Translate | Reported Speech and Attribution — Preserve Who Said What and Who Owns the Claim

If you want to translate reported speech, quotations and attribution accurately, the central question is not simply how to change pronouns or tense. The deeper job is to preserve who said what, who is responsible for the claim, how directly the source endorses it, and whether the words are quoted, paraphrased, inferred or merely attributed. A translation that blurs those distinctions can turn someone else’s claim into the author’s claim.

This matters in news translation, academic translation, legal translation, business translation, historical writing and everyday communication. Phrases such as according to, said, claimed, reported, argued, suggested, alleged, warned, admitted, noted and appeared to are not interchangeable labels for “someone spoke.” They tell readers where information comes from and how the writer positions that information.

A reliable method for translating reported speech is to build an attribution chain before rewriting the sentence. Identify the original speaker, the reporting writer, the audience, the exact proposition being attributed, the reporting verb, the level of certainty, and any shift in person, time or place. Then translate the chain so the same person still owns the same claim in the target language.

Quick answer: what is attribution in translation?

Attribution is the language that tells the reader who is responsible for a statement, observation, opinion or piece of evidence.

Compare:

The bridge is unsafe.

The engineer said the bridge is unsafe.

Residents claim the bridge is unsafe.

According to the inspection report, the bridge is unsafe.

The bridge appears to be unsafe.

These sentences may concern the same bridge, but they do not make the same communicative move.

The first presents the proposition directly.

The second attributes it to an engineer.

The third attributes it to residents and uses claim, which may create more distance.

The fourth links it to a document.

The fifth presents an inference or appearance rather than a sourced quotation.

Translation must preserve those differences.

Direct speech and reported speech are different structures

Direct speech reproduces a speaker’s words as a quotation.

Lina said, “I will finish the report tomorrow.”

Reported speech integrates the content into another speaker’s sentence.

Lina said that she would finish the report the next day.

Several elements may shift:

• I → she;
• will → would;
• tomorrow → the next day.

But these changes are not a universal formula. Some languages do not backshift tense in the same way. Some preserve deictic expressions if the reporting situation still supports them.

The translator must preserve the relationship between original utterance and reporting context.

Who owns the proposition?

This is the first question to ask.

Source:

The company said the delay was unavoidable.

The proposition the delay was unavoidable belongs to the company.

A mistranslation that effectively becomes:

The delay was unavoidable, the company said.

may still preserve attribution.

But a shortened translation:

The delay was unavoidable.

removes the source and turns an attributed justification into a direct assertion.

That is not a minor style change. It changes epistemic responsibility.

Attribution is part of meaning

Writers use attribution to separate voices.

Consider:

Researchers found that the treatment improved recovery time.

The writer reports a research finding.

Compare:

The treatment improved recovery time.

The second sentence sounds more like the writer’s own unqualified statement.

In academic and journalistic translation, deleting attribution can increase certainty and erase evidence structure.

A useful rule is:

never treat source attribution as disposable introductory wording.

Reporting verbs carry stance

English has many reporting verbs.

• said
• stated
• reported
• claimed
• argued
• suggested
• observed
• admitted
• denied
• warned
• promised
• alleged
• acknowledged
• insisted
• confirmed

Each verb packages the speech event differently.

She said the plan would work.

is relatively neutral.

She claimed the plan would work.

can imply that the statement is disputed or not independently accepted.

She confirmed the plan would work.

suggests prior uncertainty and subsequent verification.

She insisted the plan would work.

adds persistence or resistance to doubt.

A translation that maps all of these to one generic “said” may flatten stance.

“Claim” is not always neutral

The word claim deserves special attention.

In academic prose:

Smith claims that…

may simply introduce a proposition associated with Smith.

In journalism or ordinary speech:

The company claims that…

can imply distance or unresolved verification.

Context matters.

The translator should not automatically strengthen claim into “proved” or weaken it into completely neutral “said.”

Look at the surrounding tone.

“Allege” marks unproven accusation

The prosecution alleged that the funds were diverted.

The verb alleged indicates that the statement is an accusation or assertion not presented as established fact in that sentence.

Replacing it with:

The prosecution showed that…

would be a serious distortion.

In legal and news translation, attribution verbs often carry procedural status. Precision is essential.

