# Machine Translation

> Source: https://aiwiki.ai/wiki/machine_translation
> Updated: 2026-07-30
> Fact-checked: 2026-07-30
> Categories: Artificial Intelligence, Machine Learning, Natural Language Processing
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Machine Translation." aiwiki.ai, 30 Jul 2026. https://aiwiki.ai/wiki/machine_translation
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Machine translation** (MT) is the automatic translation of text or speech from one human language into another. It is a central task in [natural language processing](https://aiwiki.ai/wiki/natural_language_processing) and one of the oldest research problems in [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence). A translation system must preserve the source meaning while producing target-language wording that is grammatical and appropriate for its context. These objectives cannot be reduced to word replacement because languages differ in word order, morphology, idiom, reference, and the information they require speakers to express.[1][27]

Methods have changed substantially over the field's history. Early systems relied on manually written dictionaries and grammar rules. Statistical machine translation learned probabilities from bilingual corpora.[4] [Neural machine translation](https://aiwiki.ai/wiki/neural_machine_translation) trained an encoder and decoder jointly.[10] The Transformer then showed that an encoder-decoder model based on attention rather than recurrence could achieve strong reported translation results.[13] Multilingual systems can share one model across many languages, while general-purpose large language models can translate as one of many learned tasks. None of these changes eliminated the need to evaluate individual language pairs, domains, and risks. Results for well-resourced languages do not establish comparable quality for languages with little training or evaluation data.[19][21][22]

## Scope and terminology

In text-to-text MT, the input and output are written sequences. Speech translation adds acoustic input, synthesized speech output, or both. A speech system may be a cascade of [speech recognition](https://aiwiki.ai/wiki/speech_recognition), text MT, and [text-to-speech synthesis](https://aiwiki.ai/wiki/text_to_speech_ai), or it may learn some or all of those mappings in one model.[32] Simultaneous translation begins producing output before the source utterance is complete, which introduces an explicit tradeoff between waiting for context and responding with low delay.[31]

MT is commonly described by the amount and type of supervision available:

- **Supervised translation** learns from parallel corpora, in which source segments are paired with human translations.
- **Low-resource translation** concerns settings with limited parallel data, limited monolingual data, weak tools, or inadequate evaluation resources. Resource levels are properties of a particular language, domain, and translation direction, not fixed qualities of a language.
- **Multilingual translation** trains one system on multiple language pairs. Transfer between them can help some low-resource directions, but shared capacity can also create interference.[16][19]
- **Zero-shot translation** asks a multilingual model to translate a direction that was not present as a direct pair during training. The term does not imply that the model saw neither language.[16]
- **Document-level translation** supplies context beyond an isolated sentence so that the system can address discourse phenomena such as pronouns, lexical consistency, and ellipsis.[30]

Human translators remain part of many operational workflows. Machine output may be used as a draft, post-edited to a specified quality level, or rejected when the source, domain, confidentiality requirements, or potential harm makes automation unsuitable.[33][34]

## Historical development

### Early proposals and demonstrations

Ideas for mechanizing translation predate electronic computers. John Hutchins' historical survey identifies patents filed in the 1930s by Georges Artsrouni and Petr Troyanskii, including Troyanskii's proposals for an automatic bilingual dictionary and coded grammatical roles. After digital computers appeared, Warren Weaver's 1949 memorandum encouraged research that connected translation with cryptography, information theory, language structure, and statistical methods.[1]

The Georgetown-IBM experiment of 7 January 1954 was the first widely publicized public demonstration of computer translation. It translated selected Russian material into English on an IBM 701. The experiment was deliberately narrow: its authors used a vocabulary of about 250 items and six operational rules. The demonstration established that stored-program computers could manipulate bilingual text, but it was not evidence that unrestricted translation had been solved.[2]

In 1964, the United States government formed the Automatic Language Processing Advisory Committee to assess progress. Its 1966 report, *Language and Machines*, found no prospect of useful fully automatic high-quality translation in the immediate setting it examined and recommended investment in basic computational linguistics and tools for translators rather than continued support on the same terms.[3] The report strongly affected United States funding, while MT research and deployed systems continued elsewhere.[1][3]

