The English word ‘bank’ can mean a financial institution or the side of a river. Both dictionary meanings may be correct, yet choosing the wrong one changes an entire sentence. How should a researcher study the decision rather than simply praise whichever version sounds fluent?
The gold standard of translation studies is careful, reproducible analysis of how meaning travels across languages and contexts. This guide contains 22 chapters on semantics, grammar, cultural interpretation, machine translation, evaluation methods and ethical research, with published ISO standards kept distinct from draft revisions.
Quick information: what exactly is being evaluated?
| Object | Question | Avoid confusing it with |
|---|---|---|
| Translation output | Does the target text meet defined purpose and requirements? | A universally correct word-for-word replica |
| Translation service process | Were appropriate people, resources and review steps used? | Guaranteed perfect output |
| Machine translation | What did the system generate under the given conditions? | Human-verified accuracy |
| Human post-editing | What was corrected and to what specification? | Unreviewed raw output |
| Interpreting | How is meaning conveyed in real-time spoken or signed settings? | Written translation standards |
| Research sample | What evidence supports a limited pattern? | Universal performance claim |
Worked example: an error score depends on the rubric
| Fictional observation | Count | Invented penalty per occurrence | Total penalty |
|---|---|---|---|
| Major errors | 2 | 5 | 10 |
| Minor errors | 4 | 1 | 4 |
| Total | 6 | — | 14 |
This is a locally defined teaching exercise—not a scoring formula mandated by ISO 5060. Evaluation requires the reviewer to state what counts as an error, its severity, the sampling population and what score, if any, is acceptable.
Changed-context example: comprehension is not automatic causation
| Fictional reading test | Correct responses | Total participants | Observed share |
|---|---|---|---|
| Original text | 13 | 20 | 65% |
| Revised text | 18 | 20 | 90% |
| Difference | +5 | Same denominator | +25 percentage points |
Without a suitable design controlling reader groups and conditions, these numbers cannot prove the revised wording caused the entire difference. The scenario illustrates a method question, not a real research result.
Complete 22-chapter translation studies index
Meaning, grammar and cultural context
Chapters 1–6: Study how meaning moves between languages · Define source text, target text and situation · Distinguish lexical sense from context · Analyse syntax and relations, not only vocabulary · Separate literal resemblance from functional adequacy · Preserve register and speaker relationship.
Professional process and evaluation
Chapters 7–12: Study cultural references without reducing culture to stereotypes · Build a terminology glossary that can be checked · Choose a research question before building a corpus · Separate a translation memory from machine translation · Understand service standards without overstating conformity · Separate service process from output evaluation.
Tools, accessibility and research
Chapters 13–18: Worked exercise: score errors only under a defined rubric · A single critical error can outweigh many fluent sentences · Sample responsibly before generalising quality · Study evaluators as part of the measurement system · Human post-editing needs its own scope · Keep interpreting separate from written translation.
Ethics and transferred learning
Chapters 19–22: Localisation includes forms, layout and access · Consider power, consent and representation · Evaluate comprehension with appropriate study design · Build a reproducible translation-studies report.
- Study how meaning moves between languages
- Define source text, target text and situation
- Distinguish lexical sense from context
- Analyse syntax and relations, not only vocabulary
- Separate literal resemblance from functional adequacy
- Preserve register and speaker relationship
- Study cultural references without reducing culture to stereotypes
- Build a terminology glossary that can be checked
- Choose a research question before building a corpus
- Separate a translation memory from machine translation
- Understand service standards without overstating conformity
- Separate service process from output evaluation
- Worked exercise: score errors only under a defined rubric
- A single critical error can outweigh many fluent sentences
- Sample responsibly before generalising quality
- Study evaluators as part of the measurement system
- Human post-editing needs its own scope
- Keep interpreting separate from written translation
- Localisation includes forms, layout and access
- Consider power, consent and representation
- Evaluate comprehension with appropriate study design
- Build a reproducible translation-studies report
CHAPTER 1 OF 22 · All chapters
1. Study how meaning moves between languages
Translation studies asks how texts and meanings are interpreted, transformed and received across languages, cultures and institutions. The subject examines decisions and consequences, not simply whether someone can replace words in one language with words in another. The gold standard is a defensible explanation of translation choices in their contexts, supported by comparative evidence and clearly stated limitations.
An effective translator may deliver a fluent product, while a researcher might ask why two equally fluent translations differ in tone, omission or intended readership. Translation concerns practice, and Linguistics provides related tools for analysing language. The present guide focuses on questions, methods and evaluating the evidence behind a translation claim.