“Admit” presupposes something difficult or unwelcome

The company admitted the figures were incorrect.

Admitted does more than report speech. It frames the proposition as something the speaker acknowledged despite potential reluctance or disadvantage.

A neutral translation equivalent to “said” would lose that stance.

But an overly dramatic target verb meaning “confessed” could add moral or criminal force.

The translator should preserve the degree of markedness.

“Acknowledge” is not always “admit”

Compare:

The report acknowledges several limitations.

This often means the report recognises or openly states them.

It does not necessarily suggest wrongdoing.

Translating acknowledge with a word carrying guilt or confession can overstate the source.

Reporting verbs sit on a subtle scale of responsibility, evidence and attitude.

“Suggest” can report or hedge

The results suggest that sleep affects performance.

Here suggest signals an evidential relationship: the data supports an interpretation without claiming absolute proof.

She suggested that we leave early.

Here it means proposed or recommended.

Same verb, different reporting function.

Translation must identify the semantic frame before selecting the target verb.

“According to” creates a source frame

According to the annual report, revenue increased.

The sentence attributes the information to a document.

This can be neutral sourcing.

It may also create distance:

According to the company, the problem has been fixed.

The writer may not independently endorse the statement.

Do not automatically translate according to as if the writer were certifying the proposition.

It marks source, not truth.

Nested attribution

Complex texts often contain sources inside sources.

The newspaper reported that officials said the system had passed all tests.

There are at least three voices:

1. the current writer;
2. the newspaper;
3. the officials.

Who actually claims the system passed all tests?

The officials.

The newspaper reports their statement.

The current writer reports the newspaper.

A careless translation may collapse this into:

The system passed all tests.

That deletes two layers of source distance.

For sensitive texts, map nested attribution explicitly.

The attribution ladder

A useful review tool is to write:

Writer → Source A → Source B → Proposition

Example:

Writer → investigation report → witness → “the alarm did not sound”

Now check the translation.

Does the target still make clear that the witness, not the report author, directly experienced the event?

Attribution ladders prevent voice collapse.

Quotations require quotation ownership

Direct quotes create a strong responsibility boundary.

“The system is safe,” the manufacturer said.

The statement belongs to the manufacturer.

If the translation converts the quote into unmarked prose, that boundary weakens.

Likewise, punctuation and quotation conventions may differ across languages, but the ownership relation should remain clear.

The target reader should be able to answer:

Whose words are these?

Partial quotation

Journalists may quote only one phrase.

The minister described the measure as “temporary and proportionate.”

The quoted adjectives are presented as the minister’s exact wording.

A translation may have to decide whether to translate the quotation normally, preserve original-language wording, or include both depending on publication convention.

But the distinction between quoted language and the reporter’s framing should remain visible.

Scare quotes and distancing quotation marks

Quotation marks can signal more than direct speech.

The company promised a “seamless” transition.

The quotation marks may indicate that seamless is borrowed language or that the writer is distancing themselves from the description.

A translator who removes the marks may convert a sceptical reference into endorsement.

A translator who adds quotation marks where none exist may introduce scepticism.

Punctuation can carry attribution and stance.

Reported questions

Direct:

She asked, “Where is the file?”

Reported:

She asked where the file was.

The target language may restructure word order, tense and pronouns.

But the speech act remains a question.

A common learner error is to preserve direct-question syntax inside reported speech when the target grammar does not.

The translator’s job is not to preserve punctuation mechanically but to preserve the questioning event and content.

Reported commands and requests

Direct:

“Send the invoice today,” he said.

Reported:

He told her to send the invoice that day.

Now compare:

“Could you send the invoice today?” he asked.

Reported:

He asked her to send the invoice that day.

Both can lead to similar actions, but the interpersonal force differs.

told can sound directive.

asked preserves a request.

Translation should preserve speech act, not just outcome.

Promises and commitments

She promised to deliver the files by Friday.

The reporting verb carries a commitment.

If translated as:

She said she would deliver the files by Friday.

the future event remains, but the promise disappears.

That may matter in contracts, negotiations and disputes.

Words describing speech acts can create obligations.

Warnings

The engineer warned that the structure could fail.

Warned tells us the speech was intended to alert someone to danger.

