MinT

MinT（ミンティー：Machine in Translation）は機械翻訳サービスの一種でオープンソースのニューラル機械翻訳モデルに基づきます。 当サービスはウィキメディア財団のインフラ上にホスティングされており、他の組織がリリースしたオープンソース・ライセンスの翻訳モデルと競合しません。 無料の知識エコシステムのインフラにとって公開の機械翻訳サービスは鍵となります。 このページでは当該のインフラをもっと多くの人に使えるように、当サービスを判定しようとするイニシアティブを記録します。

この MinT を試用するには、コンテンツ翻訳や翻訳ウィキのサイト translatewiki.net の各プロジェクトに組み込みを使用するか、直接、テスト例を体験できます。

Machine translation can be useful in different contexts. As more products make use of MinT for different purposes, it is useful to differentiate those different contexts. In this way, when users report a bug it is more clear where it needs to be fixed.

• MinT Service. The backend service running open-source neural machine translation models.
  • MinT test instance. A basic interface to try the different translation models.
• MinT for Translators. Initiative to integrate the MinT Service with tools that support other machine translaiton services such as Content Translation and the Translate Extension.
  • MinT Client for Content Translation. Client exposing the MinT Service as one of the machine translation services available in Content Translation.
  • MinT Client for Translate extension. Client exposing the MinT Service as one of the machine translation services available in the Translate extension.
• MinT for Wiki Readers. Product to enable readers to use machine translation to read contents from other languages on a wiki.

フィードバックを提供するには協議ページに投稿してください。 改善計画はPhabricator にあがっていき、改善の案を提示したり問題点を指摘したり、タスクが始まっていたらその進捗チェック、それに関する自分なりの視点を共有してください。 完了した工程の確認もでき、以下にある進捗状況のチェック欄をご参照ください。

MinT サービスの設計では訳文を複数の機械翻訳モデルから提供します。 当初は以下のモデルを採用します:

• NLLB-200。メタの研究チームが手がけた最新モデル No Language Left Behind project です。 当モデルは言語 200 件の翻訳に対応し、その中には他の業者がサポートしていない言語も含まれます。
• OpusMT（オーパス・エムティー）。ヘルシンキ大学が開発したOPUS (Open Parallel Corpus) projectはフリーライセンスの多言語コンテンツをまとめて翻訳モデルOpusMT 翻訳モデル（オーパスMT）を訓練しています。 誰でもさまざまなプロジェクトに参加してデータをOPUSに提供すると、翻訳の質向上に手軽に貢献できます。 例えば利用者がウィキペディアの記事を訳すときにコンテンツ翻訳拡張機能を使うと、システム側は公開した訳文データを新しいリソースとして回収、同モデルの次のバージョンの翻訳の品質改善に役立てします。 あるいはまたTatoebaを使って訳文を提供すると、利用者が手軽に寄与するもう一つの���法になります。
• IndicTrans2. The IndicTrans2 project provides translation models to support over 20 Indic languages. These models were developed by AI4Bharat@IIT Madras, a research group at the Indian Institute of Technology, Madras.
• Softcatalà. Softcatalà is a non-profit organization with the goal to improve the use of Catalan in digital products. As part of the Softcatalà Translation project, translation models used in their translator service to translate 10 languages to and from Catalan have been released.
• MADLAD-400. MADLAD-400 is a multilingual machine translation model by Google Research that supports 419 languages.

MinT supports over 200 languages, with more than 70 languages not supported by other services (including 27 languages for which there is no Wikipedia yet). You can read more about the initial release of MinT and check some frequently asked questions in the summary page for the service.

The translation models have been optimized for performance using OpenNMT Ctranslate2 library in order to avoid the need for GPU acceleration. This makes it easier for organizations and individuals to build and run their own instances. For more details you can check the following:

• the source code (GitHub mirror)
• the API spec
• a test instance

MinT provides a platform to run multiple translation models. In order to support different initiatives, aspects such as sentence segmentation, language detection, pre/post-processing of contents, and rich format support has been developed on top of the plain-text based models.