CHAPTER 2 OF 22 · All chapters
2. Define source text, target text and situation
A text does not exist without a communicative situation. A safety warning, poem, marketing slogan and legal notice may require different decisions even if they contain similar words. Research should identify who produced the source, who will receive the translation, what purpose it serves and which constraints apply.
Write a source-and-target brief before comparison. Include language varieties, field, date, medium, intended readership and known publication circumstances. Do not claim that all contexts have one universal best translation. An apparently small change in audience or genre can change which options are appropriate.
CHAPTER 3 OF 22 · All chapters
3. Distinguish lexical sense from context
Many words have multiple related or unrelated meanings. The English word ‘bank’ can refer to a financial institution or the side of a river. A dictionary lookup without surrounding context can select the wrong sense while producing an otherwise grammatically plausible sentence.
Ask what the source phrase actually refers to and what contextual clues support the interpretation. A fictional instruction ‘meet at the bank’ cannot be safely translated without knowing the intended location. This is an issue of semantic selection, not a stylistic preference. Semantics explores how meaning changes with lexical and sentence-level context.
CHAPTER 4 OF 22 · All chapters
4. Analyse syntax and relations, not only vocabulary
Word order, agreement, tense, aspect, modality and negation carry meaning. A translation that preserves every individual noun but changes ‘must not open’ to ‘must open’ reverses a critical instruction. Grammatical form can vary across languages without permitting the underlying obligation to change.
Map important agents, actions, conditions and time relations before evaluating surface wording. For a technical warning, check who is prohibited from doing what, when the condition holds and whether an exception exists. Syntax provides background on grammatical structures. A word-for-word resemblance cannot substitute for preserving the intended relationships.
CHAPTER 5 OF 22 · All chapters
5. Separate literal resemblance from functional adequacy
A translation may closely resemble source wording but fail its communication purpose, or differ structurally while preserving the necessary meaning. A humourous pun, warranty statement and machine label each raise different trade-offs among lexical form, clarity, cultural reference and functional outcome.
Evaluate the text against its stated purpose and explicit constraints, rather than rewarding literalness automatically. Describe what would count as unacceptable loss. This is particularly important when a decision about style is incorrectly reported as a factual error, or when a serious change in meaning is dismissed as harmless localisation.
CHAPTER 6 OF 22 · All chapters
6. Preserve register and speaker relationship
Register concerns how language varies with situation and social relationship. A formal complaint, friendly invitation and medical explanation differ in expected tone and terminology. Choosing a slang expression where a polite official formulation is needed can change how recipients understand the speaker’s relationship to them.
Study the target audience and genre conventions without stereotyping every speaker of a language. Keep the level of formality, intended authority and emotional tone proportionate to the source and task. When alternatives are plausible, record the reasons rather than describing one stylistic preference as universally correct.
CHAPTER 7 OF 22 · All chapters
7. Study cultural references without reducing culture to stereotypes
Some words, foods, rituals, institutions and historical references have no simple one-word counterpart. A translation can preserve the original expression, explain it, adapt a reference where justified or use another strategy. Each has effects on what the reader learns and whose cultural knowledge is centred.
Research the actual referent and consult competent people where appropriate. Avoid erasing a community’s own name for an institution merely for an exotic or tourist-friendly effect. Cultural Heritage connects terminology with provenance and representation. A translation study should describe the cultural decision, not assume all members of a community interpret it identically.
CHAPTER 8 OF 22 · All chapters
8. Build a terminology glossary that can be checked
Technical, educational and legal texts depend on consistent treatment of important terms. A glossary should identify the source expression, intended concept, target equivalent, context, source authority and any approved alternatives. A loose list of word pairs may conceal differences in meaning.
Version the terminology and update it when the field or document changes. A term used in two unrelated senses may need separate entries. Do not claim that one translation is legally authoritative simply because it appears in a commercial dictionary. For regulated material, qualified reviewers and applicable institutions remain essential.
CHAPTER 9 OF 22 · All chapters
9. Choose a research question before building a corpus
A corpus is a systematically assembled set of texts used for linguistic or translation analysis. A researcher might study terminology stability, how requests are expressed in translated instructions or whether a translator tends to expand explanations. Data collection should serve a specified question rather than produce thousands of sentences with no meaningful sampling rationale.
Record source and target provenance, languages, genres, time periods and selection criteria. Parallel and comparable corpora answer different questions; missing metadata can invalidate a pattern. Copyright and privacy obligations apply to collected materials. Evidence-Based Thinking helps distinguish a pattern in a dataset from a universal law about translators.
CHAPTER 10 OF 22 · All chapters
10. Separate a translation memory from machine translation
A translation memory stores and retrieves aligned previously translated segments. Machine translation generates target-language output from a computational system. The tools can be used together, but their functions and error patterns are distinct. A high text similarity match is not proof that its translation fits a changed context.