### Rule-based systems

Rule-based machine translation represents linguistic knowledge explicitly. A typical transfer system analyzes a source sentence, maps a source-language representation to a target-language representation, and generates the target sentence. Direct systems perform less abstract transfer, while interlingual designs attempt to translate through a language-independent representation. Dictionaries contain lexical mappings and morphological information, and hand-written rules address agreement, word order, disambiguation, and generation.[1]

Rule-based systems can be controllable in a restricted domain, but building and maintaining their linguistic resources is labor intensive. Ambiguous words and constructions require context-sensitive decisions, and interacting rules can be difficult to scale across many language pairs. Historically important systems included SYSTRAN and the Canadian weather-forecast system METEO. Their practical use also showed that the feasibility of MT depends on domain, input regularity, and the acceptable level of human correction.[1]

The labels direct, transfer, and interlingual describe a spectrum rather than three implementations that every system follows. A direct system can apply bilingual substitutions and local reordering with little abstract analysis. A transfer system typically builds a language-specific analysis before applying bilingual structural rules. An interlingual system attempts a representation from which several target languages can be generated. In practice, deployed rule-based systems combined lexical, morphological, syntactic, and domain-specific procedures. Their histories caution against comparing paradigms as if each were a single model.[1]

## Statistical machine translation

Corpus-based methods shifted much of the work from writing rules to estimating models from data. In the noisy-channel formulation developed at IBM, a system selects the target sentence \(e\) that maximizes \(P(e \mid f)\) for a source sentence \(f\). Using [Bayes' theorem](https://aiwiki.ai/wiki/bayes_theorem), this can be expressed in terms of a translation model \(P(f \mid e)\) and a target [language model](https://aiwiki.ai/wiki/language_model) \(P(e)\).[4] The translation model learns relationships between the two languages from aligned bilingual text, while the language model rewards probable target-language sequences.

The IBM work introduced probabilistic word alignments as latent structures connecting words and positions across paired sentences. The later mathematical treatment described a sequence of five models with increasingly detailed assumptions about fertility, alignment, and distortion.[5] These models became important both as translation systems and as alignment components for later methods.

Sentence alignment and word alignment are related but different operations. Sentence alignment locates corresponding spans in bilingual documents. Word alignment estimates which source and target positions correspond within paired segments. Neither relation must be one-to-one: one word can translate as several, several can form one expression, and a source element can be omitted. The statistical models handle such uncertainty probabilistically rather than asserting that every pair has a unique dictionary mapping.[4][5]

Phrase-based statistical MT, which became the leading statistical approach in the 2000s, translates short sequences rather than individual words. Its "phrases" are contiguous word sequences and need not be linguistic constituents. Koehn, Och, and Marcu showed that comparatively simple phrase extraction and scoring methods could perform strongly in a shared experimental framework.[6] A decoder combines features such as phrase translation probabilities, target-language probability, reordering costs, and a length penalty.

The weights of those features matter. Minimum error rate training adjusted them to optimize an evaluation criterion on development data rather than estimating every weight independently.[7] The open-source Moses toolkit then assembled phrase extraction, scoring, tuning, and decoding into a reproducible research system. Its 2007 description emphasized a factored representation and support for multiple decoding models, helping laboratories compare methods on a common implementation.[8]

Statistical systems made their assumptions visible and could use large monolingual corpora through a separate language model. Their weaknesses included sparse estimates for rare events, large phrase tables, limited use of distant context, and complex pipelines whose stages were often trained separately. These limitations motivated end-to-end neural approaches, although statistical alignments and phrase-based systems remain useful historical and experimental baselines.

## Neural machine translation

### Encoder-decoder models

Neural MT learns a conditional distribution over target sequences with a single differentiable model. In the 2014 sequence-to-sequence system of Sutskever, Vinyals, and Le, one multilayer [long short-term memory network](https://aiwiki.ai/wiki/lstm) encoded a source sentence into a fixed-dimensional state and another decoded the translation. On the authors' WMT 2014 English-to-French test, an ensemble reached 34.8 BLEU, compared with 33.3 for the phrase-based baseline they reported. Reversing source word order also improved optimization by shortening important dependencies.[10] Those numbers belong to that test and preprocessing setup and should not be treated as universal quality levels.

