CAREER: Semantics for Statistical Machine Translation

在过去的几年里，机器翻译发生了一场革命，对大量平行双语文本进行训练的统计系统得到了广泛采用。最近的评估表明，目前经过统计训练的研究技术显著优于商业上可用的mtsystem，例如网络上可用的mtsystem。但是，即使是最先进的翻译系统也经常会产生混乱的翻译。机器翻译的进一步改进将需要统计系统架构的重大变化。我们的研究旨在通过允许统计系统处理更深层次的语义表示来提高机器翻译输出的质量。我们的方法侧重于通过使用谓词-参数结构级别的语义表示来改进统计机器翻译。这项工作建立在最近成功的统计学方法上，用于肤浅的语言理解，以及基于树的算法，用于使用源和目标句子的句法解析的机器翻译。在项目过程中，我们的目标是：首先，开发健壮的语义分析系统，能够泛化到新领域，并将其应用于大型双语语料库；其次，开发使用所得表示水平并可实际训练的翻译概率模型；第三，整合语言理解和翻译，以便有效地搜索新句子的最佳整体翻译。

The past few years have seen a revolution in machinetranslation, with the widespread adoption ofstatistical systems trained on large amounts ofparallel bilingual text. Recent evaluations have shownthat current statistically trained research technologysignificantly outperforms commercially available MTsystems such as those available on the web. But evenstate-of-the-art systems produce garbled translationsmore often than not. Further improvements in machinetranslation will require major changes in thearchitecture of statistical systems. Our research aimsto improve the quality of machine translation output byallowing statistical systems to handle deeper, semanticrepresentations.Our approach focuses on improving statistical machinetranslation by using a semantic representation at thelevel of predicate-argument structure. This workbuilds on the recent success in statistical approachesto shallow language understanding, and tree-basedalgorithms for machine translation using syntacticparses of the source and target sentences. Over thecourse of the project we aim to: first, develop robustsemantic parsing systems capable of generalizing to newdomains and apply them to large bilingual corpora,second, develop probabilistic models of translationthat use the resulting level of representation and canbe practically trained, and third, integrate languageunderstanding and translation to allow efficient searchfor the best overall translation of new sentences.