Neutralising it:

The engineer said the structure could fail.

loses the communicative function.

Conversely, translating a neutral said as warned adds urgency not present in the source.

Denials

The company denied sharing the data.

This does not mean the writer has proved the company did not share it.

It means the company rejected that proposition.

A bad translation can accidentally turn a denial into factual exoneration.

Keep denial attached to the denying speaker.

Confirmation

Officials confirmed that the airport had reopened.

Confirmed often indicates that the information was uncertain or unofficial before the statement.

The verb can also suggest authority.

The translator should preserve that evidential role where the target language permits.

Do not treat confirmed as a decorative synonym for said.

Anonymous sources

News texts use expressions such as:

a source familiar with the matter said

officials speaking on condition of anonymity said

according to people briefed on the discussion

These constructions tell readers about the type and limits of sourcing.

A translator should not simplify them to:

officials said

unless the source text actually identifies officials.

Source anonymity is part of evidence quality.

Institutional voice

Documents often speak through institutions.

The ministry said…

The court held…

The report concluded…

The committee found…

These are conventional ways of assigning propositions to institutional outputs.

A literal target rendering may sound unnatural if the target language prefers “the ministry stated” or “the court ruled,” but the institutional owner of the proposition must remain intact.

Author versus evidence

Academic prose often moves between an author’s interpretation and cited evidence.

The survey found that response times declined. The authors argue that this reflects improved training.

Two voices:

• the survey result;
• the authors’ interpretation.

A translation that merges them may make the interpretation look directly measured.

Strong academic translation preserves evidence layers.

“Find” can be evidential or judicial

The study found a correlation.

The court found the defendant liable.

These are not identical senses.

In the first, found refers to research result.

In the second, it refers to a formal legal determination.

Context determines the target term.

Reporting language is domain-sensitive.

“Conclude” and inference

The researchers concluded that the intervention was effective.

The sentence says the researchers reached that judgment.

It does not necessarily mean the current author endorses it fully.

Preserving the verb keeps the inference attached to the source.

Reported speech and tense backshift

English often shifts tense when reporting a past utterance.

Direct:

“I am tired.”

Reported:

She said she was tired.

Direct:

“I have finished.”

Reported:

She said she had finished.

But backshift depends on context.

The teacher said that water boils at 100°C under standard atmospheric pressure.

A general truth may retain present tense.

Languages differ in sequence-of-tense systems, so translators should not mechanically copy English backshift.

Preserve event time relative to speech time.

Temporal anchors in reported speech

Direct:

“I will finish tomorrow.”

Reported two days later:

He said he would finish the next day.

The translation needs to preserve when the finishing was expected relative to the original statement.

If the target text uses an absolute date instead, verify that the date is correct.

Reported speech combines attribution with temporal deixis.

Spatial anchors in reported speech

Direct:

“Bring the package here.”

Reported elsewhere:

She told him to take the package there.

Depending on language and viewpoint, bring/take and here/there may shift.

The proposition and action stay the same while the deictic frame changes.

This is why attribution and viewpoint should be reviewed together.

Pronoun shifts

Direct:

“We will send you our response.”

Reported by the recipient:

They said they would send us their response.

Every pronoun changes.

A translator who loses track of speaker and recipient can reverse ownership.

For complex dialogue, create a participant table.

Participant table

Example:

Then interpret pronouns against roles.

Even if you never show the table to the reader, it stabilises translation.

Attribution in headlines

Headlines compress grammar.

Minister: Talks to continue

Company says outage resolved

Study finds link between sleep and memory

A translator may need to expand syntax for natural target headlines.

But do not remove the attribution.

Outage resolved

is stronger than:

Company says outage resolved.

Headlines are especially vulnerable to source deletion because brevity pressures the translator.

Attribution in social media posts

Posts often embed screenshots, quotes and reactions.

“This is unacceptable,” she wrote.

He reposted the claim without comment.

Users reported seeing an error message.

Different actions carry different endorsement levels.

Reposting is not automatically endorsement.

Reporting that users observed something is not the same as verifying it.

Translation should preserve action type.

Attribution in legal translation

Legal texts distinguish:

• alleges;
• submits;
• argues;
• testifies;
• states;
• admits;
• denies;
• finds;
• holds;
• orders.

