Development of structured semantic models to improve the quality of statistical machine translation systems

现代机器翻译系统基于统计方法，试图找到给定输入句子的最可能翻译。通常，单词和单词串在翻译时没有抽象地表示它们的意思。这可能导致流畅的输出句子，但却不能充分表达输入句子的意思。提议的项目关注的是模型的发展，这些模型可以根据与原始句子的语义兼容性对可能的翻译进行排名，从而找到更好的翻译。输入和输出句子的意义用情景、语义角色和概念的概率模型表示。与前面的角色语义方法不同，不需要预先指定一般概念。相反，适当的分类将从可用的文本语料库中学习。这将允许统计机器翻译系统第一次直接利用结构化语义信息。虽然该项目的主要目标是提高机器翻译系统的质量，但新的语义模型也将使研究面向任务的语义类别及其跨语言的可泛化性成为可能，这也可能有利于计算语言学的其他领域。

Modern machine translation systems are based on statistical methods that try to find the most probable translation of a given input sentence. Typically, words and strings of words are translated without representing their meaning on an abstract level. This can lead to fluent output sentences, which nonetheless fail to adequately render the meaning of the input sentence. The proposed project is concerned with the development of models that can rank possible translations by their semantic compatibility with the original sentence and thus find better translations. The meanings of input and output sentences are represented in probabilistic models featuring situations, semantic roles and concepts. In contrast to previous approaches in role semantics, no general concepts are to be specified in advance. Instead, appropriate categories for classification will be learned from available text corpora. This would allow statistical machine translation systems to directly take advantage of structured semantic information for the first time. While the main goal of the project is to improve the quality of machine translation systems, the new semantic models will also make it possible to study task-oriented semantic categories and their generalizability across languages, which may also benefit other areas in computational linguistics.