决策式中文依存句法分析模型的改进  

Two-phase Deterministic Parser with Local Search for Chinese Dependency Parsing

作　　者：马若策[1] 戴新宇[1] 陈家骏[1] 

机构地区：[1]南京大学计算机软件新技术国家重点实验室,江苏南京210093

出　　处：《广西师范大学学报（自然科学版）》2009年第1期157-160,共4页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金资助项目(60673043);国家863高科技资助项目(2006AA010109);国家社科基金资助项目(07BYY051)

摘　　要：决策式依存分析方法是一种以统计学习分类算法为基础,在每一个分析步骤中挑选最有可能的分析动作,由最终的动作序列确定分析结果。该方法具有快速、高效,易于实现的特点。然而,各步分析动作相对独立,并且容易造成错误累积。针对这些问题,对原始算法进行了一定的改进,引进了局部搜索算法和两阶段分析。实验表明,基于局部搜索的两阶段决策式依存分析模型,在性能上好于原始决策式依存分析方法,同时保持了较低的时间复杂度。Recently proposed deterministic classifier-based dependency parsers offer attractive alternatives to generative statistical ones for their rapidity, effectiveness, and simple implementation. Having studied these deterministic classifier based parsers ,its found that the search strategies used are greedy ones ,that is ,to select the most possible analysis action at each stage,which gives rise to the problems of the isolation between each stage and error accumulation in dependency parsing. In this paper,to improve the original one, an algorithm of two-phase parsing with local search strategy. Experiments show that this model performs better than the original is proposed while maintaining the same time complexity.

关 键 词：依存句法分析 决策式算法 错误累积 局部搜索 两阶段分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]