汉语框架语义角色标注中特征模板选择法比较  

Comparison of Feature Template Selection Methods for Semantic Role Labeling of Chinese FrameNet

作　　者：宋毅君 王瑞波 SONG Yijun;WANG Ruibo(School of Modern Educational Technology,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学现代教育技术学院,山西太原030006

出　　处：《山西大学学报（自然科学版）》2021年第5期907-914,共8页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61806115);山西省高等学校教学改革创新项目(J2019040)。

摘　　要：在构建汉语框架语义角色标注模型时,多种语言学特征和不同的上下文窗口常被组合成大量的候选特征模板。从候选特征模板中高效地找到最优模板是构建优良的标注模型的关键所在。文章比较了两种特征模板选择方法:贪心选取法和正交表选取法。从汉语框架网络知识库中选取25个框架作为实验用语料,实验比较了两种特征模板选取方法的时间复杂度和模型性能。实验结果显示,正交表选取法不仅可以通过并行计算以显著降低训练时间开销,而且其模型标注性能也略微高于贪心选取法的标注性能。When constructing Chinese FrameNet semantic role labeling models,plenty feature templates are generated through com‐bining multiple linguistic features with different context windows,and efficiently finding an optimal template from the candidates with a novel search method is a key challenge in constructing a novel labeling model.Two different template selection methods are compared:greedy selection method and orthogonal-array-based selection method.All sentences of 25 frames from knowledge-base of Chinese FrameNet are extracted as our experimental corpus,and time complexities and model performances of both the methods are compared.The experimental results indicated that orthogonal-array-based method can not only reduce the time cost of training procedure in a parallel manner,but also achieve better performance than the greedy selection method.

关 键 词：汉语框架网 语义角色标注 正交表方法 贪心算法 模型性能 时间复杂度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]