语义角色标注特征工程研究-苏州大学.doc

中文名词性谓词语义角色标注研究* 李军辉+, 周国栋, 朱巧明, 钱培德 (苏州大学 计算机科学与技术学院, 江苏 苏州 215006 ) Semantic Role Labeling in Chinese Language for Nominal Predicates LI Jun-Hui+, ZHOU Guo-Dong, ZHU Qiao-Ming, QIAN Pei-De (School of Computer Science Technology, Soochow University, Suzhou 215006, China) + Corresponding author: E-mail: lijunhui@ Abstract: This paper explores semantic role labeling (SRL) in Chinese language for nominal predicates. In addition to widely adapted features from verbal SRL, various nominal predicate-specific features are also explored. Moreover, we improve the nominal SRL performance by properly integrating features derived from a state-of-the-art verbal SRL system. Finally, we detail nominal predicate recognition, which is essential in a fully automatic nominal SRL system. Evaluation on Chinese NomBank shows that proper integration of a verbal SRL system significantly improves the performance of nominal SRL. It also shows that our nominal SRL system achieves the performance of 72.67 in F1-measure on golden parse trees and golden predicates, and much outperforms the state-of-the-art nominal SRL in Chinese language. However, the performance drops to 55.14 in F1-measure on automatic parse trees and automatic predicates. This suggests the great challenge in Chinese nominal SRL as well as the disappointment of Chinese syntactic parsing. Key words: semantic role labeling; nominal predicate-specific features; verbal SRL features; nominal predicate recognition 摘 要: 本文研究了中文名词性谓词的语义角色标注(SRL).在使用传统动词性谓词SRL相关特征的基础上,进一步提出了名词性谓词SRL相关的特征集.此外,本文探索了中文动词性谓词SRL对中文名词性谓词SRL的影响,并且联合谓词自动识别实现了全自动的中文名词性谓词SRL.在中文NomBank的实验表明,中文动词性谓词SRL的合理使用能够大大提高中文名词性谓词SRL性能;基于正确句法树和正确谓词识别,中文名词性谓词的SRL性能F1值达到了72.67,优于目前国内外的同类系统;基于自动句法树和自动谓词识别,性能F1值为55.14,说明中文名词性谓词SRL任务艰巨,同时中文句法分析还不够完善. 关键词: 语义角色标注; 名词性谓词相关特征; 动词性语义角色标注特征; 名词性谓词识别 中图法分类号: TP18 文献标识码:A 作为自然语言处理的一项研究重点,语义分析能够将人类自然语言转化为计算机能够理解的形式化语言,从而实现人与计算机之间的相互理解.语义角色标注(semantic role labeling, 简称SRL)是一种浅层的语义分析,因具有任务明确、评测容易等特点而成为目前的研究热点.给定一个句子,SRL的