These verbs can indicate procedural role.

For example:

The claimant submits that…

introduces a party’s argument.

The court finds that…

introduces a judicial determination.

Confusing them can change legal status.

Attribution in scientific translation

Scientific writing distinguishes observation, inference and prior literature.

We observed…

The data indicate…

Previous studies suggest…

We hypothesise…

The model predicts…

Each phrase assigns the proposition to a different evidential mechanism.

A translation that turns all of them into “shows” can overstate certainty.

Preserve epistemic source.

Attribution and numbers

Consider:

The company estimates losses of $2 million.

That is not necessarily the same as:

Losses are $2 million.

The number is an estimate owned by the company.

Similarly:

Police said 5,000 people attended.

and:

5,000 people attended.

differ in sourcing.

Quantitative translation still requires attribution discipline.

Hearsay and second-hand information

She heard that the office would close.

He was told that the office would close.

It was rumoured that the office would close.

These structures mark information as indirect.

The target language may have evidential particles or grammatical forms for hearsay.

Do not translate them as direct fact.

Evidential adverbs

Words such as:

• reportedly;
• apparently;
• allegedly;
• supposedly;
• presumably;
• evidently

can mark source, inference or speaker distance.

The device reportedly failed during testing.

The writer is not necessarily claiming direct observation.

The device apparently failed during testing.

This may reflect inference from available evidence.

These distinctions are subtle but meaningful.

The danger of certainty inflation

A common translation failure is to turn:

appears to

into:

is

or:

reportedly

into no marker at all.

The result becomes more certain than the source.

Certainty inflation is especially risky in medicine, law, science, finance and journalism.

A good reviewer checks not only nouns and verbs but also source and certainty markers.

The danger of scepticism inflation

The opposite error also occurs.

Neutral source:

The spokesperson said the system was operational.

Over-sceptical translation:

The spokesperson claimed the system was operational.

If the target verb strongly implies doubt, the translator has inserted editorial stance.

Do not make the target more suspicious than the source.

Attribution and passive voice

It was reported that the server had failed.

The source is unspecified.

A translator may be tempted to add an agent:

Engineers reported that…

unless the context names them, that is invention.

The passive can deliberately background or omit the source.

Preserve absence when absence is meaningful.

Attribution and source ambiguity

According to Alex, Jordan said the file was incomplete.

Who believes the file is incomplete?

Jordan said it.

Alex is the source through whom the writer knows Jordan said it.

The current writer may endorse neither.

Nested attribution needs careful parsing.

A practical seven-step method

Step 1: identify the current narrator

Who is presenting the sentence to the reader?

Step 2: identify every source

Person, institution, document, dataset, witness or anonymous source.

Step 3: identify the proposition

What exactly is being attributed?

Step 4: identify the reporting verb

Said, claimed, found, argued, warned, denied, confirmed?

Step 5: identify stance and certainty

Does the source endorse, distance, doubt, infer or merely report?

Step 6: resolve pronoun, time and place shifts

Who is I/you/we? When is tomorrow? Where is here?

Step 7: test ownership in the translation

Ask:

If a reader challenges this statement, who would the target text imply is responsible for it?

That answer should match the source.

The ownership test

Source:

Analysts said the policy could reduce costs.

Target A:

The policy could reduce costs, analysts said.

Ownership preserved.

Target B:

The policy will reduce costs.

Ownership lost and certainty strengthened.

The ownership test catches both attribution deletion and modality drift.

The evidence-distance test

Place sentences on a rough scale:

I observed X.

The data indicate X.

Researchers reported X.

A witness said X.

X was rumoured.

These are not a universal ranking of truth. They are different source relations.

The translation should not move a proposition to a different evidential position without reason.

The quotation test

For every direct quotation, ask:

• Is the same person speaking?
• Is the same material inside quotation marks?
• Did I accidentally put the reporter’s paraphrase inside the quote?
• Did I accidentally move quoted evaluation into narrator voice?
• Did punctuation conventions change ownership?

This is especially important when quotations are split by attribution.

Split quotations

“The proposal is workable,” the chair said, “but the timeline is unrealistic.”

The two quoted parts belong to one speaker.

A target language may place the reporting clause differently.

