AI Glossary / Natural language and speech / Intermediate
Adequacy
Definition
Adequacy is the degree to which a translation preserves the meaning of its source text. It asks whether the translated message conveys the same information, relationships, qualifications, and intent. A translation can sound natural while changing a crucial fact, so adequacy is evaluated separately from fluency, which concerns the quality of the target-language expression.
What it means
In everyday language, “adequate” can mean merely good enough. In machine-translation evaluation, adequacy names a specific quality dimension: fidelity to the source. It does not impose one universal pass mark or scoring scale.
A faithful translation need not reproduce the source word for word. Languages differ in grammar, word order, and how they express ideas. The important question is whether the target reader receives the intended message, including details such as who did what, whether something is permitted, and which conditions or exceptions apply.
How it works
An evaluation presents the source and candidate translation to a suitable reviewer, often someone competent in both languages. The reviewer considers omissions, additions, mistranslations, and altered relationships. A defined rubric can produce a numerical judgment, a comparison between candidates, or an annotated set of errors.
Where only target-language reviewers are available, a trusted reference translation can provide a comparison, although that introduces dependence on the reference's quality and interpretation. Multiple valid translations may convey the same meaning with different wording.
Reviewers need clear instructions and examples distinguishing meaning errors from awkward phrasing. Source ambiguity and missing context should be recorded rather than silently resolved by guessing. Automated metrics can assist evaluation, but a high similarity score alone does not establish that every consequential detail has been preserved.
Three practical examples
1. Product information: Translating “available only in selected sizes” as “available in all sizes” is fluent but changes the offer. Adequacy review focuses on the scope of availability.
2. Customer support: A response saying a refund is “being reviewed” should not become a statement that the refund “has been approved.” The processing stage is part of the meaning.
3. Meeting communication: A proposal to move a meeting must remain a proposal. Translating it as a confirmed rescheduling changes the participants' understanding of what was agreed.
Think of it this way
Think of moving the contents of a carefully packed box into a different container. The shape and arrangement may change, but the important contents should arrive intact, without missing items or invented extras. The analogy has limits: meaning is context-sensitive, and translation sometimes requires making grammatical distinctions that the source leaves implicit. It is not a mechanical transfer of isolated words.
Worked example
Consider an invented source instruction whose intended English meaning is: “Deliver six replacement filters to the north entrance on Tuesday, but only after the supervisor confirms.” The following English renderings stand in for target-language translations so the meaning comparison is visible without requiring another language.
Candidate A says: “Deliver six replacement filters to the north entrance on Tuesday.” Candidate B says: “Tuesday, bring six replacement filters to the north entrance after supervisor confirmation only.” Candidate B is less polished, but it retains the condition that Candidate A omits.
For this teaching exercise, create six checks: delivery action, quantity six, replacement filters, north entrance, Tuesday, and confirmation as a prerequisite. Candidate A satisfies five of six; Candidate B satisfies all six. The resulting checklist fractions are 83.3 percent and 100 percent.
These are not standardized adequacy scores. The missing prerequisite could be more consequential than several other details combined. A real rubric should assess severity and acceptable wording, rather than treating every checklist item as equally important. The exercise shows why fluent wording and a high fraction of preserved details can still conceal a material meaning error.
Illustrative case study
Fictional case: Willow Supplies uses machine translation for draft delivery messages. Reviewers initially favor the smoothest wording, but one draft removes the phrase “after confirmation.”
The team separates meaning review from style review and asks bilingual reviewers to flag changed quantities, destinations, timing, and conditions. Editors improve awkward language only after preserving those details. This fictional workflow illustrates how adequacy can guide review; it makes no claim about actual company results or a universal scoring threshold.
Common misconceptions and mistakes
Do not equate adequacy with literalness. A literal translation can be misleading when an expression has an idiomatic meaning. Conversely, a paraphrase can be fully faithful.
Do not use fluency as a substitute for comparing the source. Natural wording can make a meaning error harder to notice. Also distinguish source fidelity from factual correctness in the world: a translation may faithfully reproduce a mistaken source statement. Correcting that statement without disclosure is an editorial intervention, not simply more faithful translation.
Business use
Adequacy gives business teams a concrete way to review what a translation communicates. It is useful when a changed date, condition, negation, or responsibility would affect a customer's next action.
An evaluation plan should specify the language pair, content type, reviewer qualifications, available context, and error-severity rules. Report examples of important failures alongside aggregate ratings. Averages can hide recurring omissions in a small but consequential subset of messages, and scores from different rubrics are not automatically comparable.
People also ask
Can a translation be adequate but not fluent?
Yes. Awkward grammar can preserve the intended message, although readability may still need improvement.
Can there be several adequate translations?
Yes. Different wording and sentence structure can convey the same source meaning.
Does a reference mismatch prove inadequacy?
No. The candidate may express the same meaning differently. Review the source, context, and nature of the difference rather than relying only on surface overlap.
Check your understanding
The source says “The service may restart tonight,” but the translation says “The service will restart tonight.” What adequacy issue arises?
Answer: The translation strengthens a possibility into a commitment. Even if the sentence is fluent, it changes the degree of certainty communicated by the source.
Related terms
- Machine translation
- Fluency
- Human evaluation
- Post-editing
- Semantic similarity
- fidelity
- reference translation
- mistranslation
Learn next
Compare adequacy and fluency on paired translation examples. Then explore error-based evaluation, paying attention to omissions, negation, modality, and whether a scoring rubric distinguishes minor wording problems from consequential meaning changes.
More in Natural language and speech
- Adaptive softmax - Adaptive softmax is a hierarchical output-layer method for efficiently modeling large, unevenly used vocabularies.
- Add gate - An add gate is the learned control in an LSTM network that regulates how much newly proposed information enters its memory state.
- Add-k - Add-k is a probability-smoothing method that adds the same positive amount k to every possible outcome's observed count before normalizing.
- Adjacency pairs - Adjacency pairs are linked conversational actions in which a first contribution creates an expectation for a particular kind of response.
- Adjective - An adjective is a word class whose members typically express properties of nouns or the things those nouns refer to. In “a quiet room,” quiet modifies room.
- Adverb - An adverb is a word class whose members commonly modify verbs, adjectives, other adverbs, or a wider proposition.
Sources
Source: Jurafsky and Martin, Speech and Language Processing, Chapter 13, Section 13.6, read for adequacy, fluency, human rating, and the role of automatic evaluation: https://web.stanford.edu/~jurafsky/slp3/13.pdf . The supplied full-book retrieval was unavailable; the accessible author-hosted chapter was used. The six-item checklist is an original teaching device, not a published metric or validated evaluation protocol.
Previous: Address space Next: Adjacency pairs