Translation
in a Hurry

The Whole Thing in One Page

Translation is often pictured as removal work. Strip away the foreign words, carry the meaning across, put it down in English, and try not to break anything. That picture fails because meaning is not an object inside a sentence. It is a pattern made from reference, implication, tone, rhythm, social distance, genre, shared knowledge and expectation. Another language divides that pattern differently. The translator cannot move it intact. The translator has to rebuild it.

Translation is therefore a hierarchy of obligations. A legal translator may protect defined terms and logical scope. A subtitler must fit speech into limited time and space. A poet may preserve metre at the cost of syntax, or syntax at the cost of rhyme. An interpreter cannot pause a negotiation for three days of research. A localiser may need a software command to remain accurate, familiar and short enough for a button. The good solution depends on the job.

Words still matter, but a version can match individual words while changing what the speaker is doing. A polite refusal may become a blunt no. A joke may become a report that somebody joked. A marked distinction between familiar and formal address may disappear in English. A smooth sentence can conceal an omitted warning. Fidelity is not closeness at one level. It is justified attention to the levels that carry the text's force.

The problem is old. The Rosetta Stone presents a decree through three related inscriptions, not a row-by-row dictionary. Translators of sacred texts argued over wording, authority and interpretation. Buddhist translation in China and Greek-to-Arabic scholarship under Abbasid patronage were collaborative enterprises supported by revisers, specialists and institutions. Print enlarged audiences and raised the stakes. Empire, law, education and nation-building made translation an instrument of rule as well as exchange.

Practice exposes the machinery. Translators work from briefs, terminology, references, earlier versions and queries. Editors, reviewers, publishers, courts, hospitals, software teams and clients alter what reaches the reader. Interpreters manage speech under pressure. Subtitles, dubbing and localisation add constraints that are not linguistic alone. A translation may be written by one person, but it is often made by a system.

Machines have entered that system, not abolished it. Statistical methods learnt patterns from bilingual data. Neural systems and language models became better at producing fluent sequences and using context. Their output still varies by language, domain, genre and test. They can omit, over-explain, invent or choose the wrong sense with composure. An automatic score can compare systems under stated conditions, but it cannot decide whether a consent form, poem or political insult has survived.

Something is always selected. What survives may be the factual claim, social action, rhythm, ambiguity, shock or ease of reading. What cannot survive locally may be compensated elsewhere, explained, borrowed, left strange or surrendered. Translation can also add access, new associations, sharper choices and an afterlife the source never had. Loss is real, but it is not the whole movement.

Once a version circulates, its choices begin to look like properties of the original. It can fix a character's voice, establish a sacred phrase, standardise a technical term or make a foreign writer sound locally contemporary. Later translators then work under that inheritance. The crossing changes what arrives and what future travellers expect to find. That is the book.

Why You Should Care

A dosage instruction, a witness statement and a line of comedy can all fail for the same reason: the words arrive, but the action does not. A patient may hear a frequency where the source gave a maximum. A court may hear certainty where a witness expressed doubt. An audience may understand the facts of a joke and feel nothing. Translation matters at the point where language becomes consequence.

Most people meet it constantly while noticing it rarely. They read international news, use imported software, watch subtitled drama, follow product instructions, quote sacred texts, compare philosophical ideas and buy books written in languages they cannot read. Even people who speak only one language inherit translated concepts and phrases. Much of what looks like direct contact with another place or period is contact with somebody's reconstruction of it.

Translation is also a laboratory for language itself. Put two versions beside each other and hidden structures appear: one language must mark how evidence was obtained; another marks social rank in pronouns; another lets the actor remain unstated. None is a code for a separate universe, but each makes some distinctions cheap and others laborious. Translation shows where grammar ends, context begins and culture supplies what no word says. It therefore offers more than access to foreign writing. It shows how much of ordinary understanding depends on expectations that speakers rarely notice until another language refuses to share them.

Learning how that reconstruction works changes reading. A translated novel stops being a transparent pane and becomes a double achievement. You can ask why a sentence sounds spare, old, comic or formal, and whether that quality belongs to the source, the translator, the publisher's house style or all three. Competing translations cease to look like failed attempts to reach one hidden answer. They become arguments about which features deserve priority.

It also improves judgement outside literature. Public institutions cannot treat language access as a decorative service. In medicine, law, asylum, policing and social care, interpreters mediate questions, uncertainty, consent and credibility. The relevant standard is not whether the conversation sounded fluent. It is whether the people involved could perform the required act with a defensible degree of accuracy, independence and understanding. Qualifications, preparation, role boundaries and review matter because the costs are unevenly distributed.

Translation also exposes power. Languages do not meet on equal terms. Some have large publishing markets, abundant terminology, official status and vast digital corpora. Others have fewer trained specialists, less commercial demand and weak representation in the data used to build machine systems. Institutions decide which books receive translators, which voices receive interpreters, which names are respelt and which forms of difference are edited away. Refusing translation can exclude. Translating can also classify, domesticate, censor or recruit.

Then there is artificial intelligence. Current systems can produce useful drafts and strong finished-looking text at extraordinary speed. That combination is valuable and dangerous. Weak output used to announce itself through broken grammar. A fluent error recruits the reader into trusting it. The sensible question is neither whether machines translate nor whether humans are uniquely gifted. It is which system performs this task, for these languages, in this domain, with what evidence, oversight, confidentiality and consequence.

This book will not teach you a language or turn you into a professional translator. It will give you the model needed to commission, compare and read translations without demanding the impossible. You will learn what a translator can preserve, why one success may require another sacrifice, how history built the institutions around the work, where machines help, and where confidence outruns proof. The reward is not suspicion of every translated sentence. It is a better kind of trust: alert to choices, proportionate to stakes and capable of asking what made the crossing.

The Core Ideas

Translation becomes clearer when treated as a sequence of constrained decisions, not a hunt for replacement words. Survival is never measured at one level alone. A feature can disappear locally and reappear through compensation, explanation, performance or a later version. The model must therefore explain what is chosen, what conflicts, who shapes the result and how a version changes the life of the text it carries.

A Translation Is a Hierarchy of Obligations

A sentence can carry facts, rhythm, politeness, ambiguity, technical precision, humour, historical flavour and a recognisable voice at the same time. Another language may offer no arrangement that preserves all of them together. The translator must rank which features are load-bearing for this text and this audience. The ranking may be explicit, habitual or imposed by the medium and institution. It governs the work even when nobody writes it down.

Consider a familiar problem. Many languages mark a distinction between intimate or familiar address and more formal address in their ordinary second-person pronouns. Contemporary English usually gives both as you. A translator can leave the distinction implicit, add a name or title, alter the surrounding register, or explain the change elsewhere. Each solution saves something and interferes with something else. A neat one-word replacement does not exist because the target language has organised the social relation differently.

The same conflict appears across genres. A contract may require the same defined term every time, even where repetition sounds clumsy. A novelist may vary a repeated word because the source variation matters more than terminological consistency. A subtitler has to preserve enough of a character's action and tone while cutting material that cannot fit the available reading time. A poet may decide that a line's rhythm performs the emotion more faithfully than its surface syntax. None of these translators is exempt from accuracy. They are answerable to different orders of need.

This is why the instruction “be faithful” is incomplete. Faithful to what? The sequence of words, the proposition, the speaker's intention, the reader's likely response, the genre, the sound, the legal effect or the historical distance? In easy passages these demands align and the question disappears. In difficult passages they compete. Professional judgement begins where the translator makes that competition visible to themselves and, where necessary, to the client or reader.

A hierarchy does not grant unlimited freedom. It makes the obligations explicit. If a translator chooses readability over syntactic closeness, factual relations still have to remain sound. If metre is protected in a poem, an invented image cannot be smuggled in without consequence. If a technical term is replaced by a plain-language description, the description must still distinguish the intended concept from its neighbours. A defensible translation can explain why its losses were necessary and why its gains matter.

The source itself may also require a priority decision. Ancient works can survive in differing manuscripts. A modern document may contain a broken sentence, an undefined abbreviation or copied wording that contradicts a diagram. The translator has to decide whether to reproduce, query, correct, annotate or suspend judgement. Quietly repairing every defect makes the target cleaner than its source. Reproducing every defect can make the translator look incompetent. Source criticism and query management therefore belong inside fidelity.

The hierarchy can be recorded rather than guessed. A literary translator may state that the edition preserves abrupt syntax and repeated key words, while a technical brief may rank safety, terminology and consistency above style. Translator's prefaces, client instructions and editorial queries are not administrative clutter when they disclose the governing choices. They let readers and reviewers test the version against its declared aim. Without such a record, an elegant result can conceal a changed priority halfway through the work.

The useful question is therefore not “Is this close?” but “What is this version protecting, and was that the right protection for the job?” That question makes comparison possible. Two translations can disagree sharply without one being careless. They may have accepted different obligations to the source and reader. The disagreement becomes intelligible once the hidden hierarchy is brought into view.

Meaning Is More Than What a Sentence Refers To

A sentence does more than describe the world. It can warn, flatter, threaten, refuse, joke, apologise, promise, imply or test a relationship. Linguists distinguish the relatively encoded content of words and structures from the meaning people infer in context. Translators need both. A version that preserves the proposition but changes the social action has carried the shell and lost the event.

“It is cold in here” may report a temperature. Said to somebody beside an open window, it may function as a request. The words do not contain the request as a detachable object. Speaker, setting, shared knowledge and convention produce it. A translation that turns the line into an explicit command might help one audience understand the action, but it would also remove the indirectness. That indirectness may show tact, hierarchy, resentment or uncertainty. The translator must decide whether to preserve the wording, the request, the manner of requesting, or some workable combination.

Register creates the same pressure. A surgeon, a child, a bureaucrat and a friend can refer to the same event in different linguistic forms. Their vocabulary and sentence shape place them socially. Flatten those differences and characters begin to sound alike. Inflate ordinary speech into formal prose and a witness may appear more certain or educated than they were. In legal and medical settings, such changes can affect credibility, consent and risk rather than literary texture alone.

