Multilingual Statistical Machine Translation

这个提议是关于从一种自然语言到另一种自然语言的机器翻译，比如从英语到法语的翻译。作为一个双语和多元文化的国家，翻译成多种语言对加拿大来说具有重要的社会意义。为了促进加拿大与世界的互动，翻译对加拿大也具有重要的战略意义。在过去的二十年里，机器翻译领域取得了巨大的进步，基于规则的方法被统计机器翻译所取代。统计机器翻译解决方案现在由谷歌、SDL和IBM等大公司提供。然而，所有当代翻译系统都依赖于特定语言对（如法语-英语）之间的大量数据，而忽略了许多资源匮乏的语言，如因纽特语、海地克雷约尔语、马拉雅拉姆语和芬兰语（数百万人使用的语言）。本提案旨在利用语言学对语族的见解，并利用这些信息开发多语言统计机器翻译模型和算法。我们的目标是将信息从资源丰富的源-目标语言对（如法语-英语）转移到资源较少的语言对（如英语-因纽特语）。本文解决了统计机器翻译研究中尚未解决的一个基本问题：在不受资源限制的情况下，学习任意语言对之间的翻译模型。

This proposal is about machine translation from one natural language to another, such as translation from English to French. As a bilingual and multicultural nation, translation into multiple languages is socially important to Canada. Translation is also strategically important to Canada in order to facilitate our interactions with the world. There have been tremendous advances in the field of machine translation over the last two decades with rule-based approaches being replaced by statistical machine translation. Statistical machine translation solutions are now being offered by major companies like Google, SDL and IBM. However all contemporary translation systems rely on large amounts of data between specific language pairs such as French-English, ignoring many low-resource languages, e.g. Inuktitut, Haitian Kreyol, Malayalam, and Finnish (languages that are spoken by millions of people). This proposal aims to use linguistic insights about language families and use this information to develop multilingual statistical machine translation models and algorithms. We aim to transfer information from source-target language pairs with a lot of resources, such as French-English to low-resource language pairs, such as English-Inuktitut. This proposal addresses a basic unresolved issue in statistical machine translation research: learning a translation model between any pair of languages regardless of resource limitations.