Polarity items in translation are expressions whose grammaticality or interpretation depends on the polarity and semantic environment around them. English examples include any, ever, yet, at all and negative additive either, as well as idioms such as lift a finger. Cambridge work on polarity notes that negative polarity items are constrained to negative or related licensing environments, while positive polarity items favour affirmative contexts; questions, conditionals, comparatives and other environments show why the system is more complex than simply ‘negative words.’
People searching for negative polarity items, NPI licensing, positive polarity items, any ever yet at all, free-choice any, downward entailment are solving a licensing-and-scope problem. The same English form any can mean an existential inside negation or a free-choice expression under permission, while another language may use completely different lexical items for those readings. Translation precision therefore requires identifying the item’s licensor and semantic reading before choosing a target equivalent.
This article owns the polarity-items precision layer for eduKateSG. It does not replace Negation and Scope, Quantifiers, Modality or Negative Auxiliaries. Its narrow task is to preserve polarity-sensitive licensing, distinguish NPI, PPI and free-choice readings, track operator scope, and prevent translators or AI systems from adding/removing negation merely to make a target polarity word sound grammatical.
Quick Read
For every polarity-sensitive item write ITEM TYPE, LICENSOR, SCOPE and READING. NPI any under negation is not free-choice any under permission. Ever in a question is not ordinary affirmative temporal ever. Translate the licensed reading, then choose a target item that is grammatical in that environment.
One-Sentence Answer
Translate polarity items by preserving the item–licensor relationship and scope, not by assigning one target word to any, ever, yet or their cross-linguistic counterparts.
Why This Precision Layer Matters
Polarity errors change truth conditions easily. Any can become every; few can become none; a question can gain invented negation; a free-choice permission can become an existential possibility. Because target sentences can still sound grammatical, polarity needs a dedicated operator-level review.
What Polarity Items Are
The first task is to identify the licensing environment. Polarity-sensitive expressions are words or phrases whose acceptability or interpretation depends on the semantic and pragmatic environment in which they occur. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating an item from its dictionary meaning without checking whether the target environment licenses the same expression. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to mark the item as NPI, PPI, free-choice or language-specific polarity expression before choosing a target. For example, English any, ever and at all behave very differently in ordinary affirmative statements than under negation or questions. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Negative Polarity Items
This is a semantic-syntactic dependency rather than ordinary lexical selection. Negative polarity items are expressions that are licensed in negative or otherwise downward-entailing/non-assertive environments according to language-specific constraints. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is using the ordinary affirmative equivalent in a context where the target requires a special polarity form. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to identify the licensor before translating the item. For example, ever is natural in ‘I haven’t ever seen it’ but restricted in simple affirmative ‘I ever saw it’. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Positive Polarity Items
A precise translation should classify the polarity item before choosing its target form. Positive polarity items prefer affirmative environments and can resist the scope of negation. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating a PPI as though it were an NPI or allowing negation to take the wrong scope. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to test how the item behaves under negation. For example, English some can prefer a reading outside negation in ‘I didn’t see someone’. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Any as an NPI
The same surface word can realise more than one polarity reading. English any commonly appears in negative and interrogative environments where it does not simply mean ‘every’ or ‘free choice.’ Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is choosing a universal or unrestricted target form in a licensed NPI context. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to identify whether any is existential under a licensor. For example, ‘I didn’t see any students’ means there was no student I saw. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Any as Free Choice
The target language may lexicalise the source contrast with completely different words. Any can also have a free-choice reading such as ‘choose any book,’ which differs from NPI any. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is treating free-choice any as negative polarity existential. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to test whether the sentence grants unrestricted choice across alternatives. For example, ‘You may take any book’ is closer to whichever book you choose. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Ever
The first task is to identify the licensing environment. Ever is strongly polarity-sensitive in many uses and can appear in questions, negatives, conditionals and superlatives. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating ever uniformly as ‘at any time’ even when the target uses a different construction. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to identify the licensing environment and temporal alternatives. For example, ‘Have you ever been there?’ asks whether there is any time in the relevant span. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Yet
This is a semantic-syntactic dependency rather than ordinary lexical selection. Yet often depends on negation, questions or incomplete time frames and interacts with aspect. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is using a plain temporal adverb that implies completion or affirmative occurrence. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to map temporal expectation plus polarity. For example, ‘She hasn’t arrived yet’ encodes non-arrival up to a reference point with expectation still open. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
At All
A precise translation should classify the polarity item before choosing its target form. At all is an emphatic polarity item that widens or strengthens the relevant scale/domain in licensed environments. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is dropping the emphasis or translating it as literal totality. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to identify what dimension is being minimised/widened. For example, ‘I don’t understand it at all’ strengthens zero understanding. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Either after Negation