Meaning also includes presupposition and implication. A question such as “Have you stopped taking the tablets?” assumes that the person had been taking them. A political phrase may rely on a slogan familiar to local listeners. Irony can reverse the apparent evaluation without reversing the words. Pronouns such as this, there and we depend on who is speaking and from where. Ambiguity may be deliberate, as in diplomacy, comedy or fiction. A translator who clarifies every uncertainty can misrepresent a text by making it more decisive than the source.

Grammar divides experience in ways that are easy to overlook from inside one language. Languages differ in which relations they require speakers to mark, including number, gender, evidential source, aspect and politeness. This does not imprison thought, but it changes what a sentence must make explicit. English may force a choice that the source left open, while another language may require information English speakers normally infer. Each compulsory choice creates a small risk of interpretation posing as transfer.

Some languages regularly require information that English can leave optional. A grammatical form may distinguish whether a statement was witnessed, inferred or reported. English can often say the same thing without marking the source. Translating into the more explicit system may force a choice the original sentence did not make. Translating in the other direction may erase a distinction that helped the listener judge confidence. The difficulty is not exotic vocabulary. It is that grammar has placed evidence inside the ordinary shape of the sentence.

Meaning is therefore distributed across text and situation. Dictionaries help with conventional senses, but they cannot identify the current act without context. Corpora show how expressions behave across real examples, but the translator still has to judge this example. The stronger the translation, the more it preserves not merely what the sentence says, but what the speaker is doing by saying it.

Equivalence Depends on Purpose

Translation theory has spent centuries asking what counts as an equivalent. The most useful answer is modest: equivalence is a relation established for a purpose, not an identical substance waiting to be found. Two expressions may be equivalent for one task and poor substitutes for another. A bilingual road sign, a lyric, a patent claim and an easy-read health leaflet need different forms of sameness.

Roman Jakobson described translation across signs as “equivalence in difference”. He distinguished rewording within a language, translation between languages and interpretation between verbal and non-verbal sign systems. The categories matter because they reveal that reformulation begins before a second language appears. Paraphrasing a technical explanation for a child already requires decisions about what must remain and what can change. Interlingual translation intensifies that problem rather than creating it from nothing.

Twentieth-century approaches gave different names to the priority question. Eugene Nida contrasted greater attention to source form with an orientation towards a comparable response in the receiving audience, developed through his work on Bible translation. Katharina Reiss linked methods to text functions, while Hans Vermeer placed the intended purpose, or skopos, at the centre of the commission. These models are often reduced to slogans. Used carefully, they make one point: a translation cannot be assessed apart from what it is meant to do.

A museum label may need to be brief, factual and usable by visitors who are standing up. A scholarly edition of the same source may preserve technical distinctions, variants and uncertainty that would overwhelm the wall text. A performance translation must be speakable. A search keyword has to match the terms users are likely to type. A safety instruction must trigger correct action under pressure. Calling all of these versions “free” or “literal” hides the criteria that matter.

Purpose does not excuse distortion. A client cannot make a false statement accurate by writing a convenient brief. Translators have ethical and professional duties to flag impossible instructions, corrupt source text, conflicting terminology and risks that the commissioner has missed. Skopos is a way to organise choices, not permission to produce whatever achieves a desired effect. The source, audience, institution and likely consequence continue to constrain the work.

Many commissions also have more than one purpose. A government website may need legal consistency, plain language, accessibility and public confidence. A literary edition may serve new readers while remaining useful to students. These aims can pull in different directions, so naming a purpose does not end the argument. It begins a ranking process. The translator and commissioner need to decide which failure would be most serious, who bears it and which compromises should be disclosed.

Purpose also helps draw borders among translation, adaptation and localisation. The categories overlap, but labels matter. A version may abridge a novel for younger readers, replace references for a new stage production or redesign an interface for local conventions. Such work can be excellent, yet the audience should not be led to believe that every change was demanded by the source language. The more the commission permits alteration of content, sequence or setting, the stronger the case for naming the result accordingly. Transparency protects both the maker and the source.

The concept also explains why the “best translation” may be a category error. Best for first-time readers, actors, lawyers, historians, worshippers, children or language learners? Best at preserving lineation, argument, humour or archival trace? Once the purpose is stated, quality becomes more testable. Until then, praise often means only that a version suited the reviewer's own unstated use.

Form Carries Information

Readers often separate content from form as though the first were the message and the second its wrapping. In many texts, form is part of the message. Sound can make a phrase memorable or comic. Word order can delay a revelation. Repetition can create ritual, obsession or authority. Punctuation controls pace and attachment. Typography can distinguish voices. Verse form can establish expectation and then break it. A translation that reports the content while neutralising these patterns has preserved an account of the work rather than the work's full operation.

Poetry makes the conflict obvious. Homer's Iliad begins, in a close prose rendering, “Sing, goddess, the anger of Achilles, Peleus' son.” The Greek places anger first. Alexander Pope's rhymed 1715 version begins, “Achilles' wrath, to Greece the direful spring / Of woes unnumber'd, heavenly goddess, sing!” Pope keeps wrath and invocation but moves the command, adds Greece and spring, expands one Greek line into a couplet and replaces hexameter with English heroic couplets. The close version exposes source order; Pope makes a new English poem. Asking which is more accurate still leaves the governing obligation unstated.

Rhyme, metre, line length, image, syntax and tone rarely align across languages. Retaining a rhyme scheme may force unusual vocabulary or altered emphasis. Abandoning it may release semantic precision while removing the pressure that shaped the original. A prose version can clarify argument and imagery, yet it cannot claim that line breaks and sound were disposable. Multiple translations remain valuable because each lets a different structure survive.

Wordplay is the concentrated case. A pun depends on one form supporting more than one sense. The target language may not possess the same accident. The translator can replace it with a different pun, explain it, preserve one meaning, move the joke to a nearby line or allow the loss to stand. This practice is often called compensation: recreating an effect elsewhere when the original position cannot carry it. Compensation is not a licence to decorate. It is an attempt to balance the text's pattern across a larger span.

Spoken form adds timing, breath and body. An interpreter has to manage pauses, repairs, unfinished sentences and overlap. A dubbed line must fit a performance and often the visible movement of a speaker's mouth. Subtitles reduce speech to readable text while competing with the image for attention. They usually cannot retain every repetition, hesitation or term of address. The omission may be necessary, but it still changes characterisation and rhythm. Good audiovisual translation decides what the viewer must hear in order to experience the scene rather than receive a transcript.

Signed languages expose another mistake: language is not speech written differently. Natural sign languages use handshape, movement, location, facial expression and spatial grammar. Translation between a signed and a spoken language crosses both linguistic and sensory organisation. A written subtitle may provide access to verbal content without reproducing the full expressive resources of the signed performance.

Visual form can be equally stubborn. An acrostic places meaning in initial letters. A lipogram creates pressure by excluding one part of the alphabet. A comic distributes speech across panels, balloons and images whose order controls timing. A poem may use blank space as pause or fracture. Translating the words without noticing the visible rule can remove the work's organising intelligence. Recreating it may require extensive changes elsewhere, and even then the target alphabet or page design may make exact imitation impossible.

Form survives when translators treat it as evidence. They identify what a pattern is doing before deciding how closely to imitate its surface. Sometimes the right choice is formal resemblance. Sometimes it is a new device that performs a related job. Sometimes no adequate substitute exists. Honest translation names the sacrifice rather than pretending content escaped untouched.

Culture Must Be Rebuilt, Not Decorated

Languages carry assumptions about family, status, institutions, food, religion, humour, time and ordinary conduct. These are not ornaments added after grammar. They help sentences make sense. A kinship term may encode distinctions English leaves to context. A school qualification, court office or form of land tenure may have no institutional twin. An everyday food can evoke class, season or childhood while its dictionary label evokes nothing. Translation has to rebuild enough of that world for the target text to function.

One option is borrowing: retaining the source term and letting readers learn it. Another is a descriptive phrase, a local analogue, a footnote, a glossary or a small addition in the sentence. Each changes the reading experience. Borrowing preserves difference but may slow or exclude. Substitution produces speed but can create a false equivalence. Explanation adds knowledge while altering pace and voice. There is no universal ranking because the right amount of resistance depends on genre, audience and purpose.

Lawrence Venuti uses domestication and foreignisation to describe broad tendencies. A domesticating version reduces signs of foreignness and reads according to target-language expectations. A foreignising version retains more visible difference and may disturb those expectations. The terms are useful when treated as tendencies rather than moral camps. A children's instruction may need familiar phrasing. A novel may be harmed if every object, form of address and sentence rhythm is made comfortably local. Most translations mix the two.

Names show how practical and political the choice can become. Scripts may require transliteration, yet several systems can represent the same sounds differently. Historical conventions may survive long after pronunciation changes. Some names carry meanings that matter inside a story; others should remain names. Exonyms such as English place names can be familiar without being neutral. Institutions, publishers and governments often standardise spellings, which means the chosen form may reflect authority as well as sound.

Dialect presents a sharper danger. Replacing a regional or class-marked voice with a target dialect can preserve social markedness while importing the wrong history. A rural speaker may acquire comic or backward associations that the source did not assign. Neutralising the voice removes inequality, intimacy or identity. There is no clean answer, but there are careless ones. The translator should understand what the source variety signals and avoid treating marginal speech as a costume.

Idioms reveal the same trap at a smaller scale. A literal version may reproduce an image that target readers treat as bizarre rather than ordinary. A local idiom may preserve ease and function while importing a different animal, object or social memory. Sometimes plain language is the least misleading answer. Sometimes the source image matters because it returns later or belongs to a character's world. An expression that seems disposable in isolation may be structural across the work, so cultural judgement requires attention beyond the sentence.

The label untranslatable often means that no concise equivalent carries the same network of associations. It does not mean communication has stopped. The translator can borrow, paraphrase, stage the context or let readers encounter a productive gap. What cannot survive in one word may survive across a sentence, a chapter or the reader's growing familiarity. Culture crosses through reconstruction, not through replacing every unfamiliar thing with something already known.