Compressing a complete sentence into one fixed vector made long sequences difficult. Bahdanau, Cho, and Bengio instead let the decoder form a different weighted combination of encoder states for each output step. This learned soft alignment, now called [attention](https://aiwiki.ai/wiki/attention), improved the model's ability to use relevant source positions while generating each target token.[11] The method did not remove ambiguity or guarantee faithful translation, but it changed the encoder-decoder interface and became a standard part of recurrent neural MT.

Google's 2016 GNMT research system demonstrated the operational scale of this approach. The paper described eight encoder and eight decoder LSTM layers, attention, residual connections, WordPiece units, and a coverage-related decoding penalty.[12] In the paper's human side-by-side evaluation, the authors reported an average reduction of about 60 percent in translation errors relative to Google's phrase-based production system. That result was an internal comparison over the tested language pairs and sentence samples, not a general measure of all languages or domains.[12]

### The Transformer

The 2017 Transformer removed recurrence and convolution from the core sequence model. Its encoder and decoder use multi-head self-attention, feed-forward layers, residual connections, normalization, and positional encodings. Because training does not have to process sequence positions one after another as a recurrent network does, the architecture allows substantially more parallel computation.[13]

In the original NeurIPS paper's WMT 2014 experiments, the larger Transformer model obtained 28.4 BLEU on English-to-German and 41.0 BLEU on English-to-French under the authors' reported configurations.[13] These benchmark-specific results helped establish the architecture for translation. [Transformers](https://aiwiki.ai/wiki/transformers) subsequently became common throughout language modeling, but the architecture alone does not determine translation quality. Training data, objectives, tokenization, model size, decoding, and evaluation all remain consequential.

## Training and generation

### Data preparation

A supervised MT corpus normally consists of aligned source and target segments. Building it can require document matching, sentence alignment, language identification, deduplication, filtering, normalization, and removal of corrupted or mismatched pairs. Web-mined corpora can increase coverage, but errors in language labels, alignments, encodings, and translations become training signals unless they are detected. For languages with little digital text, data acquisition and evaluation can be as important as model architecture.[19]

Models usually operate on subword tokens rather than a closed vocabulary of whole words. Sennrich, Haddow, and Birch showed that decomposing words into variable-length units allowed a neural system to represent unseen compounds and productive morphology instead of mapping every unseen word to one unknown symbol.[35] SentencePiece later provided a language-independent implementation that can learn a fixed-size subword vocabulary directly from raw sentences without requiring language-specific pre-tokenization.[15] [Tokenization](https://aiwiki.ai/wiki/tokenization) still affects sequence length, name handling, script coverage, and the comparability of reported scores.

Cleaning choices can change which communities and varieties a model represents. Removing data judged noisy may improve average benchmark performance while disproportionately discarding nonstandard spelling, code-switching, minority scripts, or text from languages that an imperfect identifier confuses. Conversely, retaining unverified mined pairs can teach wrong-language output and semantic mismatches. Data documentation should therefore identify collection methods, filters, deduplication, language labels, licensing, and known coverage gaps.[19]

Parallel data is not the only usable signal. In back-translation, a reverse-direction model translates target-language monolingual sentences into synthetic source text. The resulting synthetic-source and authentic-target pairs augment the forward model's training data. Sennrich, Haddow, and Birch found this useful for improving neural translation with monolingual data.[14] The quality, domain, and sampling of synthetic data matter, and synthetic output can reproduce the reverse model's errors.

### Optimization and decoding

During training, an autoregressive decoder commonly predicts the next target token from the source and preceding reference tokens. At inference time, it conditions on its own earlier outputs. A simple greedy decoder takes the most probable token at each step; [beam search](https://aiwiki.ai/wiki/beam_search) keeps several partial hypotheses. A wider beam does not guarantee a better translation, and scoring often includes length or coverage adjustments.[12][27]

Training is typically evaluated on held-out development data, while a separate test set is reserved for final comparison. Domain overlap, duplicate examples, translationese, and test-set contamination can overstate generalization. Scores also depend on the reference translations and on preprocessing. A responsible report therefore identifies the language direction, domain, test set, metric version, tokenization, model checkpoint, and decoding settings rather than presenting an unqualified "accuracy" number.[22][23]

Adaptation can specialize a general model to a domain, language variety, terminology set, or style. Methods include continued training, [fine-tuning](https://aiwiki.ai/wiki/fine_tuning), terminology constraints, retrieval, and reranking. Adaptation can improve in-domain performance while reducing performance elsewhere, so held-out evaluation should include both the intended domain and relevant failure cases.

