Editing Machine translation (section)

==Approaches==
{{See also|Hybrid machine translation|Example-based machine translation|}}

Before the advent of [[deep learning]] methods, statistical methods required a lot of rules accompanied by [[morphology (linguistics)|morphological]], [[syntax|syntactic]], and [[semantics|semantic]] annotations.

===Rule-based===
{{Main|Rule-based machine translation}}

The rule-based machine translation approach was used mostly in the creation of [[dictionaries]] and grammar programs. Its biggest downfall was that everything had to be made explicit: orthographical variation and erroneous input must be made part of the source language analyser in order to cope with it, and lexical selection rules must be written for all instances of ambiguity.

====Transfer-based machine translation====
{{Main|Transfer-based machine translation}}

Transfer-based machine translation was similar to [[interlingual machine translation]] in that it created a translation from an intermediate representation that simulated the meaning of the original sentence. Unlike interlingual MT, it depended partially on the language pair involved in the translation.

====Interlingual====
{{Main|Interlingual machine translation}}

Interlingual machine translation was one instance of rule-based machine-translation approaches.  In this approach, the source language, i.e. the text to be translated, was transformed into an interlingual language, i.e. a "language neutral" representation that is independent of any language. The target language was then generated out of the [[interlinguistics|interlingua]]. The only interlingual machine translation system that was made operational at the commercial level was the KANT system (Nyberg and Mitamura, 1992), which was designed to translate Caterpillar Technical English (CTE) into other languages.

====Dictionary-based====
{{Main|Dictionary-based machine translation}}

Machine translation used a method based on [[dictionary]] entries, which means that the words were translated as they are by a dictionary.

===Statistical===
{{main|Statistical machine translation}}
Statistical machine translation tried to generate translations using [[statistical methods]] based on bilingual text corpora, such as the [[Hansard#Translation|Canadian Hansard]] corpus, the English-French record of the Canadian parliament and [[Europarl corpus|EUROPARL]], the record of the [[European Parliament]]. Where such corpora were available, good results were achieved translating similar texts, but such corpora were rare for many language pairs. The first statistical machine translation software was [[CANDIDE]] from [[IBM]]. In 2005, Google improved its internal translation capabilities by using approximately 200 billion words from United Nations materials to train their system; translation accuracy improved.<ref>{{cite web |url=http://blog.outer-court.com/archive/2005-05-22-n83.html |title=Google Translator: The Universal Language |publisher=Blog.outer-court.com |date=25 January 2007 |access-date=2012-06-12 |archive-date=20 November 2008 |archive-url=https://web.archive.org/web/20081120030225/http://blog.outer-court.com/archive/2005-05-22-n83.html |url-status=live }}</ref>

SMT's biggest downfall included it being dependent upon huge amounts of parallel texts, its problems with morphology-rich languages (especially with translating ''into'' such languages), and its inability to correct singleton errors.

Some work has been done in the utilization of multiparallel [[text corpus|corpora]], that is a body of text that has been translated into 3 or more languages. Using these methods, a text that has been translated into 2 or more languages may be utilized in combination to provide a more accurate translation into a third language compared with if just one of those source languages were used alone.<ref>{{Cite conference |last=Schwartz |first=Lane |date=2008 |title=Multi-Source Translation Methods |url=https://dowobeha.github.io/papers/amta08.pdf |conference=Paper presented at the 8th Biennial Conference of the Association for Machine Translation in the Americas |archive-url=https://web.archive.org/web/20160629171944/http://dowobeha.github.io/papers/amta08.pdf |archive-date=29 June 2016 |access-date=3 November 2017 |url-status=live}}</ref><ref>{{Cite conference |last1=Cohn |first1=Trevor |last2=Lapata |first2=Mirella |date=2007 |title=Machine Translation by Triangulation: Making Effective Use of Multi-Parallel Corpora |url=http://homepages.inf.ed.ac.uk/mlap/Papers/acl07.pdf |conference=Paper presented at the 45th Annual Meeting of the Association for Computational Linguistics, June 23–30, 2007, Prague, Czech Republic |archive-url=https://web.archive.org/web/20151010171334/http://homepages.inf.ed.ac.uk/mlap/Papers/acl07.pdf |archive-date=10 October 2015 |access-date=3 February 2015 |url-status=live}}</ref><ref>{{Cite journal |last1=Nakov |first1=Preslav |last2=Ng |first2=Hwee Tou |date=2012 |title=Improving Statistical Machine Translation for a Resource-Poor Language Using Related Resource-Rich Languages |url=https://jair.org/index.php/jair/article/view/10764 |journal=Journal of Artificial Intelligence Research |volume=44 |pages=179–222 |arxiv=1401.6876 |doi=10.1613/jair.3540 |doi-access=free}}</ref>