Translation Is Made by Systems

The solitary translator at a desk is real, but incomplete. Most consequential translations pass through a system. A commissioner defines the audience and deadline. The source may arrive with errors or missing context. Terminology lists and previous versions constrain choice. Software segments text and retrieves earlier matches. Reviewers query inconsistency. Designers impose space. Lawyers, clinicians, engineers or subject specialists resolve concepts. The published version is the record of those interactions, even when one name appears on the cover.

Historical translation was often organised in the same way. Buddhist texts translated into Chinese could involve recitation, oral explanation, drafting, checking and editorial work by groups with different linguistic and doctrinal expertise. The Graeco-Arabic translation movement under Abbasid patronage depended on scholars, patrons, scribes and successive revisions rather than one mythical room producing all knowledge. Translation associated with medieval Toledo likewise occurred through varied networks, languages and institutions. The recurring lesson is that major transfers require money, access, labour and authority.

At the turn of the twentieth century, Lin Shu made the system visible. Knowing no foreign language, he listened while collaborators orally relayed novels and recast them in classical Chinese. Across about twenty years, Lin worked with nineteen assistants on more than 180 book-length translations. Source-language knowledge remained essential because his collaborators supplied it, yet credit gathered around Lin. The published voice emerged from oral mediation, his prose, commercial selection and print. Translation was the partnership, including its unequal visibility.

Interpreting makes the system visible because the participants share the event. Court and medical interpreters work within rules about accuracy, impartiality, confidentiality and role. They need enough briefing to recognise names and technical terms, yet they may be denied documents. Turn length, acoustics, fatigue and the behaviour of speakers affect the result. A qualified interpreter can still be placed in a poor system. Conversely, a well-designed process cannot turn an untrained bilingual relative into a neutral professional.

Commercial translation adds translation memories, termbases, machine output and quality checks. Reusing an approved sentence can improve consistency and speed, especially in repetitive documentation. It can also import an old error or a phrase suited to another context. Machine translation may supply a draft for human post-editing. The usefulness of that arrangement depends on language pair, domain, source quality, risk and the standard required. “Human in the loop” says little until the human's authority, time and competence are specified.

Workflow tools also change the unit of attention. A translation memory presents text in segments, often sentences or smaller units, which helps reuse and counting but can weaken awareness of argument, voice and reference across a document. Platform interfaces may hide page layout, images or neighbouring strings. Payment by word and deadlines can reward speed at the segment level while the text's coherence depends on slower reading. These effects do not make tools harmful by nature. They show that technology distributes attention. A sound system restores context through previews, reference files, document-level checks and permission to revisit an earlier choice when later material changes its meaning.

Quality assurance belongs to the system as well. A second linguist may revise the translation against the source, while another review checks target-language suitability or specialist accuracy. Automated checks can flag missing numbers, inconsistent terms or broken tags. None proves that the communication works. A process standard can allocate competence and review, but it cannot settle a disputed interpretation or rescue a defective brief. Good systems combine repeatable controls with room for reasoned challenge.

Systems decide what receives translation at all. Publishers select a fraction of world literature. States choose which languages receive forms, hearings and public information. Platforms support some scripts and interfaces earlier than others. Censors remove passages. Editors may erase difficulty in the name of sales. Data-rich languages receive stronger machine tools, which can attract more use and generate more data. Translation access and translation quality are therefore distributed through institutions, markets and policy.

Recognising the system does not dissolve individual responsibility. It places responsibility where decisions are made. When an error appears, the cause may be a mistranslated term, an impossible deadline, missing reference material, faulty software, weak review or pressure to accept an unsafe source. Better translation comes from improving the chain, not from treating the last person who touched the words as the whole process.

The Version Changes the Original's Afterlife

Translation is often judged as a copy standing beside a stable original. In circulation, the relationship becomes more complicated. Most readers may know the work only through one version. Its diction, omissions and explanations begin to shape what the author is thought to sound like. A translated phrase can become a proverb, a political slogan, a theological formula or the standard name for a technical idea. The version enters the source's public identity.

Sacred and legal texts show the force of this effect because communities build interpretation around wording. A familiar phrase can acquire authority through repetition even when later scholarship favours another rendering. Literary translations do something similar through voice. One Homer may sound ceremonious, another rapid, another contemporary. Readers then argue about whether Homer possesses those qualities as though no translator had helped establish them. The source has not changed, but its accessible afterlife has.

Translation can also alter the receiving language. Borrowed terms fill gaps, while new sentence patterns and genres expand what writers think possible. Philosophical and scientific vocabularies often develop through choices made while carrying concepts across languages. A successful term can standardise discussion. A bad one can misdirect it. Either way, later speakers inherit the decision without meeting the translator who made it.

This influence explains retranslation. Languages change, scholarship advances, audiences shift and previous versions become historically marked. A new translator may recover censored passages, revise a colonial convention, restore formal difficulty, use a different manuscript or write for performance rather than study. Retranslation is not proof that earlier work failed. It is evidence that no version exhausts the source and that every audience arrives with new requirements.

The frame around a translation joins this afterlife. A title can redirect emphasis before the first page. A cover can market tragedy as romance or philosophy as self-help. Notes can make difference legible, while their absence can make the target culture seem self-explanatory. A translator's name may be prominent, hidden or omitted. These surrounding choices are called paratexts, and they influence whether readers meet the work as foreign, ancient, difficult, canonical or familiar. The crossing includes its customs desk.

Power enters here too. Languages with strong publishing and educational institutions can circulate their interpretations widely. A translation into a dominant language may mediate later translations into smaller ones, so one set of decisions travels far beyond the first target audience. Reviews, school syllabuses, prizes and film adaptations can confirm that version's authority. The route by which a text becomes “world literature” is partly a route through gatekeepers.

Every translation begins by ranking obligations. The afterlife is where that ranking ceases to be private. Once a version is read, taught and quoted, its selections pressure the next translator. Familiar wording becomes a benchmark. Omissions may disappear from public memory. New readings arise in response. What crossed first changes what later crossings are expected to carry.

The subtitle asks what survives. The deepest answer is that survival is not unchanged possession. A text survives abroad by being rebuilt often enough to remain active. Some features fall away, others return in new versions, and some effects appear only after the crossing. Translation preserves by transforming, then leaves the transformed work available for correction, argument and another crossing.

How It Actually Works

Before the dictionary

The Rosetta Stone is a useful correction to the idea that early translation resembled a modern parallel text. Carved in 196 BCE, it presents a priestly decree in hieroglyphic and Demotic Egyptian scripts and in Greek. The inscriptions communicate related official content for different literate communities, but they are not a line-by-line vocabulary key. Their value to later decipherers came from overlap, known names, repeated structures and comparison with other evidence. Translation helped produce the object; decipherment required much more than substituting Greek words for Egyptian signs.

Multilingual states depended on spoken mediation as well as inscriptions. Envoys, merchants, soldiers, tax officials, captives and border communities carried messages between languages long before professional roles were standardised. Their work was often inseparable from diplomacy, trade or administration. A mediator could explain local practice, soften an insult or control access to a ruler. That made language knowledge valuable and politically dangerous. Trust attached to the person because no recording allowed the audience to check each decision afterwards.

Writing created a different kind of control. Scribes could compare versions, compile word lists, preserve formulae and send the same order to several regions. Yet written sources introduced their own uncertainties. Which manuscript was copied? Was an unfamiliar word a technical term, a spelling variant or an error? Did the target community share the institution being named? The translator's task already included source criticism and audience design, even where no theory used those labels.

The archive records these mediators unevenly. A royal inscription may preserve a ruler's name and erase the scribes who shaped its versions. A treaty may identify envoys but omit the interpreter. Surviving texts therefore exaggerate the visibility of commissioners and authors while hiding much of the language labour. Absence of a named translator is not evidence that translation happened automatically.

Surviving inscriptions also show translation serving more than curiosity. Rulers and temples used multiple languages to claim authority, collect revenue and address different populations. Communities used translation to preserve worship, law and memory under changing political conditions. Much of the surviving ancient record therefore shows translation organised by power, but that archive cannot stand for every exchange. States and temples made durable records; ordinary oral mediation usually vanished. Power shaped which translations were funded, preserved and legible to us. It did not determine every sentence. Translators still had to solve grammar, genre and cultural mismatch one phrase at a time.

Sacred texts and scholarly teams

Sacred translation made wording consequential because communities could treat a version as scripture in its own right. The Greek translation of Jewish scriptures known as the Septuagint developed over time, beginning with the Pentateuch in Hellenistic Egypt. A later tradition told of seventy-two translators producing an agreed version under miraculous conditions. The story expresses a claim about authority. It is not a reliable workshop report. The corpus contains differences of style and approach that fit a longer, more varied process.

Jerome's late fourth-century arguments show the dispute from another direction. Defending his Latin translation practice, he said that he usually rendered sense for sense rather than word for word, while making a special qualification for scripture. The formula did not end the literal-versus-free debate. It named a conflict translators had long faced: source order may carry authority or artistry, yet copying it can produce a target sentence that misstates the sense. Religious institutions added doctrinal review, inherited wording and communal memory to the linguistic problem.

In China, Buddhist translation made collaboration visible. Major projects associated with Kumārajīva in the late fourth and early fifth centuries involved people who could recite or explain source texts, others who drafted Chinese, and audiences or editors who tested doctrinal and stylistic choices. The process varied, and the surviving accounts should not be turned into one fixed production line. The important point is that a celebrated translator's name can stand for coordinated labour across languages, philosophy and literary form.

The Graeco-Arabic translation movement of the eighth to tenth centuries depended on similar infrastructure. Patrons funded access to manuscripts and skilled labour. Scholars worked across Greek, Syriac and Arabic. Hunayn ibn Ishaq described seeking multiple copies, comparing them and revising earlier versions while translating medical works associated with Galen. Baghdad mattered, but the work was not confined to one building called the House of Wisdom. It was a network of courts, scholars, libraries, professions and changing commissions.