Reorder if necessary, but keep the contrast and speaker ownership intact.

Translation of interviews

Interviews use alternating voices.

Labels such as:

Interviewer:

Participant:

are attribution devices.

If labels are removed or mistranslated, turns can be assigned to the wrong speaker.

For long interviews, speaker identity is a structural feature, not formatting trivia.

Translation of meeting minutes

Minutes often distinguish:

• reported;
• noted;
• agreed;
• decided;
• requested;
• actioned.

The committee noted the risk.

does not mean:

The committee accepted the risk.

The committee agreed to review the risk.

creates a decision.

Translation should preserve organisational action status.

Translation of customer complaints

The customer said the package arrived damaged.

This attributes the condition to the customer’s report.

If the company has not independently inspected the package, translating it as:

The package arrived damaged.

may remove evidential distance.

Operational records often depend on who reported what.

Translation of witness statements

Witness translation requires unusual discipline.

A translator should not “improve” uncertain language.

I think the car was blue.

must not become:

The car was blue.

I heard a noise that sounded like glass breaking.

must not become:

I heard glass break.

The witness’s uncertainty and perception boundary are evidence.

AI translation and attribution collapse

AI systems tend to optimise fluency. In summarising or translating, they may delete repeated phrases such as:

he said

according to the report

researchers suggest

because the text seems smoother without them.

That can be dangerous.

When using AI for translation, instruct the system to preserve every attribution and uncertainty marker, then review them manually.

Search the target for source-bearing language.

Translation memory and reporting verbs

Translation memory can improve consistency, but reporting verbs are context-sensitive.

A stored translation of claim from academic writing may be too sceptical in another domain.

A stored translation of submit from general English may be wrong in legal argument.

Terminology consistency should not override sense.

A source ledger

For complex documents, build a source ledger.

Example:

Then review the target document against the ledger.

This is especially useful when a document has many voices.

Attribution and fairness

Accurate attribution protects readers and speakers.

It prevents:

• opinions from becoming facts;
• allegations from becoming findings;
• findings from becoming rumours;
• denials from becoming exonerations;
• uncertain estimates from becoming exact values;
• quoted views from becoming author views.

That is why attribution is not a cosmetic feature. It is part of informational integrity.

A connection to vocabulary

Reporting verbs are a strong example of vocabulary depth.

Knowing that say, claim, argue, admit, confirm, allege all involve communication is not enough.

A learner must understand stance, evidence, register, collocation and social function.

The Vocabulary Learning Hub supports that deeper network of distinctions.

A connection to How English Works

Reported speech combines clause structure, tense, pronouns, modality, deixis and discourse voice.

The broader How English Works architecture helps explain those systems before translation carries them into another language.

The more clearly a translator can model English voice structure, the less likely attribution is to drift.

A connection to the master translation architecture

This article sits inside the Master Art of Translation system.

The master problem is preserving meaning relationships.

Attribution is one of those relationships:

speaker → proposition → reporting frame → reader

A good translation preserves the chain.

Final checklist

Before delivering a translation containing reported speech, ask:

1. Who owns each proposition?
2. Which words are directly quoted?
3. Which material is paraphrased?
4. What does the reporting verb imply?
5. Is the source neutral, sceptical, inferential or committed?
6. Did pronouns shift correctly?
7. Did time and place references shift correctly?
8. Did I preserve uncertainty?
9. Did I accidentally convert an allegation into a fact?
10. Did I accidentally convert an author’s statement into somebody else’s?
11. Did I remove an anonymous-source marker?
12. Did I strengthen “said” into “confirmed” or “claimed” without evidence?
13. Could the target reader identify who is responsible for every important claim?

If those answers are controlled, the translation is much more likely to preserve informational responsibility.

Final principle

Reported speech translation is not a tense-conversion exercise.

It is a problem of voice architecture.

You are carrying a claim from one language to another while keeping attached:

• the speaker;
• the source;
• the speech act;
• the level of certainty;
• the evidence distance;
• the original time and place anchors;
• the difference between quotation and paraphrase.

When those relationships survive, the translation does more than sound natural.

It tells the reader not only what was said, but who said it, how it was said, and how far the current writer is willing to stand behind it.

Worked example: one event, four attribution levels

Imagine a factory alarm sounded at 2 p.m.