Terminology constraints are especially important where a term has an approved translation or must remain unchanged. A glossary can influence preprocessing, decoding, post-processing, or human review, but the system still has to inflect and place the term correctly. Copying a glossary entry mechanically can produce agreement or word-order errors. Evaluation should test terms in complete sentences and include names, units, punctuation, markup, and other elements that must survive translation.

## Multilingual and low-resource translation

A multilingual model shares parameters across language pairs. Johnson and colleagues trained one neural system for multiple directions by adding a token that specified the requested target language. Their experiments also produced zero-shot translations for pairs that were not directly present in training, although the paper found that these translations could use an unintended language and were weaker than supervised directions.[16]

Multilingual pretraining broadened this idea. mBART pretrained an entire encoder-decoder model as a denoising autoencoder on monolingual text in 25 languages, then fine-tuned it on translation data. The authors reported the largest benefits in low-resource and document-level settings among their tests.[17] [mT5](https://aiwiki.ai/wiki/mt5) applied a text-to-text pretraining objective to Common Crawl data covering 101 languages. Its authors also documented "accidental translation," in which the model produced text in the wrong language after task fine-tuning, and proposed mixing unlabeled multilingual data into fine-tuning as a mitigation.[18]

[No Language Left Behind](https://aiwiki.ai/wiki/no_language_left_behind), reported by the Meta research team in 2022 and later published in *Nature*, targeted 200 languages. It used mined parallel data, a human-translated FLORES-200 benchmark, and conditional computation through a sparsely gated [mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts). The authors evaluated more than 40,000 directions and reported a 44 percent relative BLEU improvement over the prior systems used in their comparison.[19] The average does not mean every direction improved by that amount, and BLEU differences across distinct test setups are not interchangeable.

Coverage research continued beyond 200 languages. A Meta preprint released in March 2026 describes Omnilingual MT models and an English-to-1,600-language evaluation. The work combines public multilingual corpora with newly curated bilingual data and studies both decoder-only and encoder-decoder specializations.[20] Its authors emphasize that recognizing text in a language is easier than reliably generating faithful translations in that language. The 1,600 figure refers to the paper's evaluated English-to-language setting, not uniform evidence for every pair among those languages, and the report remained a preprint at the research cutoff used for this article.[20]

Low-resource MT has no single architectural fix. Parallel data may be scarce, noisy, domain-limited, or unavailable; writing systems and language identifiers may be underrepresented; and qualified evaluators may be difficult to recruit. Transfer from related languages, multilingual training, synthetic data, lexicons, and [transfer learning](https://aiwiki.ai/wiki/transfer_learning) can help, but benefits vary. Evaluation must involve speakers and appropriate text rather than assuming that a high-resource result transfers automatically.[19][21]

## Large language models as translators

A general-purpose [large language model](https://aiwiki.ai/wiki/large_language_model) can be prompted to translate without a translation-specific interface. This can be useful for instructions about tone, terminology, or context, and one prompt can combine translation with explanation or transformation. It also creates evaluation complications: model versions may change, prompting affects output, and a fluent answer can omit, add, or reinterpret source content.

Broad multilingual evidence does not support a claim that one LLM is universally best. Robinson and colleagues evaluated ChatGPT on English-to-target translation into 203 non-English language varieties from FLORES-200. In the 201 directions also covered by NLLB, NLLB had the higher chrF score in 169 (84.1%).[21] This is a pairwise comparison on that benchmark, not a statement about all systems or every language variety.