Many of these transfers passed through an intermediate language. A Greek work might reach Latin through Arabic, with each stage carrying earlier interpretation, terminology and error. Later scholars sometimes recovered another source and revised the inherited version. The chain complicates the idea of one original facing one translation. Texts can survive through families of versions whose members preserve different evidence.

From the twelfth century, translators in and around Toledo helped move works from Arabic into Latin and Castilian. Gerard of Cremona's Latin version of Ptolemy's Almagest, made from Arabic, became one important route into western European astronomy. The familiar phrase “Toledo School of Translators” can suggest a single permanent academy. Evidence points instead to varied collaborations and patrons across time. As before, knowledge travelled through institutions, intermediate languages and repeated revision rather than a clean leap between two self-contained cultures.

Print, empire and the public version

Print did not remove variation, but it changed its economics. A manuscript version might circulate through a limited chain of copies, each open to alteration. A printed translation could reach larger audiences in a more stable form, accumulate corrections across editions and become a public standard. The translator's decisions were now reproduced with the printer's errors, the editor's interventions and the publisher's commercial judgement. Stability made comparison easier and inheritance harder to escape.

Religious conflict turned vernacular translation into a fight over access and authority. In sixteenth-century Europe, translators of scripture worked amid disputes about which source texts counted, whether ordinary readers should encounter them directly and who could license the wording. William Tyndale's English New Testament of 1526 was enabled by print and condemned by ecclesiastical and secular authorities; Tyndale was executed for heresy in 1536. Later English Bibles absorbed much of his phrasing. A version can be suppressed in one generation and become inherited language in another.

Empires used translation to administer populations, classify law, negotiate with elites and support missionary work. Grammars and dictionaries could record languages while forcing them into imported categories. Interpreters could give colonised people access to officials while serving the machinery that ruled them. Translated education could widen literacy or displace local knowledge. The same infrastructure could later support anti-colonial newspapers, political argument and new national literatures. Translation is an instrument whose effects depend on who commissions it, who controls circulation and who can answer back.

Modern nation states expanded the demand for standard terminology. Schools, courts, armies, railways, medicine and mass bureaucracy needed repeatable forms. International organisations added another scale. The European Union now recognises twenty-four official languages, requiring extensive translation, terminology management and interpreting. That does not mean every internal conversation occurs in every language. It means citizens, law and institutional communication operate under a formal multilingual commitment whose daily maintenance is largely invisible.

Print also strengthened the paratext around translation. Prefaces defended methods, glossaries taught foreign terms, and title pages named or concealed translators. Editions could announce that they were corrected, authorised, modernised or made from an original language. Such labels sold trust while making claims that readers were rarely able to test. The modern habit of comparing editions grew from this abundance: stable versions made disagreement visible.

The public version therefore emerges from selection. A publisher chooses which foreign novel deserves investment. A ministry chooses which languages receive a form first. An editor chooses whether the translator's name appears on the cover. Search systems and booksellers classify the result. By the time a reader reaches the opening line, many crossings have already been refused, funded, shortened or framed.

The modern written workflow

A robust written translation begins before drafting. The commissioner should identify the audience, purpose, format, deadline, reference material, confidentiality requirements and consequences of error. The translator inspects the source, raises queries and decides whether the job requires subject expertise, software support or another specialist. A poor source may need correction by its author before translation. Starting immediately can feel efficient while locking uncertainty into every later stage.

Research then establishes terminology and context. Translators use dictionaries, corpora, parallel documents, legislation, product interfaces, specialist publications and direct queries. A termbase records approved concepts and distinctions. A translation memory stores segments from earlier work and retrieves possible matches. Neither tool decides the current sentence. A previous segment may contain an error, refer to a different product or use wording no longer approved. Reuse is evidence, not authority.

Drafting moves between local and whole-text decisions. The translator identifies the proposition, the action, the register and the function of form, then writes a target passage that can perform the required job. Difficult points may remain provisionally marked until later context resolves them. Names, numbers, units, cross-references and formatting need separate attention because fluent prose can hide a single consequential mismatch. Consistency matters, but mechanical consistency can be wrong when one source word serves several concepts.

Revision requires distance from the first solution. The translator checks the draft against the source and then reads it as target-language communication. Many professional workflows add bilingual revision by another qualified person, followed where relevant by specialist review, proofreading or functional testing. ISO 17100 sets process and competence requirements for translation services within its scope, including revision by another person. It does not certify that every resulting interpretation is correct, and it does not cover every form of language work.

Version control matters because translation is cumulative. A late source edit can invalidate an approved paragraph. Several reviewers can make conflicting changes. A query answered in email may never reach the next edition. Good project records connect each target version to the correct source, preserve decisions and distinguish a linguistic correction from a client's stylistic preference. Without that trail, consistency becomes memory under deadline.

Quality assurance can flag omissions, inconsistent terms, unmatched numbers, punctuation, tags and forbidden wording. A software localiser may test whether text fits buttons and whether variables still work. A legal team may reconcile defined terms across documents. A publisher may read the book aloud. Final approval should return to the declared priorities: does the text perform its purpose, and are the remaining compromises acceptable at this level of risk?

Speech, screens and constrained time

Interpreting removes the luxury of unlimited revision. In consecutive interpreting, the speaker pauses while the interpreter renders a portion, often using notes. In simultaneous interpreting, the interpreter listens and speaks with a short lag. Both require analysis rather than echoing. Word order may force the interpreter to wait for information, restructure the sentence and correct a false start while the next sentence is already arriving.

The International Military Tribunal at Nuremberg in 1945 and 1946 made large-scale simultaneous interpreting central to a highly visible multilingual trial conducted in German, English, French and Russian. It did not invent the method, but it demonstrated an institutional system of booths, equipment, teams and rotation under extreme legal pressure. The lesson was organisational as much as linguistic. Speed depended on preparation, technology, disciplined turn-taking and enough interpreters to manage fatigue.

Court, healthcare and public-service interpreting add role constraints. The interpreter should render what participants communicate rather than become an informal advocate, investigator or family decision-maker. Yet neutral performance still requires interventions: asking for repetition, flagging an unknown term, managing overlap or stating that an utterance cannot be heard. Institutions that deny briefing, use an inappropriate language variety or treat any bilingual person as interchangeable create risk before the interpreter begins.

Audiovisual translation works against the clock and image. Subtitle guides set limits for timing, line length and reading speed, but the numbers vary by platform, audience and language. The subtitler condenses while preserving plot, character and the relation between speech and picture. Dubbing adds performance, synchrony and casting. Audio description translates selected visual information into speech. Game and software localisation must preserve commands, variables, humour, legal notices and user expectations inside a working product.

Signed-language interpreting brings further arrangements. The interpreter may need sightlines, lighting and enough space for participants to see one another. Some encounters require relay interpreting through an intermediary language or specialist communicator. Remote platforms can widen access while introducing delay, poor framing or lost visual cues. Modality and technology are part of the communicative event, not a neutral pipe around it.

These forms show why a transcript is not the universal ideal. Repeating every hesitation can make subtitles unreadable. Removing every hesitation can make a frightened witness sound composed. The correct degree of compression depends on what the medium already supplies and what the audience needs to act or understand.

From rules to statistical systems and language models

Early machine translation often relied on hand-built dictionaries and grammatical rules. Such systems could behave predictably in narrow domains but were costly to extend and brittle around ambiguity. Statistical machine translation shifted attention towards patterns learnt from bilingual corpora. Phrase-based systems estimated likely correspondences and target sequences from large collections of aligned text. They improved coverage while inheriting the biases, omissions and domain limits of their data.

Neural machine translation learnt distributed representations and generated target sentences through neural networks. The Transformer architecture introduced in 2017 made attention the main mechanism for relating positions in a sequence and became foundational to many later systems. Quality improved sharply for numerous high-resource language pairs, and output became more fluent. Fluency reduced one visible warning sign. A grammatical sentence can still select the wrong sense, drop a qualification or invent a relation that the source does not support.

Large language models can translate through prompting, use surrounding context and respond to instructions about tone or terminology. They can also paraphrase beyond the source, normalise ambiguity, omit repeated material or produce unsupported detail. Performance varies across languages, scripts, domains and document lengths. A model strong on public English-French news may be weak on a low-resource language, a regional variety, confidential litigation or a scanned medical form. The label “AI translation” hides these differences.

Evaluation has to match the claim. BLEU, introduced in 2002, compares overlap between system output and one or more reference translations using sequences of words. It can support aggregate system comparison under suitable conditions. It does not measure every valid translation, resolve meaning or certify an individual sentence. Modern shared evaluations such as the annual WMT campaigns combine automatic measures with human assessment across selected languages and domains. Their results describe tested conditions, not translation as a whole.

The 2025 WMT general task shows what that caution looks like. It reported results across thirty language pairs and sixty systems, using document-level test sets drawn from four or five domains, while human assessment covered a selected half of the language pairs. Its organisers also argued that easy test sets no longer separated leading systems well enough. The finding is important, but bounded. It does not create one global winner, and it says little about an untested dialect, private contract or spoken encounter. A strong benchmark narrows uncertainty for specified material. It does not turn a changing collection of systems into a solved field.

Document context remains a practical weakness. Systems often process units smaller than the whole work, even when a larger context window is available. Names, pronouns, terminology and tone can drift across sections. Longer input does not guarantee that the model will use the relevant information or preserve every repeated detail. Evaluation at sentence level can therefore miss document-level failure.

Human review should be proportional to consequence. Machine output may be useful for triage, search, gist or a first draft when its limits are acceptable. Publication, contracts, medical communication and other high-stakes uses may require qualified review, specialist validation and secure handling. Post-editing standards can define workflows, but the reviewer still needs enough time and authority to reject the output. Automation changes where effort is spent. It does not remove responsibility for what crosses.

