层级分类概率句法分析

对已有的句法分析中引入知识的方法进行了归纳分析,认为多种句法分析方法都可被看作是基于特征标记的分类,然后分析了其中的欠分类和过分类问题.在此基础上,提出一种层级分类短语结构文法和一种层级分类概率句法分析方法(hierarchically classified probabilistic context-free grammar),并设计了一种通过对实例进行聚类来消除句法规则的分类歧义方法.还进一步将层级分类扩展到概率上下文相关句法分析方法,利用上下文相关性的层级分类来解决引入上下文相关时的数据稀疏性问题.通过上述一系列方法有效地克服了过分类与前分类之间的矛盾.

Hierarchically Classified Probabilistic Grammar Parsing

This paper analyzed various existing approaches of structural grammar parsing, and addressed the problem of over-classification and under-classification. Then a hierarchically classified phase structure grammar (HC-PSG) and a hierarchically classified probabilistic context-free grammar (HC-PCFG) parsing are proposed to respond to this challenge. A measure of class clustering is designed to eliminate the classification ambiguity of grammar rules. The HC approach implements a general learning rule from a small number of phrase instances. An instant clustering method is used to disambiguate rules learned from corpus. The HC method is also extended to context sensitive grammar parsing to improve performance. It employs the classification of the context relevancy to handle the problem of corpus sparsity. By all the means, it can leverage the conflicts between under-classification and over-classification.