The same surface word can realise more than one polarity reading. Additive either can function as an NPI counterpart of too/also under negation. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating negative either as disjunctive choice. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to test whether it means ‘also not’ rather than A-or-B. For example, ‘I don’t like it either’ adds the speaker to a negative proposition. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Lift a Finger and Idiomatic NPIs
The target language may lexicalise the source contrast with completely different words. Languages often contain idiomatic polarity expressions such as lift a finger, sleep a wink or budge an inch. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating the idiom literally into a target language that does not license the same minimiser. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to identify the scalar-minimiser function. For example, ‘He didn’t lift a finger’ means he made no effort. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Negation as a Licensor
The first task is to identify the licensing environment. Ordinary sentential negation is a central licensing environment for many NPIs. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is placing the NPI outside the semantic scope of negation during target restructuring. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to bracket NOT[…NPI…]. For example, ‘I didn’t ever call’ differs from an interpretation where ever escapes negation. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Questions
This is a semantic-syntactic dependency rather than ordinary lexical selection. Questions can license many NPIs even though they are not semantically negative in the ordinary sense. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is adding target negation merely because an NPI is present. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to recognise interrogative licensing separately from negative meaning. For example, ‘Did anyone call?’ contains an NPI-like any expression without asserting nobody called. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Conditionals
A precise translation should classify the polarity item before choosing its target form. If-clauses and other conditional environments commonly license polarity-sensitive items. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating the NPI as though the condition were a negative clause. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to identify the non-assertive/hypothetical environment. For example, ‘If anyone calls, tell me’ does not imply nobody will call. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Before Clauses
The same surface word can realise more than one polarity reading. Before can license NPIs in some languages and contexts because of its semantic relation to an unrealised/later event. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is inserting explicit negation into the target. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to analyse the temporal relation and licensing independently. For example, ‘Leave before anyone notices’ contains no overt not. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Without
The target language may lexicalise the source contrast with completely different words. Without-phrases create a negative-like semantic environment that can license NPIs. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is doubling the negativity by adding an extra not in the target. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to treat without as the licensor. For example, ‘She left without saying anything’ is not ‘without not saying’. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Comparatives
The first task is to identify the licensing environment. Comparative clauses can license certain NPIs in English and other languages under particular analyses. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating the item as a normal affirmative indefinite without checking target grammar. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to identify the comparative operator as part of the licensing environment. For example, ‘better than I ever expected’ permits ever. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Superlatives
This is a semantic-syntactic dependency rather than ordinary lexical selection. Superlatives frequently occur with ever, as in ‘the best I have ever seen.’ Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is assuming ever requires an overt negative and mistranslating it. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to recognise the superlative environment. For example, ever ranges over relevant experience times. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Only
A precise translation should classify the polarity item before choosing its target form. Only creates a downward-entailing or exclusion-related environment that can license some polarity items. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating the NPI while ignoring only’s scope. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to bracket ONLY and the alternatives it excludes. For example, ‘Only John ever calls’ differs from simple ‘John calls’. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Few and Minimising Quantifiers
The same surface word can realise more than one polarity reading. Quantifiers such as few can license NPIs because they create downward-entailing environments. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is turning few into absolute negation. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to preserve the small-but-not-zero quantity while keeping target polarity licensing natural. For example, ‘Few people have ever seen it’ does not mean nobody has. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Every and Licensing Complexity
The target language may lexicalise the source contrast with completely different words. Certain NPIs can appear in restricted parts of universally quantified structures depending on scope and semantic direction. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is applying a simple ‘negative licenses NPI’ rule and missing quantifier structure. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to test the exact position of the item relative to the quantifier. For example, licensing under every-type determiners can be asymmetric between restrictor and scope. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Downward Entailment
The first task is to identify the licensing environment. A major semantic account connects NPI licensing with downward-entailing contexts, where inference runs from sets to subsets. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating the formal concept as a target word rather than using it to diagnose grammar. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to test subset inference as a licensing diagnostic. For example, from ‘no students came’ one can infer ‘no tall students came’. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Nonveridicality