The WMT24 general translation shared task compared dedicated systems, commercial services, and LLM-based entries across specific language directions and domains. Its conclusion named Claude 3.5 Sonnet the best overall system in that shared task and credited it with wins in nine language-pair tasks.[22] That is a bounded result under WMT24's test sets and human-evaluation protocol, not evidence of universal superiority across languages, domains, prompts, or later model versions. WMT25 expanded the challenge with harder, multi-domain and document-level test material. Its organizers reported 60 evaluated systems, including submitted systems and collected LLM or online outputs, and used professional error-span annotation alongside automatic metrics.[23] These shared tasks support conditional comparisons, not a permanent global ranking.

Translation-specific models and general LLMs differ in deployment cost, controllability, supported context, terminology behavior, and data governance. The right comparison is made on representative documents, with the same source text and explicit quality criteria. Brand-level claims without a dated model version, language direction, domain, prompt, and evaluation protocol are not reproducible.

## Evaluation

Translation quality has several dimensions. A fluent sentence can be unfaithful, and a literal sentence can preserve information while sounding unnatural. Evaluation may consider accuracy, completeness, terminology, grammar, style, locale conventions, document consistency, harmful output, and fitness for the intended use.[26][34]

### Automatic metrics

[BLEU](https://aiwiki.ai/wiki/bleu_bilingual_evaluation_understudy) compares machine output with one or more reference translations using modified n-gram precision and a brevity penalty. Its creators proposed it as a fast corpus-level measure that correlated with human judgments in their experiments.[9] BLEU is sensitive to reference choice and tokenization, cannot fully credit valid paraphrases, and should not be interpreted as a percentage of sentences translated correctly.

chrF computes an F-score over character n-grams. The original study found strong system-level correlations on WMT data and noted that the metric is language independent and does not require tokenization.[24] Character units can be helpful when words have rich morphology, but chrF remains a reference-overlap measure.

COMET is a learned evaluation framework that can use the source sentence, candidate translation, and reference. Its original models used multilingual pretrained representations and were trained to predict human judgments.[25] Learned metrics may correlate better with a particular set of judgments while also inheriting model, training-data, and version dependencies. Reports should identify the exact checkpoint and not compare scores from incompatible metric versions.

No automatic metric is a complete substitute for human review. Metrics can be used together for development, but optimizing one metric can favor its blind spots. Segment-level scores are noisier than corpus-level aggregates, and statistically significant differences may still be too small to matter for a particular use.

Reproducible metric reporting requires more than the metric name. BLEU implementations can differ in tokenization, casing, smoothing, reference handling, and aggregation. Learned metrics can differ by checkpoint even when they share a family name. Test sets may contain multiple domains or document boundaries that are lost when sentences are shuffled. A comparison should publish the test-set version, preprocessing, metric implementation and signature, system outputs, and uncertainty analysis when possible.[22][23][25]

### Human evaluation

Human evaluation also depends on its design. Freitag and colleagues compared crowd judgments on isolated sentences with expert Multidimensional Quality Metrics annotation in document context. Their study found that the expert, context-aware protocol changed system rankings and gave a clearer advantage to human translations.[26] This result shows why annotator expertise and supplied context must be reported.

ISO 5060:2024 specifies guidance for analytic evaluation of human translations, post-edited machine translations, and unedited machine output. It uses error categories and penalties to produce translation quality measures.[34] A valid evaluation plan defines the purpose of the translation, sampling, evaluator qualifications, error taxonomy, severity levels, adjudication, and acceptance threshold before results are inspected.

## Speech and simultaneous translation

Speech translation compounds the uncertainties of speech recognition, translation, and speech generation. A cascade can expose intermediate transcripts and allow components to be replaced separately, while a direct or unified model can share representations across modalities. Error accounting should distinguish failures caused by acoustics, source transcription, semantic transfer, and target speech generation.

The SeamlessM4T Nature paper describes a single model supporting automatic speech recognition, speech-to-text, speech-to-speech, text-to-text, and text-to-speech tasks for language sets that vary by modality, with coverage of up to 100 languages.[32] The phrase "up to" matters: it does not mean that every modality supports the same 100 languages or every possible direction. Its reported benchmark improvements are experimental results for named datasets and baselines, not a guarantee for live conversations.