How we know

Translation leaves several kinds of evidence, none complete on its own. Parallel texts let scholars compare wording, omissions and structure. Manuscript variants, drafts, letters, prefaces and marginal notes reveal some decisions, while institutional archives identify patrons, reviewers and circulation. Interpreting is harder to reconstruct before recording because the performance disappears. Later transcripts and audio preserve output but not every cognitive choice or unspoken constraint.

Claims about current machine systems require separate caution. Published benchmarks test selected data, language pairs, prompts, domains and scoring methods. Human evaluations differ in instructions and sampling. Commercial models change, and proprietary training data or system details may be undisclosed. A result from one pair or field should not become a universal ranking.

The strongest account therefore combines textual comparison, historical records, practitioner knowledge and controlled evaluation. Where motives or private decisions are absent, this book does not invent them. Where a famous origin story rests on later tradition, it is identified as tradition. Translation can be studied closely, but the surviving record is itself a partial version of the work.

What People Get Wrong

Translation attracts myths because readers see the finished target text and much of the decision-making has vanished. The following mistakes matter because they reward the wrong evidence, hide risk and turn reasonable disagreement into accusations of incompetence.

“A good translation replaces every source word”

Languages do not divide meaning into matching pieces. One source word may require a phrase; several source words may form one target expression; grammar may carry information for which no separate word exists. Replacing items one by one can preserve a visible pattern while destroying the proposition or action. Idioms make the failure obvious, but ordinary syntax produces it constantly.

Word-level attention remains essential. Omitted negatives, numbers and qualifications can cause serious harm. The correction is not permission to ignore words. It is to treat them as evidence inside a larger unit. The translator asks what relation the words create in this sentence, text and situation, then rebuilds that relation with as much formal continuity as the task allows. Interpreters do the same under time pressure. They are not human dictionaries operating aloud.

Literal glosses still have a place. They can expose syntax, help students inspect a source and give reviewers a diagnostic layer. Trouble begins when a gloss is presented as finished communication or when visible resemblance is treated as the highest form of accuracy.

“There is one perfect translation waiting to be found”

Some solutions are plainly wrong. A medicine label that changes the dose, a contract that reverses an obligation or a novel that invents an event has failed. It does not follow that every sound source sentence has one complete target twin. Several versions may preserve the facts while differing in register, rhythm, historical distance and explanation.

The demand for one perfect version usually hides an unstated purpose. A reader wants the easiest edition, a scholar wants traceable syntax, an actor wants speakable lines, or a community wants inherited liturgical wording. Those aims can support different choices. A stronger judgement states the use and compares versions against it. Perfection would require every relevant feature to survive without interference. Between differently organised languages, that condition often cannot be met. Quality is disciplined compromise, not the discovery of a secret answer key.

This is why retranslation is healthy. New evidence, changed language and a different audience can justify another version without erasing the old one. The existence of several strong translations is often evidence of the source's richness, not proof that the profession has failed.

“If it reads fluently, it must be accurate”

Fluency is evidence that the target text conforms to familiar language patterns. It is not evidence that the source has been represented faithfully. A polished sentence can omit a warning, resolve an ambiguity, change who performed an action or make a tentative claim sound certain. Modern machine systems make this distinction urgent because they can produce confident prose even when they have selected the wrong sense.

The reverse mistake also occurs. Awkwardness can signal poor translation, but it may preserve strangeness, technical structure or a difficult voice that belongs to the source. Review therefore needs at least two views: comparison with the source and assessment as target communication. A monolingual reader can judge readability but cannot certify completeness. A bilingual reviewer can compare meaning but may still lack the specialist knowledge needed for law, medicine or engineering.

Fluency should therefore be tested as one quality dimension among several. Completeness, factual relation, terminology, register and fitness for purpose need their own checks. The more effortless a wrong sentence sounds, the less likely an ordinary reader is to stop at the point of failure.

“An untranslatable word cannot cross into another language”

A word called untranslatable usually lacks one concise target item with the same range, associations and social history. That is a mismatch, not a wall. The translator can borrow the word, explain it, use a phrase, build the context across several appearances or select one sense where the passage demands it. Readers learn unfamiliar concepts through use.

The romantic version of untranslatability treats each language as a sealed worldview. Languages do guide attention through habitual categories, and some distinctions are costly to express elsewhere. Speakers can still describe, teach and debate those distinctions. Translation may require expansion or leave residue, but partial loss is different from impossibility. Calling a term untranslatable can be useful when it warns against a cheap one-word equivalent. It becomes misleading when it turns difficulty into mysticism or markets a foreign word as access to a unique national soul.

Borrowed words themselves disprove the sealed-world claim. Languages repeatedly import terms, then bend their spelling, sound and meaning through use. The concept may arrive with distortion, but it enters public argument and becomes available for further correction. Crossing is a process, not an instant equivalence.

“The translator should leave no trace”

Readers often praise a translation for feeling as though it were written directly in the target language. That can be an achievement, especially where the purpose demands unobstructed use. It can also hide how much has been normalised. Syntax, names, dialect, cultural reference and difficulty may have been reshaped to fit local expectations. Smoothness is a choice with consequences.

No translator can remove all trace because choosing not to explain, not to borrow and not to disturb is still a method. The ethical aim is not theatrical visibility on every page. It is accountable presence. Credit the translator, disclose material interventions, explain unusual priorities where readers need them and resist institutions that make language labour disappear. Interpreters face the same distinction. Professional impartiality means managing the communicative role fairly, not pretending that timing, clarification and turn control never affect the event.

Erasing the translator can also misallocate authority. Readers may blame an author for a phrase selected by an editor, or quote a translator's interpretation as the author's exact wording. Visible credit gives the public a more accurate map of authorship and responsibility.

“Back-translation proves the first translation is correct”

Back-translation takes a target version and translates it back into the source language, usually without showing the second translator the original. Differences can reveal omissions, changed concepts or unexpected ambiguity. The method is useful in questionnaire adaptation, research and other controlled workflows when combined with expert review and testing.

It is not a proof machine. A poor forward translation can return to wording close to the original by chance. A good idiomatic translation can back-translate into different wording and look suspicious. The back-translator introduces another set of priorities, so the comparison now contains two transformations. Evidence from methodological studies supports team-based review rather than mechanical acceptance of matches. The right question is what the discrepancy reveals about the intended concept and audience. Counting surface differences without that analysis can reward literalness and punish effective adaptation.

A stronger process brings discrepancies back to a team that includes linguistic and subject expertise, then tests the resulting version with representative users where possible. Back-translation is most informative as a prompt for investigation. It becomes weakest when agreement is converted into an automatic pass mark.

“Artificial intelligence has solved translation”

Current machine systems can be excellent tools. They produce rapid drafts, support gist reading, handle repetitive material and perform strongly in some well-resourced language pairs and domains. That achievement should not be denied to preserve a romantic image of human work. It should also not be inflated into a universal solution.

Performance changes with language, variety, subject, source quality, context and system version. Models can omit, invent, over-explain and flatten voice while remaining fluent. Automatic scores test limited features, and benchmark results do not certify a private document or high-stakes encounter. Human output is fallible too, which is why the answer is designed assurance rather than species loyalty. Select a system for the task, protect confidential material, test representative samples, use qualified review where consequence requires it and give the reviewer authority to reject the machine's draft.

Low-resource is not one uniform condition. A language may lack parallel data, standardised spelling, digital text, domain terminology or support for its script, and speakers may be distributed across several varieties. A system's weakness can therefore reflect political and infrastructural neglect as much as linguistic difficulty. Reporting one average score hides who receives reliable service and who bears the error.

The field is still moving. Shared evaluations can identify strong systems under stated conditions, while the next release may change the order. A leaderboard is evidence about a test, not a permanent hierarchy of languages, models or uses.

Use It

Translation literacy does not require fluency in every source language. It requires better questions, proportionate evidence and a willingness to separate readability from proof. The following lenses help when reading, commissioning or reviewing work across languages.

State the hierarchy before the first sentence

Before choosing a translation or ordering one, write down what must survive. A novel for pleasure may prioritise voice, pace and cultural texture. A contract may prioritise defined relations, completeness and jurisdiction-specific terminology. A museum label has to be accurate, brief and understandable while the visitor is standing in front of an object. A game interface must also fit the screen and trigger the right action.

Do not hide behind “faithful” or “natural”. Name the audience, use, medium, level of knowledge and consequence of error. Then rank the main commitments. Where two requirements conflict, decide which failure is less acceptable. The hierarchy can change by passage: a joke, warning and recurring symbol may each need different treatment. Writing the priorities first gives translators a defensible brief and gives reviewers a standard stronger than personal taste.

Compare pressure points, not random pages

When two editions differ, select passages where languages are likely to resist each other. Look for a joke, repeated key word, form of address, technical distinction, ambiguous pronoun, line of verse or culturally loaded object. Read the translator's preface and notes, then compare how each version handles the same pressure. Even without the source language, differences reveal priorities.

Do not crown a winner from one line. A translator may compensate later, preserve a pattern across a chapter or follow a principle that becomes visible only through repetition. Check whether the version is consistent in its chosen voice and whether it signals major departures. Reviews by source-language specialists can help, but they too have purposes and preferences. The aim is to understand the choice architecture, not to turn comparison into a parlour game about who sounds nicest in English.

When translators disagree, ask what each gap reveals about the source. One version may make an implied request explicit while another preserves the indirection. One may retain a foreign term; another may choose a functional local phrase. The difference points to a feature that does not transfer cheaply. It is evidence about the crossing.

This lens works inside one language too. Rewriting a technical policy for the public, converting speech into minutes or explaining a historical text to a child involves translation-like priorities. What knowledge can be assumed? Which distinctions must remain? What tone belongs to the speaker rather than the editor? Intralingual rewriting is not identical to translation between languages, but it trains the same discipline: state the purpose, protect the load-bearing meaning and disclose material simplification.

