Machine translation, post-editing and localisation

In short

Machine translation (MT) is software that converts text from a source language into a target language; post-editing (MTPE) is a human translator correcting that raw output to an agreed quality level; localisation (l10n) is adapting a whole product — strings, images, dates, currency, legal text — so it feels native in a target locale. For Indian language graduates this is where the paid work now is: almost no commercial Hindi, Tamil, Bengali or Malayalam job today starts from a blank page, and rates, deadlines and job titles ("post-editor", "language lead", "localisation QA") are built around MT. University papers on translation and media examine MT typology (RBMT/SMT/NMT), evaluation metrics, post-editing levels and the ISO standards; interviews at Bhashini partners, IT services firms, OTT platforms and e-commerce localisation vendors test exactly this. It also matters legally and constitutionally — the Official Languages Act, 1963 and its Rules require bilingual government documents, and the National Language Translation Mission (Bhashini) is the state's answer to that volume.

Explained

What MT is, and the four generations of it

• Rule-based MT (RBMT): hand-written morphological, syntactic and transfer rules plus bilingual dictionaries. Predictable, explainable, terminology-faithful; brittle, expensive to build. Indian examples: Anusaaraka, Mantra-Rajbhasha (C-DAC/CDAC-ER), AnglaBharati.
• Example-based MT (EBMT): translates by analogy with stored aligned sentence pairs; the conceptual ancestor of translation memory.
• Statistical MT (SMT): learns phrase translation probabilities and a target-language model from parallel corpora (Moses-era, roughly 2005-2016). Output is fluent in fragments, incoherent across a long sentence; word order suffers badly for English→Hindi because English is SVO and most Indian languages are SOV.
• Neural MT (NMT): an encoder-decoder neural network, now almost always a Transformer (self-attention, 2017), trained end-to-end on parallel sentences; it emits a whole sentence conditioned on the whole source. Far more fluent, handles word order and agreement, but hallucinates — invents plausible text when uncertain — and silently drops negation, numbers and named entities.
• Variants you should be able to name: *multilingual / many-to-many NMT* (one model, many language pairs, so low-resource Indian languages borrow strength from high-resource ones — the basis of IndicTrans2); *transfer learning and fine-tuning* on a client's own corpus; *adaptive MT* that learns from each post-edit in real time; *LLM-based translation*, where a general large language model is prompted with a glossary and style guide and is better at context, tone and explanation but slower and costlier per word.
• Pivot translation: Tamil→English→Marathi when no direct Tamil-Marathi data exists; errors compound at each hop.

Why Indian languages are hard for MT

• Morphological richness: Malayalam, Tamil, Telugu and Kannada are agglutinative — one word can carry case, number, tense, person and politeness, so vocabulary explodes. Handled by subword tokenisation (BPE/SentencePiece), which splits words into frequent pieces.
• Low-resource data: parallel corpora are small and domain-skewed (parliamentary, religious, news). Mitigated by back-translation (machine-translate monolingual target text into the source to make synthetic pairs) and corpus mining (Samanantar, BPCC).
• Script and encoding: Unicode is mandatory — Devanagari U+0900-U+097F, Tamil, Bengali, Malayalam, Telugu, Kannada, Gurmukhi, Odia, Gujarati blocks; conjuncts formed with the virama; ZWJ/ZWNJ change rendering; nukta and chandrabindu variants need Unicode normalisation (NFC) before any comparison. ISCII (IS 13194) is legacy only; old font-encoded "Kruti Dev / Shree Lipi" text is not text at all until converted.
• Honorifics and T-V distinction: तू / तुम / आप, நீ / நீங்கள் — MT picks one at random unless the style guide and a tuned model force it. Gender agreement with the user's gender is unresolvable from an isolated UI string.
• Code-mixing and transliteration: real Indian input is Hinglish and Romanised Hindi/Tamil; transliteration (Roman→native script, ITRANS/InScript, Google Input Tools) is a separate task from translation.
• Named entities and numbers: lakh/crore versus million/billion, Indian digit grouping, Devanagari digits (१२३) versus ASCII digits.