Simultaneous MT must decide when to read more source input and when to emit target output. Waiting provides syntactic and semantic context; speaking early reduces delay but can force revisions or premature commitments. Monotonic Infinite Lookback attention learns an adaptive read-write schedule while retaining access to earlier encoder states, and its evaluation explicitly measures quality against latency.[31] Latency should therefore be reported with a defined metric and hardware or service setting, not as an unsupported universal response time.

## Reliability and known limitations

Machine translation errors are not limited to awkward wording. Important failure classes include:

- **Domain mismatch:** A model trained on news or web text can fail on medicine, law, software interfaces, conversation, or historical language because terminology and distributions differ.[27]
- **Rare forms and names:** Subword modeling reduces unknown tokens but does not ensure correct names, numbers, compounds, transliterations, or inflected forms.[27][35]
- **Long-range and document context:** Sentence-level input may not resolve pronouns, repeated terminology, ellipsis, or discourse relations. Public document-aligned corpora remain less abundant than sentence pairs, although context-aware systems improve some targeted phenomena.[30]
- **Hallucination:** A system may generate fluent target text that is detached from, contradicts, or adds to the source. A 2023 study built a manually annotated corpus of naturally occurring neural MT hallucinations and found that several detection methods missed substantial cases.[28] [Hallucination](https://aiwiki.ai/wiki/hallucination) detection is therefore a safeguard, not proof of faithfulness.
- **Social bias:** Gender information that is absent or ambiguous in a source can be resolved according to learned stereotypes. A broad survey documents both representational harms and performance differences in gender translation, as well as the limits of existing tests and mitigations.[29] This is one instance of [bias](https://aiwiki.ai/wiki/bias), not the only fairness concern.
- **Wrong-language output:** Multilingual models can mix languages or generate a language other than the requested target, particularly in zero-shot and weakly supervised settings.[16][18]
- **Evaluation gaps:** Many languages lack representative benchmarks and enough qualified evaluators. Automatic scores can hide meaning-changing errors that are rare but severe.[19][26]

Confidence scores do not automatically identify these failures, and fluent output should not be treated as verified. High-consequence use requires source-aware review by people qualified in the language pair and subject matter.

Some errors arise from genuine underspecification. A source may omit a subject, grammatical gender, number, politeness level, or sense that the target language forces the translator to choose. The correct response may depend on preceding sentences, a speaker's identity, or the document's intended audience. A system that receives only one sentence cannot recover information that is not present there. Supplying context can reduce some ambiguity, but it also gives the model more material from which to copy, ignore, or infer incorrectly.[29][30]

## Human workflows and standards

Post-editing is the revision of machine output by a human. ISO 18587:2017 specifies requirements for full post-editing of machine translation output and the competences of post-editors. It applies to text processed by MT and sets a service framework rather than certifying the intrinsic quality of a model.[33]

A practical workflow can include source preparation, selection of an approved system, terminology controls, machine translation, automated checks, bilingual human review, formatting review, and final approval. The depth of review should follow the consequences of an error. Informal comprehension, publication, contracts, clinical communication, safety instructions, and public policy do not share the same risk threshold.

Full post-editing is not the same as making output merely understandable. The post-editor checks the source against the entire target, corrects omissions and additions, applies terminology and style requirements, and ensures that formatting and locale conventions are fit for delivery.[33] Measuring elapsed editing time or the number of changes can inform a workflow, but neither alone proves that the final text is accurate. Quality evaluation and productivity measurement answer different questions.

Privacy and intellectual-property requirements also affect system choice. Sending text to an external service can expose confidential or regulated material depending on the service terms and configuration. These questions are separate from linguistic accuracy and should be assessed before uploading source content.

