Machine Translation Post-Editing and Quality Evaluation in Pashto and Dari

What is machine translation post-editing (MTPE)?

Machine translation post-editing (MTPE) is the human correction of machine translation output — from a neural MT engine or a large language model — until it meets an agreed quality level. The post-editor works from the source text, segment by segment, and fixes meaning, terminology, grammar and style as the specification requires. ISO 18587 sets the requirements for full post-editing and describes light post-editing in its Annex B. Human translation without MT falls under ISO 17100, which places raw MT plus post-editing outside its scope.

Key terms

Machine translation post-editing (MTPE)

Human correction of machine or AI translation output to an agreed quality level.

Light post-editing

Post-editing that makes output accurate and understandable without polishing style; described in ISO 18587, Annex B.

Full post-editing

Post-editing to a quality comparable to human translation; the subject of the ISO 18587 requirements.

ISO 18587

International standard (2017) for full human post-editing and post-editor competences; under revision as ISO/DIS 18587.

ISO 5060

International guidance (2024) for evaluating human, post-edited and raw machine translation with error types and penalty points.

MQM

Multidimensional Quality Metrics: a shared error typology and scoring model for analytic translation quality evaluation.

Critical error

An error that could cause harm or legal, safety or financial consequences. It fails the sample whatever the score.

Quality estimation (QE)

Automatic prediction of translation quality without a reference translation, used to route segments to post-editing.

COMET

A neural metric trained to predict human judgments of translation quality.

HTER

Human-targeted translation edit rate: the share of words changed in post-editing, a measure of effort.

Afghan Dari

The standard variety of Persian used in Afghanistan, distinct from Iranian Persian (Farsi) in vocabulary, spelling and calendar terms.

Solar Hijri calendar

The solar calendar on most Afghan records, with Afghan month names such as Hamal in Dari and Wray in Pashto.

Why raw machine translation fails in Pashto and Dari

Pashto and Dari are low-resource languages for machine translation: engines see far less Afghan text than English, French or Iranian Persian. Output often reads fluently and is still wrong. These are the failure modes we test for on every engagement.

1. 01

   Dari that is really Iranian Persian

   Engines trained mostly on Iranian text choose Iranian words — بیمارستان for hospital, استان for province — where Afghan readers expect شفاخانه and ولایت.

   Dari
2. 02

   Substituted Pashto letters

   Pashto uses letters Persian and Arabic do not: ټ ډ ړ ښ ږ ځ څ ڼ ګ ې ۍ ئ. Several carry gender, number or verb agreement, so a look-alike substitute changes the grammar.

   Pashto
3. 03

   Pakistani usage in Afghan Pashto

   Output mixes in Pakistani and Urdu-influenced vocabulary, such as شناختي کارت for the national ID where Afghan readers write تذکره.

   Pashto
4. 04

   Pronouns, gender and number

   A U.S. court denied a Pashto-speaking refugee’s asylum claim after a translation tool turned “I” into “we” in her written statement, Rest of World reported in 2023.

   Both
5. 05

   Kinship terms and military ranks

   Relatives’ terms vary by region and ranks have no one-to-one equivalents. Both decide outcomes in Special Immigrant Visa (SIV) and asylum files.

   Both
6. 06

   Dates and calendars

   Most Afghan records use the Solar Hijri calendar with Afghan month names — حمل in Dari, وری in Pashto — not Iran’s فروردین. Administrative documents issued since 2022 often use the lunar Hijri calendar.

   Both
7. 07

   Encoding and right-to-left layout

   Arabic and Persian code points for ی and ک look identical but break search and matching. Missing zero-width non-joiners and bidirectional errors garble mixed Latin and Arabic-script text.

   Both
8. 08

   LLM omissions and additions

   Large language models can drop a clause, add a sentence, or drift into Urdu, Arabic or Iranian Persian words, and still read fluently.

   LLM

Pashto and Dari MTPE and translation quality evaluation services

Six services, used alone or as one program. Each starts from a written specification: your content, your readers, the quality level you need and how it will be measured. Every service keeps a qualified human in the loop.

• MT readiness assessment

  Find out where machine translation is safe to use. We run a representative sample of your content through your engine or the candidates, evaluate the output, and return a go or no-go for each content type with the post-editing level it needs.

  You receive: Readiness report, content risk-tier map, recommended workflow

• Light post-editing

  For internal, high-volume and time-sensitive content. We correct meaning errors, key terminology and anything unsafe or offensive, and leave correct machine phrasing in place. Light post-editing is described in ISO 18587, Annex B.

  You receive: Post-edited files, change log

• Full post-editing

  For published, customer-facing and regulated content. Output comparable to human translation, consistent with your glossary and Afghan locale conventions, in a process aligned with ISO 18587. Second-linguist revision on request.

  You receive: Post-edited files, revision record, term base updates

• Translation quality evaluation (ISO 5060)

  Independent evaluation of machine, post-edited or human translation. Two native evaluators annotate every error by type and severity, a senior reviewer adjudicates, and you receive the score, the quality rating and the annotated evidence.

  You receive: Scorecard, annotated segments, root-cause analysis

• MT engine and LLM benchmarking

  Blind, side-by-side comparison of the MT engines, custom models and LLMs you are considering — such as GPT, Gemini, Claude, Llama-family models and NLLB-200 — on your own content and language directions. Repeated when a model updates, so quality cannot slip unnoticed.

  You receive: Benchmark report, selection recommendation, regression test set

• Evaluation datasets and metric calibration

  For AI labs and MT developers: error-annotated Pashto and Dari test sets, gold-standard reference translations, blind side-by-side system comparisons, and COMET, chrF and quality-estimation thresholds calibrated against human scores. Evaluation data only; training corpora are a separate service.

  You receive: JSONL datasets, annotation guidelines, agreement statistics

How we evaluate Pashto and Dari translation quality (ISO 5060 and MQM)

ISO 5060:2024 is international guidance for evaluating translation output — human translation, post-edited machine translation and raw machine translation. Evaluators mark each error by type and severity; penalty points produce an error score, and the score maps to a quality rating agreed before the evaluation starts. Localization teams often call this linguistic quality assurance (LQA). ISO 5060 is guidance, not a certification scheme, and it does not cover interpreting.

How we deliver a Pashto and Dari MTPE program

Five phases, each ending in something you can inspect. A program can stop after the assessment if machine translation is not the right tool.

1. 01

   Scope

   NDA, content inventory, readers, systems in use and success criteria.

   Output: Written specification

2. 02

   Pilot

   Readiness assessment and, if needed, a blind engine or LLM comparison.

   Output: Go or no-go per content tier

3. 03

   Configure

   Glossary, style guide, Afghan locale conventions, prompts and quality-estimation thresholds.

   Output: Configuration pack

4. 04

   Produce

   Light or full post-editing by tier, with revision where specified.

   Output: Post-edited files in your formats

5. 05

   Measure

   ISO 5060 evaluation on a sampling schedule and after every engine or model update.

   Output: Scorecards and trend report

Ways to engage

• Readiness assessment

  A fixed-scope answer to whether, where and how to use machine translation.

• Managed MTPE program

  Post-editing by content tier, with continuous evaluation.

• Evaluation retainer

  Scheduled ISO 5060 audits of your MT, vendor or in-house output.

• Dataset or benchmark project

  Annotated Pashto and Dari data and human evaluation for model builders.