Post-editing: levels, process and standards

• Definition: correcting raw MT output to a specified standard — not retranslating. The governing standard is ISO 18587:2017, *Post-editing of machine translation output — Requirements*; the parent standard for human translation services is ISO 17100:2015, and ISO 5060:2024 covers evaluation of translation output. ISO/TS 11669 covers translation project specification.
• Light (rapid) post-editing: fix only what breaks meaning — mistranslation, omission, added content, wrong numbers, offensive output. Accept clumsy but understandable style. Used for internal documents, support tickets, user reviews, large low-value catalogues. ISO 18587 calls the outcome "comprehensible and accurate", not publishable-quality.
• Full post-editing: meaning, grammar, terminology, register, consistency, formatting — output indistinguishable from good human translation. Used for marketing, legal, medical, UI and anything published.
• The post-editor's rule of thumb: if a segment needs more than ~3 edits per 10 words, delete it and translate from scratch — "over-editing" a bad segment costs more than retranslating it.
• Workflow: source analysis → pre-editing (fix source typos, shorten sentences, mark do-not-translate items) → MT with glossary/terminology injection → post-edit in a CAT tool → automated QA checks → bilingual review → LQA sampling → delivery → TM and glossary update → MT retraining.
• Pre-editing and controlled language pay more than post-editing effort saved: short sentences, one idea per sentence, no ambiguous pronouns, consistent terms, no idioms.
• What never goes to a public MT engine: personal data and client confidential text. Until 13 May 2027 the applicable Indian regime is Information Technology Act, 2000 section 43A with the SPDI Rules, 2011; the Digital Personal Data Protection Act, 2023 data-fiduciary duties under the DPDP Rules, 2025 (notice, consent, breach intimation, erasure) apply from that date. Use on-premise or contracted no-retention engines for sensitive content.
• Copyright: under the Copyright Act, 1957 the right to make a translation is part of the author's exclusive rights (section 14), a translation is a protected work in its own right, and the translator retains moral rights (section 57). Raw MT output is not a human work; the post-edited text is where authorship is argued.

Quality evaluation: automatic and human

• Automatic metrics compare MT output against one or more human reference translations:
  • BLEU — modified n-gram precision with a brevity penalty. Fast, reproducible, but punishes legitimate paraphrase and is unreliable for morphologically rich Indian languages; never compare BLEU across language pairs or test sets.
  • chrF / chrF++ — character n-gram F-score; much better for Tamil, Malayalam, Telugu, Kannada because it credits partial word matches.
  • TER (Translation Edit Rate) — minimum edits (insert, delete, substitute, shift) per reference word; HTER uses the actual human post-edit as the reference, so it measures real effort.
  • METEOR — recall-weighted, allows stem and synonym matches.
  • COMET / BERTScore — neural, embedding-based metrics that correlate far better with human judgement; now the research default.
• Human evaluation: *adequacy* (is the meaning there?) and *fluency* (does it read naturally?) on 1-5 or 1-100 scales; direct assessment; ranking of systems; and error typology scoring — MQM (Multidimensional Quality Metrics) and the TAUS DQF categories: accuracy (mistranslation, omission, addition), fluency (grammar, spelling, punctuation), terminology, style, locale convention, design/markup — each weighted minor / major / critical.
• LQA pass-fail in industry: a reviewer scores a random sample (commonly 1,000-2,000 words) and computes an error-penalty score per 1,000 words against a threshold.
• MT suitability test: before quoting MTPE, run a pilot of 1,000-2,000 words, measure HTER and words-per-hour, and decide the discount from data — not from a vendor's claim.

Internationalisation and localisation

• g11n = i18n + l10n + t9n. Internationalisation (i18n) is engineering done once so the product *can* be localised; localisation (l10n) is adapting it for one locale; translation is only one part of l10n; transcreation is rewriting a marketing message for cultural effect, usually priced per hour or per piece, not per word.
• Locale identifiers: BCP 47 tags built from ISO 639 language codes and ISO 3166-1 country codes — hi-IN, ta-IN, bn-IN, ml-IN, mr-IN, ur-PK vs ur-IN, and script subtags where needed (pa-Guru-IN, sd-Arab-IN). A locale is language plus region, not a language.
• i18n requirements (the exam list): Unicode (UTF-8) end to end; externalise all strings into resource files (never hard-code); no string concatenation — use placeholders (%1$s, {count}); support plural rules (CLDR categories: one/other for Hindi, and language-specific sets) and gender variants; locale-aware date (dd-mm-yyyy in India), time, number, currency (₹ with Indian lakh/crore grouping: ₹12,34,567.89), collation and sorting; text expansion allowance in the UI (English→Hindi typically 10-25% longer; Devanagari also needs more line height for matras); bidirectional (RTL) support for Urdu and Kashmiri; locale-independent internal storage of dates and amounts; pseudo-localisation testing before any translation exists.
• Standard file formats: XLIFF (1.2 and 2.x) for bilingual exchange, TMX for translation memories, TBX for term bases, SRX for segmentation rules, ITS 2.0 for translation metadata in XML/HTML; plus platform formats — gettext .po, Java .properties, Android strings.xml, Apple .strings, .resx, JSON, YAML, Markdown, InDesign IDML, SRT/VTT for captions.
• CAT tool components: segmentation into sentences; translation memory (TM) with exact, fuzzy and context (101%) matches; term base; QA checks (numbers, terminology, double spaces, untranslated segments, tag mismatch); and MT plugged in as a suggestion source. The commercial tools in Indian use are Trados Studio, memoQ, Phrase (Memsource), Smartcat, Crowdin, Lokalise, MateCat, OmegaT (free, open-source).
• Localisation QA has three layers: linguistic (meaning, terminology, register), cosmetic/UI (truncation, overlap, clipped matras, wrong font fallback), functional (locale-dependent crashes, sorting, input method, date parsing).
• Pseudo-translation / continuous localisation: in agile products, strings are pushed to translators on every sprint through a connector; "localisation debt" accumulates when English ships ahead of the other languages.

