Advances in Language Modeling through Lexico-Grammatical Methods

提出的研究解决了目前在N元语言建模的停滞，通过开发一套原则性的技术，更高层次的结构可以应用到语言建模的问题。 这些技术包括跨有限态 概率 语法 形式主义（概率特征文法和树插入文法）既实用又对词汇敏感， 基于相似性的方法直接解决大N模型的数据稀疏性问题。 一种新的多遍解析技术在解决使用这种模型时出现的一些实际计算问题方面具有潜在的实用性。 这种有前途的技术组合 - 词汇化的语法形式和多遍解析方法使它们实用，沿着基于相似性的估计以帮助解决数据稀疏问题-有潜力改进超越渐近线的语言建模 它已经达到了， 在改进语音识别、信息检索和许多其他语言处理应用方面。

The proposed research addresses the current stagnation in N- gram language modeling by developing a set of principled techniques by which higher level structure can be applied to the problem of language modeling. These techniques include trans-finite-state probabilistic grammar formalisms (probabilistic feature grammars and tree-insertion grammars) that are both practical and lexically sensitive, and similarity-based methods that address data sparsity problem for large-N models directly. A novel technique of multi- pass parsing has potential utility in addressing some of the practical computational issues that arise in using such models. This promising combination of techniques --- lexicalized grammar formalisms and the multi-pass parsing method to make them practical, along with similarity-based estimation to aid in solving data sparsity problems --- has the potential to improve language modeling beyond the asymptote that it has reached, with applications in improving speech recognition, information retrieval, and many other language-processing applications.