## Selected milestones

| Year | Milestone | Significance |
|---|---|---|
| 1949 | Warren Weaver circulates his translation memorandum | Connects computer translation with cryptography, information theory, linguistic structure, and statistical ideas.[1] |
| 1954 | Georgetown-IBM public demonstration | Demonstrates a narrow Russian-to-English system on an IBM 701 with about 250 vocabulary items and six rules.[2] |
| 1966 | ALPAC publishes *Language and Machines* | Criticizes the near-term utility and economics of then-current MT and recommends more basic research and translator aids.[3] |
| 1990 to 1993 | IBM statistical models | Formalize translation probabilities and latent word alignment learned from bilingual corpora.[4][5] |
| 2002 | BLEU | Introduces a fast corpus-level automatic metric based on reference n-gram overlap and a brevity penalty.[9] |
| 2003 | Phrase-based modeling and minimum error rate training | Establishes influential methods for phrase translation and direct tuning against an evaluation criterion.[6][7] |
| 2014 to 2015 | Sequence-to-sequence and neural attention | Shows competitive end-to-end recurrent translation and a learned source-context mechanism.[10][11] |
| 2016 | GNMT, subword NMT, and back-translation | Demonstrates large recurrent NMT and widely adopted ways to handle open vocabulary and monolingual data.[12][14][35] |
| 2017 | Transformer | Replaces recurrence in the core translation model with attention-based parallel sequence processing.[13] |
| 2022 to 2024 | NLLB research and publication | Extends a single research program and benchmark to 200 languages and more than 40,000 evaluated directions.[19] |
| 2023 to 2025 | LLM and harder shared-task evaluations | Shows that general LLM translation is competitive in some conditions but highly dependent on language, domain, and evaluation design.[21][22][23] |
| 2026 | Omnilingual MT preprint | Reports English-to-1,600-language evaluation while identifying faithful target-language generation as a remaining bottleneck.[20] |

## See also

- [Deep Learning](https://aiwiki.ai/wiki/deep_learning)
- [Recurrent Neural Network](https://aiwiki.ai/wiki/recurrent_neural_network)
- [Large Language Model](https://aiwiki.ai/wiki/large_language_model)
- [Fine Tuning](https://aiwiki.ai/wiki/fine_tuning)

