RI: Multi-Level Modeling of Language and Translation

以往的统计机器翻译(SMT)方法采用了基于短语的模型，该模型将短语表示为完全屈折变化的单词的序列，否则缺乏语言学细节。这些方法不能概括，而且本质上依赖于记忆训练数据中观察到的单词和短语的翻译。该项目旨在通过引入更复杂的模型来表示使用多层次信息的短语来提高SMT的质量。这可以包括基本的语言信息，如词性、词条和同意信息(格、数、人称)，以及更复杂的语言细节，包括语义类、论元结构、共指关系、短语边界和从句法中心传播的信息。通过使用这些信息注释所有数据并适当扩展模型，有可能从训练中学到比以前方法更多的东西。现在，如果出现其他形式的单词，就有可能学习看不见的单词的翻译；现在可以学习关于一种语言的剑序的一般事实；现在可以利用语言语境来生成语法输出。这种概括有可能产生更高质量的翻译，特别是对于只有少量训练数据的语言。因此，它比以往的SMT方法有了很大的进步。多层次模型有可能对所有语言技术产生广泛的影响。同时对不同层次的表示进行建模是描述语言的一种非常有用和自然的方式。该项目正在开发一个通用框架，用于创建语言和翻译的多层次概率模型，并探索其在翻译以外的任务中的应用，包括自然语言技术的生成、释义和自动评估。

Previous approaches to statistical machine translation (SMT) haveemployed phrase-based models which represent phrases as sequences offully-inflected words, and are otherwise devoid of linguisticdetail. Such approaches are unable to generalize and essentiallyrely on memorizing the translations of words and phrases that areobserved in training data.This project aims to improve the quality of SMT through theintroduction of more sophisticated models which represent phrasesusing multiple levels of information. This can include basiclinguistic information such as part of speech, lemmas, and agreementinformation (case, number, person), as well as more sophisticatedlinguistic detail including semantic classes, argument structure,co-reference, phrase boundaries, and information propagated fromsyntactic heads.By annotating all data with this information and extending modelsappropriately, there is the potential to learn much more fromtraining than was possible under previous approaches. There is nowthe potential to learn translations of unseen words if other forms ofthe words occur; it is now possible to learn general facts about a language'sword order; it is now feasible to use linguistic context to generategrammatical output. Such generalization has the potential to resultin much higher quality translation, especially for languages thatonly have small amounts of training data. It therefore represents asignificant advance over previous approaches to SMT.Multi-level models have the potential for wide-ranging impact onall language technologies. Simultaneous modeling of differentlevels of representation is an extremely useful and natural way ofdescribing language. This project is developing a general frameworkfor the creation of multi-level probabilistic models of language andtranslation, and exploring its application to tasks beyondtranslation including generation, paraphrasing, and the automaticevaluation of natural language technologies.