The Indian ecosystem, institutions and law

• Bhashini / National Language Translation Mission (NLTM) — MeitY's mission to build Indian-language AI; exposes speech, translation and OCR models through ULCA (Universal Language Contribution API) and the Bhashini app; crowdsourced data via Bhasha Daan.
• AI4Bharat (IIT Madras): IndicTrans2 (22 scheduled languages, open-source), Samanantar and BPCC parallel corpora, IndicBERT, IndicNLG, IndicTrans toolkit — the standard baselines you should be able to name in an interview.
• Others: C-DAC (Mantra-Rajbhasha, Hindi official-language translation), TDIL (Technology Development for Indian Languages, MeitY) and its data centre, Anuvadini (AICTE, used to translate technical textbooks and course material), National Translation Mission at CIIL Mysuru (academic/textbook translation), CSTT — Commission for Scientific and Technical Terminology — the authority for Hindi scientific and technical terms, whose glossaries you cite in a Hindi term base.
• Commercial localisation in India: Reverie Language Technologies, Process9, Devnagri AI, Lionbridge/RWS/TransPerfect India delivery centres, and the language arms of TCS, Infosys, Wipro and HCLTech; OTT and media localisation for Netflix, Prime Video, JioHotstar, SonyLIV, ZEE5.
• Official language framework: the Official Languages Act, 1963 and the Official Languages (Use for Official Purposes of the Union) Rules, 1976 require specified central-government documents to issue in Hindi and English; Department of Official Language (Rajbhasha) targets drive enormous translation volume. The Eighth Schedule to the Constitution lists 22 scheduled languages. GIGW (Guidelines for Indian Government Websites, MeitY/NIC) requires government sites to be bilingual and accessible.
• Commercials: translation and post-editing services are a supply of services under GST; the standard services rate is 18% (the slab structure since 22 September 2025 is 5% / 18% / 40%). Freelancers invoicing abroad deal with LUT/export-of-service rules; income is taxed under the Income-tax Act, 2025 (in force from 1 April 2026) as business or professional income, with TDS deducted by Indian clients on professional fees — rates and thresholds change by Finance Act, so confirm the current year's figure.

Key points

• MT - software translation of text; MTPE - human correction of it; localisation - adapting a whole product to a locale.
• RBMT → SMT → NMT → LLM - rules, then phrase statistics, then end-to-end Transformers, then prompted general models.
• Transformer / self-attention - the architecture behind all current NMT; encoder reads the source, decoder emits the target.
• Hallucination - fluent invented output; the single most dangerous NMT failure because it reads perfectly.
• Subword tokenisation (BPE/SentencePiece) - splits agglutinative Tamil/Malayalam words so the vocabulary stays finite.
• Back-translation - synthetic parallel data from monolingual target text; standard fix for low-resource Indian pairs.
• Light vs full post-editing - meaning-only versus publishable; defined by ISO 18587:2017.
• ISO 17100:2015 - human translation services (qualified translator plus independent reviser); ISO 5060:2024 - evaluation of output.
• BLEU - n-gram precision with brevity penalty, 0-100; never compared across language pairs.
• chrF - character n-gram F-score; the fairer automatic metric for morphologically rich Indian languages.
• TER / HTER - edits per reference word; HTER uses the actual post-edit, so it measures real effort.
• COMET - neural metric, best correlation with human judgement.
• MQM / DQF - error typologies (accuracy, fluency, terminology, style, locale convention) with minor/major/critical severity.
• TM fuzzy match - similarity band (75-84%, 85-94%, 95-99%, 100%, 101% context) driving the discount grid.
• i18n before l10n - externalised strings, UTF-8, placeholders, plural and gender rules, expansion space, RTL support.
• Locale = language + region - BCP 47 tags like hi-IN, ta-IN, ur-IN; ISO 639 + ISO 3166.
• XLIFF / TMX / TBX - standard exchange formats for bilingual files, memories and term bases.
• Bhashini, ULCA, IndicTrans2, Samanantar, CSTT, Anuvadini - the Indian MT and terminology stack worth naming.
• Official Languages Act, 1963 + Rules, 1976 - statutory source of India's bilingual document volume.

