9_多领域中文依存树库构建与影响统计句法分析因素之分.pdfVIP

多领域中文依存树库构建与影响统计句法分析因素之分析* 1 1 2 邱立坤 ，史林林 ，王厚峰 （1.鲁东大学文学院，山东省烟台市 264025 ；2. 北京大学计算语言学研究所，北京市 100871） 摘要：为提升依存分析并分析影响其精度的相关因素，该文构建了大规模中文通用依存树库和中等规模领 域依存树库。基于这一系列树库，通过句法分析实验考察质量、规模、领域差异等因素对中文依存分析的 影响，实验结果表明：（1）树库规模和质量均与句法分析精度成正相关关系，质量应先于规模因素被优先 考虑；（2）通用树库和领域树库之间的差异程度与前者对后者的替代性成相关关系；（3）两种树库混合使 用的效果同样与领域差异有关。 关键词：依存树库；领域迁移；依存句法分析 中图分类号：TP391 文献标识码：A Construction of Multi-Domain Chinese Dependency Treebanks and Analysis of Influencing Factors on Dependency Parsing 1 1 2 QIU Likun , SHI Linlin , WANG Houfeng (1. School of Chinese Language and Literature, Ludong University, Shandong 264025, China; 2. Institute of Computational Linguistics, Peking University, Beijing 100871, China) Abstract: To boost Chinese dependency parsing and analyze influencing factors on Chinese dependency parsing, we constructed a large-scale general treebank and several middle-scale treebanks for specific domains. Using these treebanks, we performed experiments to evaluate the influence of treebank quality, scale and domain difference upon parsing accuracy. Experimental results show that both treebank quality and scale are positively related to parsing accuracy, and quality is more important than scale. Our experiments also demonstrate that general treebanks and domain treebanks are complementary. In addition, whether a general treebank and domain treebank should be used together is also dependent on the difference between them. Key words: dependency treebank; domain adaptation; dependency parsing 1 引言 依存句法分析的目标是为给定句子中的每个词找出一个合适的父结点，并标记子结点与 父结点之间的句法关系，它是目前最常用的句法分析理论之一。作为主流依存分析方法的统 [1] 计句法分析，通常用包含大量依存句法树的树库作为训练数据，采用基于图的方