The MinT test instance is a basic interface to try the different translation models. It allow to translate contents across the selected language pairs and select the preferred translation model when multiple are available. This allows different communities to check how well the models support their language. This instance is intended for testing, so performance and availability may be reduced compared to other MinT-based products. You can check the availability status of the MinT test instance.

Translation is a common way to contribute in the Wikimedia ecosystem for multilingual users. Machine translation can provide a useful initial translation for users to review and improve. The Language team has developed tools to support translations in their workflows that can integrate different machine translation services to speed up their processes. Once MinT was available, integrating it with these tools was a logical next step to amplify their impact. MinT is available in the following projects:

Content Translation integrates several translation services to provide an initial translation. You can check which languages supported by MinT are available in Content Translation

Communities of translators use it on translatewiki.net , Wikimedia Meta-wiki, MediaWiki.org and more.

The number of topics and the amount of information a reader can learn about from Wikipedia and other wikis depends on the languages they speak. Machine translation can help people to learn more about their topics of interest when the content is not available in their language.

• Allows readers to learn more about the topics of interest from other languages.
• Clearly differentiates automatically generated content from community-created one.
• Encourages to access and contribute to community-created content when possible.

Working on the previous initiatives will help to polish and solidify the system. For now, the MinT API is only available for Wikimedia products. As the system gets ready, we'll consider a wider exposure. Providing a service that can be used by communities in innovative ways can be a very powerful tool. New initiatives to make MinT more widely available will be captured here in the future. Meanwhile, feel free to configure your own MinT instance to experiment with it.

1. Accuracy of MinT’s Translations - The accuracy of translations generated by MinT may vary. Translations may not be entirely accurate or may not always convey the intended meaning or context of the original content. Wikimedia makes no representations or warranties regarding the accuracy or adequacy of the automatically translated content.
2. Limitation of Liability - Wikimedia, its affiliates, and employees are not liable for any direct, indirect, incidental, punitive, or consequential damages, including but not limited to damages for goodwill, use, data, or any other intangible losses arising out of or in connection with the use of MinT or translations generated with MinT.
3. Creative Commons Compliance - Translations generated with MinT are considered derivative works under the applicable Creative Commons license governing the original content. Users shall comply with the terms of the applicable Creative Commons license when using translated content.
4. Terms of Use and Privacy Policy - Use of MinT is subject to Wikimedia's Terms of Use and Privacy Policy.

• Adjusted translation limits for Punjabi after community request to make them less strict due to improved quality of machine translation.
• Research on MinT for Wikipedia Readers is complete. Two reports were published at the research page
• multi-model support for the MinT test instance. Allowing communities supported by multiple translation models to try, compare assess the quality to determine which one works the best.

• Infrastructure updates to benefit from newer Python versions.

• A new larger instance has been created for the MinT. Memory quota has been increased to accommodate the needs for MinT as the usage and models available increase.

• New design concepts for exposing MinT to Wikipedia readers have been created based on the input from initial research. Multilingual prototypes have been updated to learn from the new concepts in the next round of research.
• Adjusted exposure of MinT in the translate extension to avoid showing translation suggestions for contents with wikitext markup

• Better wikitext support by improving error handling when MinT processes wikitext.
• Completed Research plan is complete and started research sessions.
• Explored New advanced API for sentence segmentation to support needs for EditCheck use case and others.
• Improved responsiveness of the MinT test instance by avoiding some translation requests to get stuck.
• MinT was set as the default translation service in Content Translation for Kurdish (ku) and Sesotho (st), languages where it is optional but frequently used.
• A new larger instance has been created for the MinT. Memory quota has been increased to accommodate the needs for MinT as the usage and models available increase.
• New design concepts for exposing MinT to Wikipedia readers have been created based on input from the initial round of research.
• Published report analyzing usage of machine translation services