Treat confidence as a claim that needs evidence

A translation can sound certain because the target language or editor prefers decisive prose. Check whether the source contains hedging, reported information, ambiguity or competing readings. In a book, notes may disclose this. In professional work, query logs and reviewer comments should show how disputed points were resolved. A smooth target sentence deserves more scrutiny when its consequence is high.

This test is especially useful for quotations. A public figure's translated words may circulate without the question, surrounding sentence or interpreter's correction. A famous line from an old text may exist in several English forms. Attribute the version when wording matters. Write “in X's translation” rather than allowing a translator's phrasing to masquerade as a direct English utterance by the author. That small habit restores part of the missing chain.

Design the interpreting encounter

Good interpreting depends on the meeting around it. Identify the language and relevant variety rather than assuming a broad label is enough. Use a qualified interpreter appropriate to the setting. Provide names, documents and specialist terms in advance where confidentiality permits. Arrange sound, sightlines and turn-taking. Speak to the other participant, not about them to the interpreter.

During the encounter, use manageable segments and allow requests for repetition or clarification. Do not ask a child or dependent relative to mediate a high-stakes conversation merely because they are bilingual. In legal, medical and public-service settings, follow the applicable local rules and professional standards. Remote interpreting can solve distance and availability problems, but test the platform, privacy and visual access. The interpreter's skill matters; the conditions decide how much of that skill can be used.

Put machines on a risk ladder

Begin with the consequence of a plausible error. For gist reading of a low-stakes message, unreviewed machine output may be enough. For a public article, product interface or customer communication, sample testing and human editing may be appropriate. For consent, legal obligation, safety or material financial consequence, use qualified language and subject review, secure handling and a process that can stop publication.

Test the system on representative material rather than a clean demonstration sentence. Include names, numbers, negation, tables, repeated terms, long-range references and local varieties. Compare more than fluency. Check omissions, unsupported additions, consistency and whether uncertainty survives. Record the model or service version because output can change. Never call a person “in the loop” unless they have time, source access, competence and authority to reject the draft. A rushed reviewer can become a rubber stamp for automation.

Compare the machine route with a realistic alternative, not with an imaginary flawless human. Record the time spent preparing, correcting and checking output, and count errors that escape review. A tool may reduce cost on repetitive product text while increasing risk on short, ambiguous messages. Re-test after model, glossary or workflow changes. Efficiency belongs to the whole process, not the speed of the first draft.

Read the frame around the text

The title page, cover, introduction, notes and translator credit tell you how a version wants to be read. Is it abridged, adapted, modernised, performed or based on a particular edition? Does the translator explain names, dialect, metre or earlier versions? Is an intermediate language involved? Silence does not prove that none of these choices exists.

For publishers and commissioners, make the frame honest. Credit translators and interpreters where the medium permits. Label machine use when it materially affects trust or responsibility. Distinguish translation from adaptation when content, setting or sequence has been changed. Preserve a record of source version, approvals and unresolved issues. Paratext can burden a casual reader if it explains every decision, so use proportion. The principle is disclosure of changes that would alter a reasonable reader's judgement of the work.

The limits

No checklist can choose the right line of poetry or settle every dispute about a legal term. Readers without the source language remain dependent on translators, reviewers and institutions. Even bilingual readers may miss historical usage, specialist concepts or the effect of a dialect outside their own experience. Full certainty is rare where evidence, manuscripts or context are incomplete.

Practical advice also changes with setting. Court and healthcare interpreting rules differ by jurisdiction. Subtitle conventions differ by language, audience and platform. A machine system's performance changes across versions and data. Standards define processes within a scope; they do not make every product safe. The lenses above improve questions and design. They do not convert a non-specialist into a certified translator, interpreter, lawyer, clinician or evaluator.

There is a further limit to the loss model itself. Treating every choice as a sacrifice can make translation sound like managed failure. Versions create access, literary forms, terminology, solidarities and arguments that did not exist in the target language before. The reader should notice damage without becoming blind to invention.

The one thing to keep

Ask what had to survive, then inspect what the chosen method made expendable.

That sentence works for a poem, hearing, subtitle, software menu and machine draft. It shifts attention from closeness in the abstract to the hierarchy governing the work. It also leaves room for gain. A translator may surrender one local effect to preserve the whole movement, or add explanation so that a foreign institution becomes usable without pretending it has a domestic twin.

The strongest trust is informed rather than absolute. It credits the people and systems that made access possible, expects choices where languages differ and demands stronger assurance as consequences rise. It can admire a graceful sentence while asking what the grace cost, or reject a literal-looking version when its action has changed. Every translation is an answer to a constrained problem. Read the answer, and read the constraints.

Terms

These terms separate processes that are often collapsed into one word. Some name relationships between source and target, while others identify professional stages, technologies or interpretive choices. None should be treated as a sealed doctrine. Their value is practical: they let a reader describe where a version changed, who reviewed it, what a machine supplied and why two sound translations can differ. Definitions are kept short, but each points back to a decision made under constraint.

Adaptation

A version that permits substantial change to content, structure, setting or audience requirements. The boundary with translation varies, so transparent labelling matters more than pretending the categories never overlap.

Back-translation

Translation of a target version back into the source language, often by an independent translator. It can expose conceptual problems but cannot prove the forward translation correct.

BLEU

An automatic machine-translation metric introduced in 2002. It measures word-sequence overlap with reference translations, adjusted for output length, and is most defensible for aggregate comparison under stated conditions.

Borrowing

Retention of a source-language word or form in the target text. Borrowing preserves visible difference and can fill a lexical gap, though readers may need context or explanation.

Calque

A loan translation made by rendering the parts of a source expression into target-language elements. Some calques become ordinary usage; others remain opaque or misleading.

CAT tool

Computer-assisted translation software that supports human work through segmentation, translation memories, termbases, quality checks and file handling. It is broader than machine translation and need not generate prose.

Compensation

Recreation elsewhere of an effect that cannot be retained in its original position. It is common with humour, sound and wordplay, but must remain accountable to the wider text.

Consecutive interpreting

Interpreting in which a speaker pauses after a segment while the interpreter renders it, often using structured notes. Segment length and turn management affect completeness and memory.

Corpus

A searchable collection of real language data, sometimes aligned across languages. Translators use corpora to examine usage, collocation, register and how terms behave in comparable contexts.

Domestication

A tendency to reduce signs of foreignness and make a translation conform more closely to target-language expectations. It can aid access while concealing difference or importing false familiarity.

Dynamic equivalence

Eugene Nida's audience-oriented concept, concerned with how a translation functions for receivers rather than with close reproduction of source form. Later summaries often oversimplify his account.

False friend

A word that resembles a term in another language but differs materially in meaning or use. Familiar appearance can make false friends harder to notice than unfamiliar vocabulary.

Foreignisation

A tendency to retain or create visible linguistic and cultural difference in the target text. It can resist assimilation, though unfamiliarity alone does not guarantee accuracy or ethical value.

Formal equivalence

Nida's source-oriented tendency, giving closer attention to form and content of the original. It is a relative orientation, not a command to replace every word mechanically.

Interlingual translation

Roman Jakobson's term for interpretation of verbal signs through another language. It is the ordinary sense of translation between languages, distinguished from rewording within one language.

Interpreting

Mediation of spoken or signed communication in real time or near real time. Interpreters analyse and reformulate meaning under constraints of memory, timing, setting and interaction.

Intralingual translation

Jakobson's term for rewording verbal signs within the same language. Plain-language rewriting, modernisation and explanation can involve translation-like choices without crossing into another language.

Localisation

Adaptation of a product or service for a target locale, including language, formats, law, imagery, interface behaviour and user expectations. Translation is one component of the process.

Machine translation

Automatic generation of target-language text from source-language input. Systems may be rule-based, statistical, neural or model-based, and their performance varies by task, data and evaluation.

Paratext

Material surrounding a text, such as title, cover, preface, notes and translator credit. Paratext frames the version and can disclose or conceal the conditions of its production.

Post-editing

Human review and correction of machine-translated output to a specified standard. Effective post-editing requires source access, competence, adequate time and authority to reject unsuitable output.

Pragmatics

The study of meaning in context, including implication, presupposition, politeness and speech acts. Pragmatic meaning explains why preserving a proposition may still change what an utterance does.

Register

Language associated with a situation, relationship, profession or level of formality. Register choices shape voice and status, and a mismatch can alter credibility, intimacy or authority.

Relay interpreting

Interpreting through an intermediate language or communicator when no direct provision is available. Each relay adds delay and another point of interpretation, so coordination is essential.

Revision

Bilingual examination of a translation against its source to identify and correct problems. In formal workflows, revision may be performed by a second qualified person before later review.

Simultaneous interpreting

Interpreting delivered while the speaker continues, with a short lag. It requires specialised skill, suitable equipment or positioning, preparation and team arrangements that manage sustained cognitive load.

Skopos

The intended purpose of a translation commission, central to the approach associated with Hans Vermeer. Purpose organises choices but does not erase source constraints or professional responsibility.

Source and target texts

The source text is the material being translated. The target text is the resulting version for the receiving language and audience. Both may exist in several editions.

Translation memory

A database of source segments paired with previous translations. It supports consistency and reuse, but retrieved matches need contextual checking and can carry earlier errors forward.

Transliteration

Representation of writing from one script in another, usually according to a system. Transliteration handles written form rather than translating meaning, though names often require both decisions.

Go Deeper

These four works lead in different directions: public explanation, linguistic foundations, systematic practice and the politics of fluent translation. Together they move from the general reader's central puzzle to professional analysis and institutional power. They disagree in emphasis, which is useful. Read them as rival instruments rather than steps towards one doctrine.

David Bellos, Is That a Fish in Your Ear? Translation and the Meaning of Everything

Begin here for a witty, wide-ranging account of what translation does in ordinary life, literature, institutions and international organisations. Bellos is especially good at dismantling the fantasy that meaning exists as a fixed packet inside words. His confidence sometimes invites argument, which is part of the value. Read it for range, examples and the pleasure of seeing familiar questions turned around. It is the closest of these recommendations to a general reader's continuation of this book. Pair its examples with a published translation you already know and test where Bellos's general claims fit or fail.