Formulas and examples

Example 1 - BLEU score. A Hindi MT output has modified precisions p₁ = 0.70, p₂ = 0.50, p₃ = 0.35, p₄ = 0.25; candidate length c = 28 words, reference r = 30 words.

• c < r, so BP = exp(1 − 30/28) = exp(−0.07143) = 0.9311.
• Geometric mean = (0.70 × 0.50 × 0.35 × 0.25)^1/4 = (0.030625)^1/4. √0.030625 = 0.175; √0.175 = 0.4183.
• BLEU = 0.9311 × 0.4183 = 0.3895 → 38.95, reported as BLEU 38.9. A good English→Hindi general-domain NMT system sits roughly in the 25-40 band; the number is only meaningful against the same test set.

Example 2 - TER and HTER. Reference has 24 words. The post-editor made 2 substitutions, 1 insertion, 2 deletions and moved one 3-word phrase (1 shift) = 6 edits.

• TER = 6/24 = 0.25 → 25%. If the post-edited text itself is 23 words long, HTER = 6/23 = 26.1%.
• Reading: roughly one edit per four words — squarely in the band where full post-editing is still cheaper than translating from scratch.

Example 3 - fuzzy match from edit distance. Source segment in TM: "Click Submit to save the form" (29 characters); new segment: "Click Submit to send the form" (29 characters). Changing "save" to "send" needs 3 character substitutions (a→e, v→n, e→d): d = 3.

• Similarity = (1 − 3/29) × 100 = 89.7% → falls in the 85-94% band, payable at (say) 60% of the full rate.

Example 4 - MTPE quote, productivity and earnings. An 18,000-word Marathi e-learning project. Measured throughput: human translation 350 words/hour, full post-editing 900 words/hour. Agency rates: ₹2.00/word HT, ₹1.10/word MTPE (a 45% discount).

• Productivity gain = (900 − 350)/350 = 157%; time saving = 1 − 20/51.43 = 61.1%.
• Client saves ₹16,200 before tax (45%); the post-editor still earns 41% more per hour — the number to quote when a client demands a 60% discount, because at ₹0.80/word the hourly becomes ₹720, barely above translating from scratch.

Example 5 - CAT weighted word count and TM leverage. A 16,000-word Tamil software release analysed in the CAT tool, at a base rate of ₹1.50/word:

• Fee = 11,650 × ₹1.50 = ₹17,475 (+18% GST = ₹20,620.50), against ₹24,000 at flat rate.
• TM leverage = (1 − 11,650/16,000) × 100 = 27.2%.

Example 6 - UI text expansion. An English button label "Save changes" is 12 characters and the button is 96 px wide at 8 px/character. The Hindi string "परिवर्तन सहेजें" measures an expansion factor of 1.25 in rendered width.

• Required width = 96 × 1.25 = 120 px → the button truncates unless it is made flexible. Devanagari also needs extra line height for upper and lower matras, so a fixed 16 px row clips the vowel signs. This is exactly what pseudo-localisation (padding every English string by ~30%) catches before translation starts.

(Speed and price bands are industry norms, not fixed rates; they vary by language pair, domain and client.)