• MinT is now supported in Content Translation for Fon, a Wikipedia that graduated recently from incubator.
• Announced sentencex library: sentencex: Empowering NLP with Multilingual Sentence Extraction - A python and js library to meet the needs of sentence segmentation for all the languages we support.
• Proposed model card for language identification as part of the creation of a LiftWing service to provide those capabilities for MinT and others.
• The new sentence segmentation approach has been exposed in Content and Section Translation to validate it with real contents. Resolved community-reported issues such as the problems translating court cases.
• MinT test instance provides consistent language names with Wikipedia by using Wikipedia APIs instead of the limited browser localization capabilities.
• Launched the Language Identification service to automatically detect in which language is written a given text. The service supports the detection of 201 languages, and anyone can access the API to use the service or read the model card for more details. Machine Learning team completed the last checks after deploying to LiftWing and evaluating that the service can "easily withstand a high amount of traffic".
• Basic support for rich text translation by supporting transferring of markup to apply styling such as words in bold from the source text into the equivalent ones in the machine translation (which lacks format since translation models operate with plain-text).
• Completed the process to enable MinT for languages with no Wikipedia yet. Translation models in MinT support 25 languages for which there is no Wikipedia. These can be tested in MinT's test instance for speakers of those languages to assess quality, and ensures that translation tools are well-equipped once wikis are created for those languages (as it has been the case with the recent graduation of Fon Wikipedia out of incubator).
• Completed the process to enable MinT for closely-related languages based on Community input. For some languages where machine translation is not available, Wikipedia editors have asked to have access to machine translation in Content Translation using a related language instead of having no support at all. With this enablement translators of Gan (gan) Wikipedia will have machine translation based on the traditional script variant of Chinese as a starting point.
• Analysis of translation activity on 55 languages for which MinT provides machine translation for the first time shows how (a) translations have increased 2X since MinT is available, and (b) deletion rates have not increased. Activity levels for these 55 wikis changed from ~500 translations/month, to 1K+ translations/month after MinT was enabled. For example, a recent peak of 2.15K translations were published in August 2023 when MinT was available for those languages, which is a significant increase from 225 translations in August 2022 when MinT was not available for them.
• Better visibility of translation quality by including a tag in translations where unedited machine translation is close to the limits. This will facilitate analysis about translation quality and limits.
• Created prototypes for upcoming research illustrating 5 concepts on how MinT can be used by Wikipedia readers and supporting the 4 languages we will conduct research in: Hindi, Chattisgarhi, Awadhi, and Korean.
• Improvements for MinT to process more predictably contents with new lines in them.

A message well received by the attendees.

• コンテンツ翻訳機能のズールー語版で MinT を既定の翻訳サービスに指定

• （コミュニティから意見を収集しながら）MinT を新たに75言語で機械翻訳に採用：62 言語ではモバイル版翻訳の経験を提供、また 機械翻訳（MT）使用報告書のデータおよび／またはコミュニティからの聞き取りにより、他の翻訳サービスの質が最適ではなかった13言語に展開。
• 前回の展開を検証： Bhojpuri 語、ラトビア語で MinT を展開できなかった問題点を識別、どちらもウィキペディアが採用する言語コードとMinT ならびに付帯の翻訳モデルのそれとの照合失敗による。
• 当初の設計の探求と試作版では MinT をウィキペディアに融合させる複数の方法を検討
• Mint 改善版で翻訳の後処理により、 文末の読み点（フルストップ）直後の余分なアキ（スペース記号）を除去、アラビア語の記法を用いる諸言語のサポートを改善
• （訳注：インド諸語対応の）IndicTrans2 モデルの統合を完了、先方モデルが対応する全23言語を有効化するかどうか確認。
• ウィキペディアのコミュニティ群を対象にした活動の初期評価 は MinT を採用した事例の第1号で、将来の調査対象、早期導入先として仮のパイロット運用ウィキの割り出しを目指します。
• ウィキメディアその他の効果のプロジェクトにおいて、多言語化（ローカライゼーション）translatewiki.net の MinT 導入ではで使うものです。