Roman Jakobson, “On Linguistic Aspects of Translation”

This short 1959 essay supplies several ideas that still organise discussion: intralingual, interlingual and intersemiotic translation; equivalence in difference; and the problem created when languages require different grammatical information. It is compressed and belongs to its period, but it rewards slow rereading. Choose it when you want the argument reduced to first principles rather than a survey of professional practice. The essay also shows why translation cannot be explained through bilingual vocabulary lists alone. Its treatment of poetry is brief but especially useful for understanding why formal organisation can resist complete reformulation.

Mona Baker and Henry Jones, In Other Words: A Coursebook on Translation, fourth edition

Choose this for a structured route through meaning at word, grammatical, textual and pragmatic levels, joined to ethics and present-day practice. The fourth edition, published in 2026, updates a long-used coursebook for digital platforms and recent technological change. It is designed for students and practising translators, so the explanations lead into authentic examples and decisions rather than stopping at theory. Read it with a language pair you know, or use its framework to analyse competing published versions. Exercises and examples make it the practical choice for moving from informed reading towards disciplined translation work.

Lawrence Venuti, The Translator's Invisibility: A History of Translation, third edition

Venuti asks what fluent English translation hides: labour, difference, publishing norms and the power of a dominant language to make foreign writing sound domestically natural. The book's opposition between domesticating and foreignising tendencies has been influential and contested. Read it for the political and historical challenge, not as a rule that foreign-sounding prose is automatically superior. It will sharpen questions about credit, canon formation and the afterlife created by the version readers come to treat as transparent. Follow it with critics of Venuti if you want to test the limits of that opposition.

Notes and Sources

The Whole Thing in One Page and Why You Should Care

The organising model. The account of translation as a hierarchy of obligations under constrained mismatch is a synthesis rather than a quotation from one school. The hierarchy is analytical: it may be explicit, habitual or imposed by a medium or institution, rather than consciously scored by one translator. Jakobson (1959) establishes reformulation and “equivalence in difference”; Nida (1964) distinguishes orientations towards source form and receiver response; Reiss and Vermeer (2014) organise choice around text function and purpose; Baker and Jones (2026) analyse meaning across lexical, grammatical, textual and pragmatic levels; Venuti (2018) supplies the institutional and cultural critique of fluent transparency. Bellos (2011) provides an accessible counterweight to the idea that meaning is a packet stored inside words.

Consequence and language access. The medical, legal and public-service examples are illustrative types, not reported incidents. They show why propositional content, modality, role and pragmatic force must be checked separately. This book does not state one international rule for court or healthcare interpreting. Applicable duties, qualifications and procedures vary by jurisdiction and institution.

Unequal language resources. The contrast between data-rich and less-resourced languages is supported by Haddow et al. (2022), whose survey separates several sources of scarcity rather than treating “low-resource” as one condition. The discussion is about unequal infrastructure, data and support, not an inherent ranking of languages.

The Core Ideas

Levels of meaning. Baker and Jones (2026) support the treatment of word meaning, grammar, cohesion, pragmatics, register and professional decision-making. The examples involving indirect requests, presupposition and social address are constructed illustrations of established linguistic problems. They are not presented as universal features of every language pair.

Equivalence and purpose. Jakobson's three categories and phrase “equivalence in difference” come from his 1959 essay. Nida's formal and dynamic equivalence are presented as relative orientations, not the later caricature of word-for-word versus unlimited freedom. Reiss and Vermeer's skopos theory is likewise bounded by commission, source, audience and responsibility. The manuscript rejects the false inference that purpose can make distortion accurate.

Form, speech and signed languages. Díaz Cintas and Remael (2021) support the account of subtitle compression, timing, reading conditions and audiovisual constraint. Gile (2009) supports the account of interpreting as analysis under processing, memory and production pressure. Sutton-Spence and Woll (1999) support the description of British Sign Language as a natural language using manual and non-manual resources and spatial grammar. One signed language does not represent all others; the example shows that translation can cross modality as well as language.

Homer's opening. The Greek text, literal word order and close prose rendering of Iliad 1.1 were checked against the Open University's teaching edition, which shows that mēnin, anger, occupies the emphatic first position. Pope's opening couplet was checked against the 1715-1720 edition preserved by the Oxford Text Archive and a public-domain transcription. The comparison concerns order, expansion, rhyme and metre. It does not establish one definitive English meaning for the line.

Culture, domestication and foreignisation. Venuti (2018) supplies the two terms and the critique of fluency in English-language literary translation. The book treats them as tendencies, not moral absolutes. Borrowing, explanation, substitution, transliteration and dialect choice are assessed by function and consequence rather than assigned one permanent rank.

Systems, institutions and selection. Cheung (2006), Gutas (1998), Burnett (2001), Hill (2013), Lung (2004), Cronin (2003) and Rafael (1993) support the claim that translation is made within organised networks of patrons, oral mediators, scribes, reviewers, officials, publishers and political institutions. Hill documents Lin Shu's work with nineteen assistants on more than 180 book-length translations over about twenty years; Lung examines the oral collaborator's role in the Chinese David Copperfield. The account does not claim that all workshops shared one structure or that institutional power determined every sentence.

Afterlife and retranslation. Venuti (2018), Bellos (2011) and Cronin (2003) support the treatment of translation as part of circulation, canon formation and the receiving language. The claim that a widely read version can reshape expectations about an original is an interpretive causal account. It is difficult to isolate translation from publishing, education, criticism, prizes and adaptation, so the manuscript presents these forces together.

Working history and process

Rosetta Stone. The British Museum's record for EA24 dates the decree to 196 BCE and identifies hieroglyphic, Demotic and Greek inscriptions. The text is described as related official content, not a row-by-row vocabulary list. Its later role in decipherment depended on comparison, proper names and evidence beyond the stone itself.

The Septuagint. Jobes and Silva (2015) support the account of a Greek Jewish scriptural corpus formed over time, beginning with the Pentateuch in Hellenistic Egypt. The tradition of seventy-two agreeing translators is treated as a later authority story rather than a literal workshop log. Variation across books prevents the legend from being used as evidence of one uniform process.

Jerome. Letter 57, addressed to Pammachius, contains Jerome's defence of rendering sense for sense rather than word for word, with his stated qualification concerning scripture. The manuscript paraphrases the argument and does not claim that Jerome invented the distinction or ended the dispute.

Chinese Buddhist translation. Cheung (2006) supports the account of collaborative work around major Buddhist projects and the importance of recitation, explanation, drafting, checking and editorial participation. Surviving descriptions vary by project and period. The manuscript therefore rejects a fixed universal production line and uses Kumārajīva as a bounded example.

Graeco-Arabic scholarship. Gutas (1998) supports the eighth- to tenth-century setting, Abbasid patronage, work across Greek, Syriac and Arabic, and Hunayn ibn Ishaq's concern with manuscripts and revision. Baghdad was a major centre, but “the House of Wisdom” is not treated as one factory that produced the whole movement.

Toledo. Burnett (2001) supports a connected but varied twelfth-century translation programme involving different patrons, languages and working arrangements. Gerard of Cremona's Arabic-to-Latin Almagest is retained as one route into western European astronomy. The label “Toledo School of Translators” is flagged because it can imply a permanent academy more coherent than the evidence permits.

Print and Tyndale. Daniell (1994) supports the 1526 English New Testament, official hostility, Tyndale's execution in 1536 and the large later inheritance of his English phrasing. The manuscript uses this case to show circulation and afterlife, not to reduce Reformation translation to one man.

Empire and administration. Rafael (1993) provides a close case of translation, conversion and colonial rule in Tagalog society. Cronin (2003) supports the wider account of translation under global and institutional power. The text also records counter-use: language infrastructure could support anti-colonial argument and new literatures rather than serving rulers in one direction only.

The European Union. The European Union's official information page stated on 2 September 2026 that the EU has twenty-four official languages. The manuscript confines the claim to formal multilingual status and does not imply that every internal exchange is reproduced in every language.

Written workflow and standards. ISO 17100:2015 remains the published international standard for translation-service processes as of 2 September 2026, although a replacement project is under development. It covers core processes, resources and applicable specifications and excludes interpreting and raw machine output plus post-editing. ISO 18587:2017 remains the published standard for full human post-editing of machine-translation output and post-editor competence, while a replacement draft is under development. Standards describe process within scope; they do not prove that an interpretation is correct.

Nuremberg interpreting. Gaiba (1998) supports the account of simultaneous interpreting at the International Military Tribunal in 1945 and 1946, including the four working languages, equipment, teams and rotation. The trial is described as a highly visible large-scale deployment, not the invention of interpreting or the first use of simultaneous technique.

Machine-translation development. Papineni et al. (2002) introduced BLEU as an automatic corpus-level metric based on modified n-gram precision and a brevity penalty. Vaswani et al. (2017) introduced the Transformer architecture. The manuscript does not infer that one architecture, metric or benchmark measures translation quality in every setting.

Current systems and document-level failure. Jiang et al. (2023) supports the need for discourse-centred and document-level evaluation. Benkirane et al. (2024) and Tang, Chatterjee and Garg (2025) support the continued study of hallucinated or unsupported translation output across resource conditions. These papers establish live failure modes, not one fixed error rate for all language models.

WMT25 and the 2026 data vintage. Kocmi et al. (2025) reports the general shared task across thirty language pairs and sixty systems. Human evaluation covered half of those pairs, with test sets spanning four or five domains, and the task moved towards whole-document source input. These figures describe that benchmark only. WMT26 had released unseen documents and completed system submissions by July 2026, but its camera-ready papers and conference still lay ahead on the 2 September verification date. WMT25 was therefore the newest complete published general-task findings report. Neither year supports a universal ranking of models, languages or professional uses.

Evidence limits. Written translation leaves parallel texts, manuscripts, drafts, prefaces and institutional records. Interpreting before recording is less directly recoverable, and archives often preserve commissioners more readily than language workers. The manuscript treats unnamed labour and private motives cautiously rather than converting archival silence into a confident story.