Version 1:

The alarm sounded at 2 p.m.

This presents the event directly.

Version 2:

The operator log records that the alarm sounded at 2 p.m.

This ties the event to a documentary source.

Version 3:

A technician said the alarm sounded at about 2 p.m.

This attributes the event to a person and adds approximation.

Version 4:

Workers reportedly heard an alarm shortly after 2 p.m.

This introduces indirect reporting and changes the event from “alarm sounded” to “workers heard an alarm.”

These sentences are related, but none should be translated into another. A translator must preserve the evidential route actually chosen by the source.

Worked example: allegation versus finding

Source A:

The regulator alleged that the supplier had concealed defects.

Source B:

The regulator found that the supplier had concealed defects.

The difference is procedural.

In Source A, concealment is asserted as an allegation.

In Source B, the regulator has reached a finding.

A translation that uses the same verb for both can erase legal status.

Now add:

The supplier denied concealing defects.

That denial is yet another speech act. The target should preserve the three separate positions:

• regulator alleges;
• regulator finds;
• supplier denies.

Do not compress a dispute into a single narrator sentence.

Worked example: academic source layering

Consider:

Previous studies suggest that bilingual experience may improve some forms of attentional control, although recent reviews caution that the effect is inconsistent.

There are multiple attribution and modality markers:

• previous studies = source group;
• suggest = evidence supports but does not establish absolutely;
• may improve = modal possibility;
• some forms = limited scope;
• recent reviews = second source group;
• caution = warning against overgeneralisation;
• inconsistent = limitation.

A translation that becomes:

Bilingual experience improves attentional control.

has deleted nearly every control built into the source.

This is why academic translation should preserve claim architecture, not just topic vocabulary.

Worked example: business communication

Source:

The vendor says the patch should resolve the issue, but our engineering team has not yet verified the fix.

There are two epistemic layers.

The vendor supplies the expectation.

The internal engineering team has not confirmed it.

If the translation becomes:

The patch will resolve the issue.

the target text creates certainty that nobody in the source claimed.

A faithful translation should keep the vendor attribution and the pending verification status.

Attribution matrix for review

For difficult material, reviewers can use four columns:

• Speaker/source
• Proposition
• Reporting verb or evidential marker
• Status in target translation

Example:

• Vendor | patch should resolve issue | says | preserved
• Engineering team | fix not yet verified | not yet verified | preserved
• Reporter | no independent endorsement | structural | preserved

This simple matrix is useful for news, minutes, research summaries and incident reports.

Attribution across paragraph boundaries

Source ownership can continue after the explicit reporting clause.

Example:

Dr Tan said the new method reduced processing time. It also required less memory and performed better on noisy data.

Who owns the second sentence?

Context may imply that these claims are still part of Dr Tan’s explanation.

A translator who treats the second sentence as the narrator’s independent assertion may change the source structure.

Look for paragraph-level scope of attribution, especially after verbs such as said, argued, explained and reported.

When repeating attribution is safer

Some target languages or editorial styles prefer repeated source markers to avoid ambiguity.

If a long paragraph moves between the author’s voice and a quoted expert’s position, adding a natural source reminder may preserve clarity even when the source language relies on looser continuation.

The goal is not word-count symmetry.

The goal is voice clarity.

Any such adjustment should remain faithful to the original ownership structure and should not add new evaluation.

When attribution can be implicit

The reverse also occurs. A language may allow source marking through morphology, particles or evidential forms rather than repeated reporting clauses.

A literal translation that repeats she said after every sentence could sound heavy and unnatural.

You can vary form while preserving function.

What must remain stable is the reader’s ability to tell:

whose proposition is this?

Final quality-control pass: search the source-bearing words

Before delivery, search the source and target for terms such as:

• said;
• according to;
• reported;
• reportedly;
• claimed;
• alleged;
• argued;
• suggested;
• found;
• concluded;
• confirmed;
• denied;
• admitted;
• warned;
• estimated;
• appeared;
• apparently;
• quoted;
• wrote;
• testified.

Do not expect one-to-one counts. Languages differ.

Use the search to locate places where evidence ownership might have disappeared.

That final pass is fast, and in high-stakes translation it can prevent some of the most consequential meaning errors.