This is a semantic-syntactic dependency rather than ordinary lexical selection. Other accounts use nonveridicality or related notions to capture questions, modals, conditionals and environments that do not commit to the truth of the embedded proposition. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is assuming all licensed contexts are semantically negative. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to ask whether the context commits to the proposition being true. For example, a question or conditional can license polarity without asserting falsity. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Licensing Strength
A precise translation should classify the polarity item before choosing its target form. Not all NPIs behave alike: some are weak and appear in a broad range of environments, while strong NPIs need more strongly negative or anti-additive contexts. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is assuming one licensing rule applies to any, ever, at all and idiomatic minimisers identically. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to classify the item by its distribution. For example, an emphatic minimiser can be more restricted than ordinary any. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Cross-Linguistic Variation
The same surface word can realise more than one polarity reading. Languages divide the polarity domain differently, with dedicated indefinites, particles, minimisers, negative concord items and free-choice forms. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is mapping English any onto one target word everywhere. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to identify the target system’s distinct polarity items. For example, a target language may use separate words for NPI existential and free-choice universal readings. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
NPI vs Negative Concord Item
The target language may lexicalise the source contrast with completely different words. Negative concord items can co-occur with negation and contribute a single semantic negation, while NPIs need licensing but are not necessarily negative concord morphology. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is treating all ‘negative-looking’ words as double negation. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to check the source language’s concord system. For example, a n-word can be licensed by negation without producing two semantic negations. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Polarity and Scope
The first task is to identify the licensing environment. The position of negation, quantifiers and polarity items determines interpretation. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is moving the item during target restructuring and changing scope. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to draw operator scope before translation. For example, NOT[see any] differs from a PPI reading outside negation. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Polarity and Modality
This is a semantic-syntactic dependency rather than ordinary lexical selection. Modal verbs can license free-choice or nonveridical readings and interact with any-like items. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating an item as NPI existential when the modal creates free choice. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to test whether all alternatives are permitted/possible. For example, ‘You can choose any seat’ differs from ‘You can’t choose any seat’. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Polarity and Aspect
A precise translation should classify the polarity item before choosing its target form. Items such as yet and ever interact with perfect and progressive aspect. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is translating polarity correctly but losing temporal/aspectual expectation. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to map polarity and aspect separately. For example, ‘hasn’t finished yet’ differs from ‘never finished’. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Polarity and Register
The same surface word can realise more than one polarity reading. Some polarity items are emphatic, colloquial, formal or idiomatic. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is normalising a vivid minimiser into a bland indefinite or over-dramatising a neutral any. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to match target register after licensing is correct. For example, ‘not a damn thing’ and ‘nothing at all’ differ in register despite similar polarity force. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Translation into English
The target language may lexicalise the source contrast with completely different words. A source may use a dedicated polarity indefinite where English selects any, ever, yet, at all, no one or an idiom according to environment. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is choosing one English item from a dictionary without testing acceptability in the target clause. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to generate several target candidates and test licensing. For example, the correct equivalent can change between negative, interrogative and free-choice contexts. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
AI Any/Every Confusion
The first task is to identify the licensing environment. Models can translate a free-choice form as existential any or a negative-polarity form as universal every. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is changing quantification and truth conditions. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to classify NPI vs free-choice before generation. For example, ‘anyone may enter’ differs radically from ‘did anyone enter?’. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
AI Polarity Repair
This is a semantic-syntactic dependency rather than ordinary lexical selection. Models may ‘fix’ an unusual source construction by adding or removing negation to make a target polarity item grammatical. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is changing the proposition to rescue target syntax. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to choose a different target polarity expression rather than altering source polarity. For example, if target any is unlicensed, replace the item—not the truth conditions. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Back-Translation Licensing Check
A precise translation should classify the polarity item before choosing its target form. A final check should identify both the polarity item and its licensor in the target. Licensing is sensitive to scope and construction, so a good translation checks the operator that makes the item possible rather than merely matching dictionary entries.
The main translation risk is checking only whether a similar word appears. Because target polarity systems differ, literal substitution can create an ungrammatical sentence or, worse, a grammatical sentence with different quantification. The translator should preserve proposition polarity first, then repair lexical choice.
A practical diagnostic is to write ITEM + LICENSOR + SCOPE + READING. For example, the target succeeds when the same existential/free-choice/emphatic reading is licensed naturally. If the obvious target cognate or indefinite is not licensed, choose another natural target construction instead of changing negation, question force or modality.