What People Get Wrong

Word replacement and the perfect-version myth. Jakobson (1959), Nida (1964), Bellos (2011) and Baker and Jones (2026) support the rejection of one-to-one lexical transfer and context-free equivalence. The correction still allows plainly wrong translations and does not turn every variant into an equally good answer.

Fluency and accuracy. The machine-translation discussion draws on Jiang et al. (2023), Benkirane et al. (2024), Tang, Chatterjee and Garg (2025) and the bounded WMT25 results. Fluency is treated as one target-language property, not source fidelity. Human review is also presented as fallible and must be designed rather than invoked as a magic safeguard.

Untranslatability and invisibility. Bellos (2011) and Venuti (2018) provide contrasting pressure on these topics. A missing concise equivalent can require borrowing or expansion without making communication impossible. Visible difference can matter, but awkwardness by itself is not proof of ethical or accurate translation.

Back-translation. Brislin (1970) established a widely used cross-cultural research procedure. Behr (2017) analyses its shortcomings as a quality-testing method. The manuscript therefore retains back-translation as a diagnostic within team review and rejects surface return to the source wording as proof.

Use It

Practical lenses. The priority hierarchy, pressure-point comparison, confidence test, interpreting design, machine-risk ladder and paratext check are editorial syntheses derived from the sources above. They are decision aids, not legal, clinical or certification standards. The risk ladder deliberately asks about consequence, representative testing, secure handling, reviewer competence and stop authority rather than naming one tool as safe for a whole category.

Limits and gains. The final lens preserves both sides of the subtitle. Translation involves loss because forms and contexts do not align. It can also create access, terminology, literary possibilities and political connection. Cronin (2003), Bellos (2011) and Venuti (2018) support the broader account, while the precise balance remains text- and institution-specific.

Terms

The glossary follows standard usage in translation studies and professional practice. Jakobson (1959) supports intralingual and interlingual translation; Nida (1964) supports formal and dynamic equivalence; Reiss and Vermeer (2014) support skopos; Venuti (2018) supports domestication and foreignisation; ISO 17100:2015 and ISO 18587:2017 support revision and post-editing; Papineni et al. (2002) supports BLEU. CAT tool, termbase and translation-memory definitions follow the workflow described by Baker and Jones (2026). Category boundaries such as translation, adaptation and localisation remain partly conventional, so transparent labelling is given priority over false precision.

Go Deeper

Edition details for Bellos (2011), Jakobson (1959), Baker and Jones (2026) and Venuti (2018) were checked against publisher, university or scholarly catalogue records. The four recommendations have different jobs: accessible overview, compact linguistic foundation, systematic coursebook and political-historical challenge. They are not presented as a settled canon or as mutually consistent accounts.

Bibliography

Primary sources, standards and original evidence

British Museum. “The Rosetta Stone.” Collection and research material for museum number EA24. Accessed 2 September 2026.

European Union. “Principles, Countries, History: Languages.” Official European Union information page. Accessed 2 September 2026.

International Organization for Standardization. ISO 17100:2015. Translation Services: Requirements for Translation Services. Geneva: ISO, 2015, with Amendment 1, 2017. Status checked 2 September 2026.

International Organization for Standardization. ISO 18587:2017. Translation Services: Post-editing of Machine Translation Output: Requirements. Geneva: ISO, 2017. Status checked 2 September 2026.

Jerome. “Letter 57: To Pammachius on the Best Method of Translating.” In Nicene and Post-Nicene Fathers, second series, volume 6, edited by Philip Schaff and Henry Wace, translated by W. H. Fremantle, G. Lewis and W. G. Martley. Buffalo, NY: Christian Literature Publishing Company, 1893.

Homer. The Iliad: of Homer. Translated by Alexander Pope. 6 vols. London: W. Bowyer for Bernard Lintott, 1715-1720. Digital edition consulted through the Oxford Text Archive. Accessed 2 September 2026.

The Open University. “Introducing Homer's Iliad: Meter and Word Order.” Greek text, literal word order and prose rendering of Iliad 1.1-7. Accessed 2 September 2026.

Jakobson, Roman. “On Linguistic Aspects of Translation.” In On Translation, edited by Reuben A. Brower, 232-239. Cambridge, MA: Harvard University Press, 1959.

Nida, Eugene A. Toward a Science of Translating: With Special Reference to Principles and Procedures Involved in Bible Translating. Leiden: E. J. Brill, 1964.

Papineni, Kishore, Salim Roukos, Todd Ward and Wei-Jing Zhu. “BLEU: A Method for Automatic Evaluation of Machine Translation.” In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, 311-318. Philadelphia: Association for Computational Linguistics, 2002. DOI 10.3115/1073083.1073135.

Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. “Attention Is All You Need.” In Advances in Neural Information Processing Systems 30. 2017.

Kocmi, Tom, et al. “Findings of the WMT25 General Machine Translation Shared Task: Time to Stop Evaluating on Easy Test Sets.” In Proceedings of the Tenth Conference on Machine Translation, 355-413. Suzhou, China: Association for Computational Linguistics, 2025. DOI 10.18653/v1/2025.wmt-1.22.

Conference on Machine Translation. “WMT 2026” and “General Machine Translation Task.” Official task pages. Status checked 2 September 2026.

Modern studies and reference works

Baker, Mona, and Henry Jones. In Other Words: A Coursebook on Translation. Fourth edition. London: Routledge, 2026. DOI 10.4324/9781003440857.

Behr, Dorothée. “Assessing the Use of Back Translation: The Shortcomings of Back Translation as a Quality Testing Method.” International Journal of Social Research Methodology 20, no. 6 (2017): 573-584. DOI 10.1080/13645579.2016.1252188.

Bellos, David. Is That a Fish in Your Ear? Translation and the Meaning of Everything. London: Particular Books, 2011.

Benkirane, Kenza, Laura Gongas, Shahar Pelles, Naomi Fuchs, Joshua Darmon, Pontus Stenetorp, David Ifeoluwa Adelani and Eduardo Sánchez. “Machine Translation Hallucination Detection for Low and High Resource Languages using Large Language Models.” In Findings of the Association for Computational Linguistics: EMNLP 2024, 9647-9665. Miami, Florida, USA: Association for Computational Linguistics, 2024. DOI 10.18653/v1/2024.findings-emnlp.564.

Brislin, Richard W. “Back-Translation for Cross-Cultural Research.” Journal of Cross-Cultural Psychology 1, no. 3 (1970): 185-216. DOI 10.1177/135910457000100301.

Burnett, Charles. “The Coherence of the Arabic-Latin Translation Program in Toledo in the Twelfth Century.” Science in Context 14, nos. 1-2 (2001): 249-288. DOI 10.1017/S0269889701000096.

Cheung, Martha P. Y., ed. An Anthology of Chinese Discourse on Translation, Volume 1: From Earliest Times to the Buddhist Project. Manchester: St Jerome Publishing, 2006.

Cronin, Michael. Translation and Globalization. London: Routledge, 2003.

Daniell, David. William Tyndale: A Biography. New Haven: Yale University Press, 1994.

Díaz Cintas, Jorge, and Aline Remael. Subtitling: Concepts and Practices. London: Routledge, 2021.

Gaiba, Francesca. The Origins of Simultaneous Interpretation: The Nuremberg Trial. Ottawa: University of Ottawa Press, 1998.

Gile, Daniel. Basic Concepts and Models for Interpreter and Translator Training. Revised edition. Amsterdam: John Benjamins, 2009.

Gutas, Dimitri. Greek Thought, Arabic Culture: The Graeco-Arabic Translation Movement in Baghdad and Early Abbasid Society. London: Routledge, 1998.

Haddow, Barry, Rachel Bawden, Antonio Valerio Miceli Barone, Jindřich Helcl and Alexandra Birch. “Survey of Low-Resource Machine Translation.” Computational Linguistics 48, no. 3 (2022): 673-732. DOI 10.1162/coli_a_00446.

Hill, Michael Gibbs. Lin Shu, Inc.: Translation and the Making of Modern Chinese Culture. New York: Oxford University Press, 2013.

Jiang, Yuchen Eleanor, Tianyu Liu, Shuming Ma, Dongdong Zhang, Mrinmaya Sachan and Ryan Cotterell. “Discourse-Centric Evaluation of Document-level Machine Translation with a New Densely Annotated Parallel Corpus of Novels.” In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 7853-7872. Toronto, Canada: Association for Computational Linguistics, 2023. DOI 10.18653/v1/2023.acl-long.435.

Jobes, Karen H., and Moisés Silva. Invitation to the Septuagint. Second edition. Grand Rapids: Baker Academic, 2015.

Lung, Rachel Wai Chu. “The Oral Translator's ‘Visibility’: The Chinese Translation of David Copperfield by Lin Shu and Wei Yi.” TTR: Traduction, Terminologie, Rédaction 17, no. 2 (2004): 161-184. DOI 10.7202/013277ar.

Rafael, Vicente L. Contracting Colonialism: Translation and Christian Conversion in Tagalog Society under Early Spanish Rule. Durham, NC: Duke University Press, 1993.

Reiss, Katharina, and Hans J. Vermeer. Towards a General Theory of Translational Action: Skopos Theory Explained. Translated by Christiane Nord. London: Routledge, 2014.

Sutton-Spence, Rachel, and Bencie Woll. The Linguistics of British Sign Language: An Introduction. Cambridge: Cambridge University Press, 1999.

Tang, Zilu, Rajen Chatterjee and Sarthak Garg. “Mitigating Hallucinated Translations in Large Language Models with Hallucination-focused Preference Optimization.” In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 3410-3433. Albuquerque, New Mexico: Association for Computational Linguistics, 2025. DOI 10.18653/v1/2025.naacl-long.175.

Venuti, Lawrence. The Translator's Invisibility: A History of Translation. Third edition. Abingdon: Routledge, 2018.