摘要
在头驱动句法分析模型下,基于经典插值平滑算法,提出了以统计空间中平均事件数为基础的直接插值平滑建模原则,并应用经典的误差理论分析了该原则的合理性.基于该原则并借鉴语言模型中其他插值平滑算法对模型的零点进行假设的方法,在头驱动句法分析模型下,重新构造了4种平滑算法.实验数据显示,新平滑算法在高于经典平滑算法性能的同时,显著降低了自由参数的扰动程度,从实验的角度证明了该平滑建模原则的有效性.
Based on the classical smoothing technology, this paper proposes a smoothing approach within head-driven parsing, which directly calculates interpolation weight from the average occurrences of event in the training sample and is proved by the statistic theory of errors. By using this approach and deriving zero-value assumption from other smoothing technologies, this paper proposes four smoothing algorithms for head-driven parsing. Experiments indicate that these four smoothing algorithms have higher performance than the Baseline algorithm and reduce the disturbing curve of the optimized parameter significantly, which prove the effectiveness of the proposed approach.
出处
《软件学报》
EI
CSCD
北大核心
2009年第11期2915-2924,共10页
Journal of Software
基金
国家自然科学基金Nos.60736014
60773069
国家高技术研究发展计划(863)No.2006AA010108 ~~
关键词
句法分析
平滑算法
插值平滑
头驱动句法分析
parsing
smoothing algorithm
interpolation smoothing
head-driven parsing