Where it is used

• Government and public sector: Rajbhasha translation cells in ministries, banks and PSUs under the Official Languages Act, 1963; Bhashini-powered services, judgment translation for courts (SUVAS/e-courts initiatives), bilingual GIGW-compliant websites.
• IT and product localisation: Indian-language UI for UPI apps, insurance and banking portals, government apps, and global software localised into hi-IN, ta-IN, bn-IN, te-IN, mr-IN; roles — localisation project manager, language lead, linguistic QA, terminologist, MT evaluator.
• E-commerce and consumer internet: product titles and descriptions at Amazon, Flipkart, Meesho translated at a scale only MT plus light post-editing can carry; review translation; voice and chatbot content.
• Media and OTT: subtitle and dub-script workflows seeded with MT and post-edited; news agencies and digital publishers running Hindi/Tamil/Telugu editions off MT with editorial post-editing.
• Education and research: AICTE's Anuvadini translating engineering textbooks, NPTEL/SWAYAM course material, CIIL's National Translation Mission, NCERT material in scheduled languages.
• Vendors and freelance market: RWS, Lionbridge, TransPerfect, Reverie, Process9, Devnagri AI; freelance MTPE on Smartcat, Phrase, ProZ, with CAT-tool proficiency as the entry gate.

Interview questions

• Q: What is the difference between translation, localisation and internationalisation? A: Translation converts text between languages; localisation adapts the whole product — text, formats, images, legal and cultural content — to a locale; internationalisation is the prior engineering that makes the product localisable at all.
• Q: Why does neural MT fail in ways that are harder to catch than statistical MT? A: NMT output is grammatically fluent even when wrong, so omissions, negation flips, wrong numbers and hallucinated content read naturally and survive a quick proofread.
• Q: When would you choose light post-editing over full post-editing? A: When the content is high-volume and low-value and only needs to be understood — support tickets, internal documents, user reviews, long-tail catalogue text — and it will not be published as the brand's own copy.
• Q: What are the limitations of BLEU, especially for Indian languages? A: It rewards surface n-gram overlap against one reference, so legitimate paraphrase and free word order are penalised, and rich morphology means near-correct inflected words score zero; chrF or COMET correlate better.
• Q: What does a 92% fuzzy match mean and why is it paid less? A: The segment is 92% similar to one already in the translation memory by edit distance, so the translator edits rather than translates it; the discount grid pays a fraction of the full rate for that band.
• Q: Which ISO standard governs post-editing, and what does it require? A: ISO 18587:2017 — it defines the post-editing process, the competences of a post-editor and the distinction between full and light post-editing, with ISO 17100:2015 covering human translation services.
• Q: How would you decide whether MT is usable at all for a new client's content? A: Run a pilot of 1,000-2,000 representative words, measure HTER and post-editing throughput against the from-scratch baseline, and price the discount from that data; if HTER is high or the content is creative, MT is rejected.
• Q: What is Bhashini and what is IndicTrans2? A: Bhashini is MeitY's National Language Translation Mission platform exposing Indian-language speech and translation models through ULCA; IndicTrans2 is AI4Bharat's open-source multilingual NMT model covering the 22 scheduled languages.
• Q: A UI string is "Delete {count} files". What localisation problems does it carry? A: Plural rules differ by language so one form is not enough; the placeholder must stay intact and may need reordering; and the Hindi or Tamil rendering will be longer, so the control must allow expansion.
• Q: Can you upload a client's confidential document to a free online MT engine? A: No — free engines may retain and reuse input, which breaches the NDA and, for personal data, Indian data-protection obligations; use a contracted no-retention or on-premise engine.
• Q: Who owns the copyright in a translation? A: Under the Copyright Act, 1957 the right to translate is the original author's, so permission is needed, but the resulting translation is itself a protected work and the translator holds moral rights in it.
• Q: Your MT output uses तू for a banking app's user. How do you fix it systemically? A: Fix it in the style guide and term base, then enforce it by fine-tuning or glossary injection and an automated QA check — not by hand-editing every segment, which will regress on the next release.

Common mistakes

• Treating post-editing as proofreading: reading only the target text, so omissions, negation flips and wrong numbers pass — always read bilingually against the source.
• Over-editing to personal taste on a light post-editing brief, or preferentially rewriting acceptable MT, destroying the productivity the discount was priced on.
• Quoting BLEU as an absolute quality grade, or comparing BLEU between different language pairs and test sets.
• Confusing localisation with internationalisation in exams, and confusing a *language* with a *locale* (hi is not hi-IN).
• Breaking placeholders, tags and escape sequences while editing, or hard-coding translated strings and concatenating sentences — the commonest cause of localisation build failures.
• Ignoring text expansion, line height and RTL, so Devanagari matras clip and Urdu layout breaks even though the translation is correct.
• Pasting client or personal data into public MT engines, and citing ad-hoc Hindi technical terms instead of the CSTT glossary or the client's approved term base.

Related topics

Translation · Subtitling and dubbing · Editing and proofreading · Content writing · Computational linguistics and NLP for Indian languages · Terminology management and lexicography