Suppose a prior segment concerned a river bank and a new one concerns finance. Reusing an apparently close segment can introduce a wrong referent. Check meaning, surrounding text and current domain, not just a software percentage. Treat machine output as material requiring appropriate review instead of a factual authority on its own.
CHAPTER 11 OF 22 · All chapters
11. Understand service standards without overstating conformity
Professional translation workflows may require agreed processes, resources, competences and revision. ISO 17100:2015 describes requirements for translation services within its scope and has an amendment published in 2017. It does not apply to interpreting or to raw machine-translation output plus post-editing under that standard’s stated scope.
A new edition of ISO 17100 is under development as of October 2026, which does not automatically replace the published text. Quoting a standard’s name does not show a specific organisation conforms to it. Identify the precise activity and version before using a compliance claim.
CHAPTER 12 OF 22 · All chapters
12. Separate service process from output evaluation
A sound service process and a good translation output are related but not interchangeable. A project may follow documented steps and still contain an error; a single fluent sample may be produced without a robust repeatable process. The research question determines whether process evidence, output quality or user comprehension is the outcome.
The published ISO 5060:2024 provides guidance on human evaluation of human, post-edited machine and unedited machine translation output, including error types, penalty points and sampling. Its scope differs from ISO 17100. Do not merge both into one invented certification test.
CHAPTER 13 OF 22 · All chapters
13. Worked exercise: score errors only under a defined rubric
Imagine forty fictional translated segments reviewed under a locally defined teaching rubric. The reviewer identifies two major and four minor errors. Suppose the rubric assigns five points to a major error and one point to a minor error. The illustrative penalty is 2 × 5 + 4 × 1 = 14 points.
This is NOT an ISO 5060-prescribed calculation or an official quality rating. Without defining the sampling process, severity criteria, denominators and threshold, the number 14 conveys little. A serious review publishes its rubric and explains which errors were counted, whether multiple errors occur in one segment and how disagreement was resolved.
CHAPTER 14 OF 22 · All chapters
14. A single critical error can outweigh many fluent sentences
A translation with almost every sentence correct might still contain a critical reversal of a safety instruction. Counting successful sentences and averaging all deviations can hide consequences that require separate escalation. Some requirements are non-compensatory because one failure cannot be offset by smooth wording elsewhere.
A review should distinguish comprehension, severity and intended use. In educational contexts, a minor stylistic difference might be acceptable; in medical, legal or industrial use, a material meaning change can require qualified intervention. The purpose of risk-aware evaluation is accurate judgment, not treating every text as equally high stakes.
CHAPTER 15 OF 22 · All chapters
15. Sample responsibly before generalising quality
Reviewing twenty conveniently selected pages cannot establish quality throughout a much larger multilingual corpus without suitable design. Easy, repetitive segments may overstate performance while difficult clauses and rare terms remain untested. Sampling should represent the content and claims being made.
Document the population, sampling method, languages, document types and changes since the sampled version. ISO 5060 discusses sampling, but no simple sample size works for every purpose. Distinguish an observed error count from a confidently estimated population error rate. Statistics supports sound interpretation.
CHAPTER 16 OF 22 · All chapters
16. Study evaluators as part of the measurement system
Two qualified reviewers may disagree about nuance, severity or domain relevance. That variation does not mean evaluation is impossible, but it requires clear instructions and evidence. A guideline that says ‘mark bad translation’ without criteria cannot support a repeatable quality comparison.
Use calibration examples, documented disagreements and escalation to relevant subject experts. Compare reviewers’ interpretations and revise unclear categories. A stable rubric supports judgment but should not erase legitimate alternative renderings where several solutions satisfy the purpose.
CHAPTER 17 OF 22 · All chapters
17. Human post-editing needs its own scope
Post-editing machine-generated translation involves assessment and correction of output, potentially under different workflow requirements from translation produced directly by a human. The degree of revision needed depends on purpose, risk and agreed specifications. Fluency alone cannot establish that the output preserved source meaning.
ISO 18587:2017 covers full human post-editing of machine translation within its scope. A proposed revision, ISO/DIS 18587, was in development during 2026; a draft is not a published replacement. When describing requirements, distinguish the issued edition from ongoing standards work.
CHAPTER 18 OF 22 · All chapters
18. Keep interpreting separate from written translation
Interpreting enables spoken or signed communication in real time or near real time and creates its own demands on working memory, turn taking, setting and ethics. Written translation allows different review and revision possibilities. A general theory may inform both, but their operational standards and research designs differ.
