一种基于语义的上下位关系抽取方法

分类体系主要由上下位关系组成,传统的基于模板的上下位关系抽取方法分为两类:第一类方法只使用高质量的模板导致低召回率;第二类方法使用所有可用的模板导致低精度。根据模板的质量将其分为更细粒度的强句法模板和弱句法模板。为了提高弱模板的精度,将弱模板和概念/实体结合构建语义模板。结合强句法模板和语义模板,提出一套新颖的框架从语料中抽取上下位关系,具有高精度和召回率的特点。在中英文语料上进行的实验,实验结果证明了框架的有效性。

HYPERNYMY RELATION EXTRACTION BASED ON SEMANTICS

Taxonomies mainly compose of hypernymy relations. The traditional pattern-based methods for hypernymy relation extraction are usually divided into two categories. One only uses high-quality patterns, which causes low recall rate. The other uses all available patterns, which leads to low precision. According to the quality of patterns, it can be divided into more fine-grained strong syntactic pattern and weak syntactic pattern. In order to improve the accuracy of weak syntactic pattern, a semantic pattern was constructed by combining with concept/entity. Combining strong syntactic pattern with semantic pattern, we proposed a novel framework for extracting hypernymy relations from corpus with high precision and recall rate. Experiments on Chinese and English corpus demonstrate the effectiveness of the framework.