RI: Small: Statistical Machine Translation Through a Tree Adjoining Grammar with Flexible Parsing Operations

我们的研究涉及到统计机器翻译的句法方法的发展，该方法将树邻接语法(标记)形式扩展到翻译问题，并将翻译直接框架化为句法分析问题。该模型不对短语词典中的词条施加任何限制，从而保留了基于短语的翻译系统的灵活词条；它允许直接合并句法语言模型。用于将树片段组合成完整解析树的操作是标签中发现的标准解析操作的泛化；具体地说，它们被修改为高度灵活，潜在地允许对初始片段进行任何可能的排列(重新排序)。这使得模型在捕捉源语言和目标语言之间的词序差异方面有了很大的自由。灵活的句法分析操作的使用带来了一些挑战，这是我们研究的一个主要焦点。首先，模型需要高效的译码算法。其次，灵活的解析操作使模型能够捕捉复杂的重新排序现象，但也引入了许多虚假的可能性。我们正在研究基于源语言句子或源语言句子的句法树中的信息学习的概率约束的使用，从而纳入来自源语言的句法信息。该项目的最终目标是开发新的翻译模型，以提高翻译的流畅性或语法，提高语义信息(例如，谓词-论证结构)在翻译中的保留程度，并改进对源语言和目标语言之间不同词序的处理。

Our research involves the development of a syntactic approach forstatistical machine translation that extends a tree adjoining grammar(TAG) formalism to the translation problem, and frames translationdirectly as a parsing problem. The model imposes no constraints onentries in the phrasal lexicon, thereby retaining the flexible lexicalentries of phrase-based translation systems; it allows straightforwardincorporation of a syntactic language model. The operations used tocombine tree fragments into a complete parse tree are generalizationsof standard parsing operations found in TAG; specifically, they aremodified to be highly flexible, potentially allowing any possiblepermutation (reordering) of the initial fragments. This allows themodel a great deal of freedom in capturing differences in word orderbetween source and target languages.The use of flexible parsing operations raises a couple of challengesthat are a major focus of our research. First, efficient decodingalgorithms are required for the models. Second, flexible parsingoperations allow the model to capture complex reordering phenomena,but in addition introduce many spurious possibilities. We areinvestigating the use of learned, probabilistic constraints based oninformation in the source-language sentence, or in a parse tree forthe source-language sentence, thereby incorporating syntacticinformation from the source language.The end goal of the project is to develop new models for translationthat improve the fluency or grammaticality of translations, improvethe degree to which semantic information (e.g., predicate-argumentstructure) is preserved in translation, and improve the treatment ofdiffering word orders between source and target languages.