. clues

Constraint manipulation: write one version of the sentence that makes clues almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

39. exam

Constraint manipulation: write one version of the sentence that makes exam almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

40. storm

Constraint manipulation: write one version of the sentence that makes storm almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

41. pressure

Constraint manipulation: write one version of the sentence that makes pressure almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

42. responsibility

Constraint manipulation: write one version of the sentence that makes responsibility almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

43. reflect

Constraint manipulation: write one version of the sentence that makes reflect almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

44. conclusion

Constraint manipulation: write one version of the sentence that makes conclusion almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

45. thunder

Constraint manipulation: write one version of the sentence that makes thunder almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

46. load

Constraint manipulation: write one version of the sentence that makes load almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

47. story

Constraint manipulation: write one version of the sentence that makes story almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

48. refund

Constraint manipulation: write one version of the sentence that makes refund almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

49. trail

Constraint manipulation: write one version of the sentence that makes trail almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

50. empty

Constraint manipulation: write one version of the sentence that makes empty almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

51. game

Constraint manipulation: write one version of the sentence that makes game almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

52. unclear

Constraint manipulation: write one version of the sentence that makes unclear almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

53. data

Constraint manipulation: write one version of the sentence that makes data almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

54. issue

Constraint manipulation: write one version of the sentence that makes issue almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

55. sea

Constraint manipulation: write one version of the sentence that makes sea almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

56. performance

Constraint manipulation: write one version of the sentence that makes performance almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

57. medicine

Constraint manipulation: write one version of the sentence that makes medicine almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

58. accident

Constraint manipulation: write one version of the sentence that makes accident almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

59. quietly

Constraint manipulation: write one version of the sentence that makes quietly almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.

60. field

Constraint manipulation: write one version of the sentence that makes field almost inevitable and one that leaves five or more plausible continuations. Keep the grammar comparable so the main difference is contextual constraint.

Human-versus-model question: ask whether people and a language model would likely rank the same continuations. Differences can reveal experience, world knowledge or register assumptions not captured identically by both systems.

Learning decision: begin teaching with the high-constraint version, then verify independent knowledge using the low-constraint version.