=== Neural MT ===
{{Main|Neural machine translation}}

A [[deep learning]]-based approach to MT, [[neural machine translation]] has made rapid progress in recent years. However, the current consensus is that the so-called human parity achieved is not real, being based wholly on limited domains, language pairs, and certain test benchmarks<ref>Antonio Toral, Sheila Castilho, Ke Hu, and Andy
Way. 2018. Attaining the unattainable? reassessing claims of human parity in neural machine translation. CoRR, abs/1808.10432.</ref> i.e., it lacks statistical significance power.<ref>{{Cite arXiv |eprint=1906.09833 |first1=Graham |last1=Yvette |first2=Haddow |last2=Barry |title=Translationese in Machine Translation Evaluation |date=2019 |last3=Koehn |first3=Philipp|class=cs.CL }}</ref>

Translations by neural MT tools like [[DeepL Translator]], which is thought to usually deliver the best machine translation results as of 2022, typically still need post-editing by a human.<ref>{{cite journal |last1=Katsnelson |first1=Alla |title=Poor English skills? New AIs help researchers to write better |journal=Nature |pages=208–209 |language=en |doi=10.1038/d41586-022-02767-9 |date=29 August 2022|volume=609 |issue=7925 |pmid=36038730 |bibcode=2022Natur.609..208K |s2cid=251931306 |doi-access=free }}</ref><ref>{{cite web |last1=Korab |first1=Petr |title=DeepL: An Exceptionally Magnificent Language Translator |url=https://towardsdatascience.com/deepl-an-exceptionally-magnificent-language-translator-78e86d8062d3 |website=Medium |access-date=9 January 2023 |language=en |date=18 February 2022}}</ref><ref>{{cite news |title=DeepL outperforms Google Translate – DW – 12/05/2018 |url=https://www.dw.com/en/deepl-cologne-based-startup-outperforms-google-translate/a-46581948 |access-date=9 January 2023 |work=Deutsche Welle |language=en}}</ref>

Instead of training specialized translation models on parallel datasets, one can also [[Prompt engineering|directly prompt]] generative [[large language model]]s like [[Generative pre-trained transformer|GPT]] to translate a text.<ref name="Hendy2023">{{cite arXiv |last1=Hendy |first1=Amr |last2=Abdelrehim |first2=Mohamed |last3=Sharaf |first3=Amr |last4=Raunak |first4=Vikas |last5=Gabr |first5=Mohamed |last6=Matsushita |first6=Hitokazu |last7=Kim |first7=Young Jin |last8=Afify |first8=Mohamed |last9=Awadalla |first9=Hany |title=How Good Are GPT Models at Machine Translation? A Comprehensive Evaluation |date=2023-02-18 |eprint=2302.09210 |class=cs.CL}}</ref><ref>{{cite news |last1=Fadelli |first1=Ingrid |title=Study assesses the quality of AI literary translations by comparing them with human translations |url=https://techxplore.com/news/2022-11-quality-ai-literary-human.html |access-date=18 December 2022 |work=techxplore.com |language=en}}</ref><ref name="arxiv221014250">{{Cite arXiv|last1=Thai |first1=Katherine |last2=Karpinska |first2=Marzena |last3=Krishna |first3=Kalpesh |last4=Ray |first4=Bill |last5=Inghilleri |first5=Moira |last6=Wieting |first6=John |last7=Iyyer |first7=Mohit |title=Exploring Document-Level Literary Machine Translation with Parallel Paragraphs from World Literature |date=25 October 2022|class=cs.CL |eprint=2210.14250 }}</ref> This approach is considered promising,<ref name="WMT2023">{{cite conference |last1=Kocmi |first1=Tom |last2=Avramidis |first2=Eleftherios |last3=Bawden |first3=Rachel |last4=Bojar |first4=Ondřej |last5=Dvorkovich |first5=Anton |last6=Federmann |first6=Christian |last7=Fishel |first7=Mark |last8=Freitag |first8=Markus |last9=Gowda |first9=Thamme |last10=Grundkiewicz |first10=Roman |last11=Haddow |first11=Barry |last12=Koehn |first12=Philipp |last13=Marie |first13=Benjamin |last14=Monz |first14=Christof |last15=Morishita |first15=Makoto |date=2023 |editor-last=Koehn |editor-first=Philipp |editor2-last=Haddow |editor2-first=Barry |editor3-last=Kocmi |editor3-first=Tom |editor4-last=Monz |editor4-first=Christof |title=Findings of the 2023 Conference on Machine Translation (WMT23): LLMs Are Here but Not Quite There Yet |url=https://aclanthology.org/2023.wmt-1.1 |journal=Proceedings of the Eighth Conference on Machine Translation |location=Singapore |publisher=Association for Computational Linguistics |pages=1–42 |doi=10.18653/v1/2023.wmt-1.1|doi-access=free }}</ref> but is still more resource-intensive than specialized translation models.