Worked Examples
Worked Example 1: NPI Any under Negation
Source problem: Source: ‘I didn’t see any students.’
Diagnosis: Any is existential under negation; the sentence denies the existence of a seen student.
Target strategy: Use the target’s ordinary negative indefinite/concord strategy.
Verification: Do not translate any as every.
Worked Example 2: Free-Choice Any
Source problem: Source: ‘You may choose any seat.’
Diagnosis: Any ranges freely over alternatives under permission.
Target strategy: Use a free-choice/unrestricted target form.
Verification: The target must mean every seat is an allowable choice, not that there exists some unspecified seat.
Worked Example 3: Ever in a Question
Source problem: Source: ‘Have you ever visited?’
Diagnosis: Ever is licensed by the interrogative and quantifies over relevant times.
Target strategy: Use the target’s question-compatible temporal indefinite.
Verification: No negation should be added.
Worked Example 4: At All
Source problem: Source: ‘I don’t understand it at all.’
Diagnosis: At all strengthens the lower end of the scale under negation.
Target strategy: Use an emphatic negative expression appropriate to target register.
Verification: Plain ‘I don’t understand it’ may lose emphasis; universal ‘all’ may be wrong.
Practical Diagnostic Workflow
Create a POLARITY CARD: ITEM; ITEM TYPE=NPI/PPI/FREE-CHOICE/CONCORD/IDIOM; LICENSOR; CLAUSE POLARITY; SCOPE; QUANTIFICATION; REGISTER. Then test the target candidate inside the completed target sentence, because polarity acceptability cannot be judged from the isolated word.
Run the affirmative/negative/question/conditional test. Move the item through these four environments and observe whether its reading changes. This quickly exposes items that look equivalent in a dictionary but belong to different polarity classes.
Verification Method
From the target, identify the polarity item, its licensor and its semantic reading. Check that no new negation, universality, free-choice reading or emphatic force has appeared. The target passes when the same alternatives and truth conditions are licensed naturally.
Recovery Method
If a target item is unlicensed, do not alter the source proposition. Replace the target polarity item or restructure the clause. If any/every or NPI/free-choice meaning has been confused, return to the operator and rebuild quantification before editing style.
Quality-Assurance Checklist
- Classify NPI/PPI/free-choice/concord.
- Identify the licensor.
- Check negation scope.
- Check questions and conditionals.
- Check comparatives/superlatives/only/few.
- Distinguish any existential from free-choice any.
- Preserve ever/yet temporal behavior.
- Preserve at-all/minimiser emphasis.
- Distinguish negative additive either from disjunction.
- Check cross-linguistic polarity inventories.
- Distinguish NPI from negative concord item.
- Preserve modality and quantifier scope.
- Do not add negation to license a target word.
- Audit AI any/every confusion.
- Audit AI polarity repair.
Frequently Asked Questions
What is a negative polarity item?
An expression whose distribution is restricted to negative or related licensing environments, such as English any/ever/at all in many uses.
Why can NPIs occur in questions?
Polarity licensing is broader than overt negation; interrogative and other non-assertive/downward-entailing environments can license many NPIs.
Is any always an NPI?
No. Any also has free-choice uses such as ‘choose any book.’
What is a positive polarity item?
An expression that prefers affirmative environments and can resist being interpreted inside negation.
What licenses ever?
Negation, questions, conditionals, superlatives and other constructions can license or support ever, depending on the use.
Is a negative concord word an NPI?
The categories can overlap historically/functionally, but negative concord items participate in a grammatical concord system and should not automatically be treated as ordinary NPIs.
Why can’t I translate polarity items word-for-word?
Languages divide existential, free-choice, negative, emphatic and concord meanings differently, so the appropriate target item changes with environment.
Why does AI make polarity errors?
A model can prefer a familiar target word and then modify negation or scope to make it fit, silently changing truth conditions.
Where This Fits in the eduKate Translation Architecture
Route proposition-level negation to the live Negation and Scope owner; quantifier logic to the Quantifiers and Scope owner; modal/free-choice environments to Modality and Hedging; Finnish/Estonian connegative morphology to the live Negative Auxiliaries owner.
Final Principle
A polarity item is licensed by its environment. Preserve the operator, scope and reading first; then choose the target word. Never change the proposition just to make a familiar polarity expression fit.