## References

1. Hutchins, John. "The History of Machine Translation in a Nutshell." Revised January 2014. https://aclanthology.org/www.mt-archive.info/10/Hutchins-2014.pdf
2. Hutchins, John. "The First Public Demonstration of Machine Translation: The Georgetown-IBM System, 7th January 1954." 2006. https://open.unive.it/hitrade/books/HutchinsFirst.pdf
3. Automatic Language Processing Advisory Committee. Language and Machines: Computers in Translation and Linguistics. National Academy of Sciences, 1966. https://doi.org/10.17226/20813
4. Brown, Peter F., et al. "A Statistical Approach to Machine Translation." Computational Linguistics 16, no. 2, 1990. https://aclanthology.org/J90-2002/
5. Brown, Peter F., et al. "The Mathematics of Statistical Machine Translation: Parameter Estimation." Computational Linguistics 19, no. 2, 1993. https://aclanthology.org/J93-2003/
6. Koehn, Philipp, Franz Josef Och, and Daniel Marcu. "Statistical Phrase-Based Translation." HLT-NAACL 2003. https://aclanthology.org/N03-1017/
7. Och, Franz Josef. "Minimum Error Rate Training in Statistical Machine Translation." ACL 2003. https://aclanthology.org/P03-1021/
8. Koehn, Philipp, et al. "Moses: Open Source Toolkit for Statistical Machine Translation." ACL 2007 Demonstrations. https://aclanthology.org/P07-2045/
9. Papineni, Kishore, et al. "BLEU: A Method for Automatic Evaluation of Machine Translation." ACL 2002. https://aclanthology.org/P02-1040/
10. Sutskever, Ilya, Oriol Vinyals, and Quoc V. Le. "Sequence to Sequence Learning with Neural Networks." Advances in Neural Information Processing Systems 27, 2014. https://papers.nips.cc/paper/5346-sequence-to-sequence-learning-with-neural-networks
11. Bahdanau, Dzmitry, Kyunghyun Cho, and Yoshua Bengio. "Neural Machine Translation by Jointly Learning to Align and Translate." ICLR 2015. https://arxiv.org/abs/1409.0473
12. Wu, Yonghui, et al. "Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation." 2016. https://arxiv.org/abs/1609.08144
13. Vaswani, Ashish, et al. "Attention Is All You Need." Advances in Neural Information Processing Systems 30, 2017. https://papers.nips.cc/paper/7181-attention-is-all-you-need
14. Sennrich, Rico, Barry Haddow, and Alexandra Birch. "Improving Neural Machine Translation Models with Monolingual Data." ACL 2016. https://aclanthology.org/P16-1009/
15. Kudo, Taku, and John Richardson. "SentencePiece: A Simple and Language Independent Subword Tokenizer and Detokenizer for Neural Text Processing." EMNLP 2018 System Demonstrations. https://aclanthology.org/D18-2012/
16. Johnson, Melvin, et al. "Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation." Transactions of the Association for Computational Linguistics 5, 2017. https://aclanthology.org/Q17-1024/
17. Liu, Yinhan, et al. "Multilingual Denoising Pre-training for Neural Machine Translation." Transactions of the Association for Computational Linguistics 8, 2020. https://aclanthology.org/2020.tacl-1.47/
18. Xue, Linting, et al. "mT5: A Massively Multilingual Pre-trained Text-to-Text Transformer." NAACL 2021. https://aclanthology.org/2021.naacl-main.41/
19. NLLB Team. "Scaling Neural Machine Translation to 200 Languages." Nature 630 (2024): 841-846. https://doi.org/10.1038/s41586-024-07335-x
20. Omnilingual MT Team, et al. "Omnilingual MT: Machine Translation for 1,600 Languages." arXiv preprint, March 2026. https://arxiv.org/abs/2603.16309
21. Robinson, Nathaniel R., et al. "ChatGPT MT: Competitive for High- (but Not Low-) Resource Languages." WMT 2023. https://aclanthology.org/2023.wmt-1.40/
22. Kocmi, Tom, et al. "Findings of the WMT24 General Machine Translation Shared Task: The LLM Era Is Here but MT Is Not Solved Yet." WMT 2024. https://aclanthology.org/2024.wmt-1.1/
23. Kocmi, Tom, et al. "Findings of the WMT25 General Machine Translation Shared Task: Time to Stop Evaluating on Easy Test Sets." WMT 2025. https://aclanthology.org/2025.wmt-1.22/
24. Popovic, Maja. "chrF: Character n-gram F-score for Automatic MT Evaluation." WMT 2015. https://aclanthology.org/W15-3049/
25. Rei, Ricardo, et al. "COMET: A Neural Framework for MT Evaluation." EMNLP 2020. https://aclanthology.org/2020.emnlp-main.213/
26. Freitag, Markus, et al. "Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translation." Transactions of the Association for Computational Linguistics 9, 2021. https://aclanthology.org/2021.tacl-1.87/
27. Koehn, Philipp, and Rebecca Knowles. "Six Challenges for Neural Machine Translation." First Workshop on Neural Machine Translation, 2017. https://aclanthology.org/W17-3204/
28. Guerreiro, Nuno M., Elena Voita, and Andre F. T. Martins. "Looking for a Needle in a Haystack: A Comprehensive Study of Hallucinations in Neural Machine Translation." EACL 2023. https://aclanthology.org/2023.eacl-main.75/
29. Savoldi, Beatrice, et al. "Gender Bias in Machine Translation." Transactions of the Association for Computational Linguistics 9, 2021. https://aclanthology.org/2021.tacl-1.51/
30. Pal, Proyag, Alexandra Birch, and Kenneth Heafield. "Document-Level Machine Translation with Large-Scale Public Parallel Corpora." ACL 2024. https://aclanthology.org/2024.acl-long.712/
31. Arivazhagan, Naveen, et al. "Monotonic Infinite Lookback Attention for Simultaneous Machine Translation." ACL 2019. https://aclanthology.org/P19-1126/
32. SEAMLESS Communication Team. "Joint Speech and Text Machine Translation for Up to 100 Languages." Nature 637 (2025): 587-593. https://doi.org/10.1038/s41586-024-08359-z
33. International Organization for Standardization. "ISO 18587:2017: Translation Services: Post-editing of Machine Translation Output: Requirements." 2017. https://www.iso.org/standard/62970.html
34. International Organization for Standardization. "ISO 5060:2024: Translation Services: Evaluation of Translation Output: General Guidance." 2024. https://www.iso.org/standard/80701.html
35. Sennrich, Rico, Barry Haddow, and Alexandra Birch. "Neural Machine Translation of Rare Words with Subword Units." ACL 2016. https://aclanthology.org/P16-1162/