Do not apply ISO 17100 or ISO 5060 to interpreting as though the standards covered it. Identify the actual modality, participants and purpose. In healthcare, legal or emergency contexts, appropriate qualified interpreters and applicable rules may be required; this educational article does not teach the user to replace those professionals.
CHAPTER 19 OF 22 · All chapters
19. Localisation includes forms, layout and access
A translation may preserve literal words while leaving the target audience unable to use the product because dates, currency, units, reading order or interface constraints were not adapted appropriately. Localisation can involve broader design and cultural decisions, but these should not silently alter substantive obligations.
Test the translated interface with actual target-language content at realistic text size. Check expansion, line breaking, controls, numeric formats and the meaning of linked help. Typography and Technical Communication explain why readable delivery is part of the end user’s outcome rather than merely decoration.
CHAPTER 20 OF 22 · All chapters
20. Consider power, consent and representation
Translation can determine whose explanation becomes visible in courts, schools, museums and public services. Decisions about what to simplify or omit can alter the authority of speakers and communities. Research should consider copyright, confidentiality, consent and the possibility that people affected by a translation were excluded from its evaluation.
When collecting interviews, texts or sensitive translation data, protect people and document permissions. Treat cultural or political disagreements with accurate attribution and fair context. Avoid assuming a translation automatically represents an entire community merely because one fluent speaker endorsed it.
CHAPTER 21 OF 22 · All chapters
21. Evaluate comprehension with appropriate study design
A translator may produce a faithful target sentence while intended readers misunderstand its practical purpose. A usability or comprehension study can reveal this, but the test must use representative readers and carefully defined tasks. Familiarity, education and interface design may all affect results.
Suppose thirteen of twenty fictional readers correctly identify an instruction before a revision and eighteen of twenty do afterward. Rates are 65% and 90%, a 25-percentage-point difference. This is descriptive, not automatically causal: reader groups and other conditions may differ. Use ethical study methods and explain what the results can and cannot prove.
CHAPTER 22 OF 22 · All chapters
22. Build a reproducible translation-studies report
A sound research report identifies source and target texts, purpose, corpus selection, translation versions, reviewer competence, scoring criteria, uncertainty and relevant ethical considerations. It makes alternative explanations visible, rather than advertising one model or method as universally superior.
Teaching can begin with ‘bank’ and the importance of context, then progress through negation, genre and data interpretation. The immutable Secondary 1 Mathematics Tutor Clementi Small Groups Tutorials page models the general habit: define the problem, justify each step and test an altered case. It is a learning reference, not a translation standard.
Translation studies evidence audit
| Evaluation item | Documented evidence | Common mistaken claim |
|---|---|---|
| Source meaning | Context, purpose, variety and version | One dictionary meaning fits every case |
| Target adequacy | Meaning, register, genre and critical constraints | Literal resemblance proves correctness |
| Process | Applicable standard, reviewer and workflow | Named standard certifies every output |
| Output evaluation | Sampling, error definitions and severity | Invented score treated as official ISO formula |
| Machine output | System version, human checks and limitations | Fluency means factual accuracy |
| Research result | Method, comparison, uncertainties and rights | Small sample taken as universal claim |
Frequently asked questions
What does translation studies investigate?
How translation decisions, meanings, processes and social contexts work, and how they can be researched or evaluated.
Is the most literal translation always best?
No. Audience, genre, purpose and semantic accuracy can require different structures.
Does ISO 17100 cover interpreting?
No. Its stated scope concerns specified translation services and excludes interpreting.
Does ISO 5060 prescribe the fictional error score in this guide?
No. The worked numbers use an explicitly invented teaching rubric.
Can machine translation be used without review for high-stakes tasks?
Its suitability depends on risk, requirements and competent review; fluent output alone does not establish accuracy.
Is a draft ISO revision already binding as the published standard?
No. Distinguish the currently issued standard from a draft under development.
What makes a translation evaluation reproducible?
Defined text population, versions, evaluator criteria, sampling, error rubric and transparent interpretation.
International standards and status in October 2026
The published ISO 17100:2015 concerns translation service processes and has a 2017 amendment; a revised edition remained under development as of October 2026. ISO 5060:2024 offers guidance for human evaluation of human, post-edited and unedited machine translation outputs, including evaluator competences and sampling. ISO 18587:2017 concerns full human post-editing of machine-translation output; an ISO/DIS 18587 revision was in development during 2026. None of these is claimed to certify this article. Its scores, sample texts and reader study are fictional.
Connected Gold Standard subjects
Translation · Linguistics · Semantics · Syntax · Literacy · Information Literacy · Cultural Heritage · Technical Communication · Technical Writing · Typography · Education · Evidence-Based Thinking · Statistics · Documentation