Is this about low-resource languages?

No. Resource availability concerns data and tools; cold here concerns patterns of reference.

Does English never omit participants?

No. English allows omission in some constructions and conversational contexts; the comparison is a tendency, not an absolute prohibition.

Can a translator always add a pronoun safely?

No. Adding one may require a choice the available context does not justify.

AI Glossary / Natural language and speech / Intermediate

Cold languages

Definition

Cold languages is a metaphorical label used in some linguistic discussions for languages with relatively sparse explicit reference, where listeners often infer omitted participants from context. It concerns referential density, not climate, emotional warmth or linguistic quality.

What it means

A sentence describes an event with participants: someone sends something to someone, for example. Languages differ in how often those participants must or tend to be expressed. When the context already establishes a participant, some constructions allow the corresponding pronoun or noun phrase to remain unspoken.

Jurafsky and Martin use cold for languages such as Chinese and Japanese that often leave substantial reference to inference. The contrasting label hot describes more explicit reference. These are broad descriptive tendencies in a particular discussion, not rigid boxes covering every utterance.

For AI, the practical issue is recovering who or what a message concerns. A translation into a language that requires an explicit subject may need to supply one, even when the source did not state it. Guessing can change responsibility, ownership or the identity of an actor.

How it works

A reader or listener combines grammar with discourse context to recover an omitted argument. Earlier mentions, the current topic, the event being discussed and the participants in the conversation can all help. The source wording may still leave more than one interpretation possible.

A language system therefore needs to distinguish an omitted reference from an absent piece of meaning that should never be invented. Document-level context can be useful, but more context does not guarantee certainty. Translation may require an explicit choice, a different phrasing or a clarification step.

Pro-drop describes the possibility of omitting pronouns. Referential density concerns how much reference is explicitly expressed in actual use. Languages permitting omission can differ considerably in frequency and conditions. Neither a single short sentence nor a raw pronoun count is enough to establish a comprehensive classification of a language or its speakers.

Three practical examples

A translation system must recover an unstated subject before producing an English sentence about who approved an order.
A conversation assistant tracks the active topic so an omitted object is interpreted as the product already under discussion.
A corpus researcher measures overt and omitted arguments under a stated annotation scheme rather than equating fewer words with less meaning.

Think of it this way

Think of colleagues sharing a whiteboard full of context. One can say “sent yesterday” because the relevant report and sender are already understood. A newcomer may need those details stated explicitly.

This analogy illustrates dependence on shared context. It does not make English fragments a complete model of another language's grammar, nor imply that an entire language is always less explicit.

Worked example

Separating omission from an unsupported guess

Use an invented English gloss solely to illustrate the reasoning, not as a claim about a particular Japanese construction: “Mina prepared the report. Sent yesterday.” The second fragment leaves its sender and object unstated.

Suppose a reviewed context record says Mina sent that report. A suitable complete rendering can make both references explicit: “Mina sent the report yesterday.” Now change the context record: Mina prepared it, but Omar was responsible for dispatch. The same short gloss no longer supports assuming Mina was the sender.

A teaching annotation might mark two possible argument positions in the second fragment and zero overt noun phrases there. That local count describes the fragment under this chosen scheme. It neither proves a universal density threshold nor resolves the references. The interpretation requires evidence about the discourse.

Illustrative case study

A fictional supplier translates short customer updates into a shared operations language. Several messages omit participants already established earlier in the conversation. A sentence-only pipeline supplies a subject from its most common pattern, occasionally assigning an action to the wrong team.

Reviewers compare the translations with the surrounding exchange and mark where the subject is supported, uncertain or genuinely unspecified. The team evaluates whether preserving conversational context reduces these particular errors, while keeping a review path for remaining ambiguity.

The lesson is to preserve the source's commitments. A fluent sentence is not adequate if it silently changes who did the work. The case is illustrative and does not report measured accuracy gains.

Common misconceptions and mistakes

Cold means unfriendly or unemotional. The term concerns explicit reference in this linguistic use.
These languages lack pronouns. The issue is their expression or omission in particular contexts.
Every omitted participant is obvious. Ambiguity can remain even with context.
Pro-drop languages all behave alike. Grammatical conditions and actual frequencies differ.
More explicit wording makes a language better. The labels do not rank languages or speakers.

Business use

This concept matters when translated updates feed task assignment, summaries or customer records. Test whether the system preserves actors, recipients and objects across sentence boundaries. Include examples where the obvious recent name is not the intended actor.

Keep context available where the workflow permits it, and record unresolved reference rather than inventing responsibility. In evaluation, distinguish grammatical fluency from reference accuracy. Product documentation will often be clearer using terms such as omitted arguments or context-dependent reference, with cold language retained only when explaining the source terminology.

Check your understanding

Why should a system not automatically choose the most recently mentioned person as the omitted subject?

Answer: Recency is only one clue. The discourse may establish another actor, or leave the actor uncertain. A grammatical output can still misrepresent the source.

Related terms

Learn next

Study zero-pronoun resolution and document-level translation. With a qualified bilingual reviewer, compare isolated sentences with their full conversations, noting which references become recoverable and which remain genuinely ambiguous.

Sources

Jurafsky and Martin, Machine Translation, section 13.1.3: https://web.stanford.edu/~jurafsky/slp3/13.pdf. The source's referential-density and cold/hot terminology.

Ri and colleagues, Zero-pronoun Data Augmentation for Japanese-to-English Translation: https://aclanthology.org/2021.wat-1.11/. Primary research on context and missing pronouns. The report example is an invented English gloss, not a language-specific grammatical example.

Previous: Cointegration Next: Cold start (recommender systems)