摘要
语义角色标注是自然语言处理的基础课题之一,目前自动标注的效果尚未达到实用水平。主要存在两大问题:首先语义角色颗粒度的大小不好确定;其次静态的谓词词典难以覆盖动态的语料标注问题,特别是缺乏对句子的语义角色省略情况的处理机制。因此,本文基于一种新的整句抽象语义表示方法(AMR)来研究谓词的语义角色问题,根据5,000句中文AMR标注语料统计出了谓词词典的动态覆盖情况;并通过与中文命题库(CPB)的对比发现,AMR能更完整地标注谓词的核心语义角色,且在语义关系的设置上做到了颗粒度粗细相融:核心语义关系颗粒度粗,非核心语义关系颗粒度较细,整体表征能力强;此外,允许增补概念的规定解决了语义角色省略的情况。最后得出,AMR作为整句的语义表示方法,在语义角色标注方面具备独特的优势,需要进一步加强AMR语料库的建设,为中文句子语义处理奠定基础。
Semantic role labeling is one of the fundamental issues in natural language processing.The effect of the current automatic annotation has not yet reached a practical level.There are two major hurdles.First,the granularity of semantic roles is not easily determined.Second,it is hard for static predicate frame lexicon to cover dynamic annotation problems,especially the lack of processing mechanism for the omission of semantic roles.Therefore,based on the new representation of the meaning of sentences,Abstract Meaning Representation(AMR),we address the problems of semantic roles of predicates.We sum up the dynamic coverage of the predicate dictionary according to a Chinese AMR corpus with 5,000 sentences.Then,we contrast the corpus and Chinese Proposition Bank(CPB).We find that AMR is able to annotate core semantic roles with better recall,and AMR fuses fine-grained and coarse-grained semantic roles.The core semantic roles are coarse-grained and the non-core semantic roles fine-grained,so that its overall representational ability is strong.Moreover,the likelihood of AMR to add new concepts can solve the problem of semantic role omission.We conclude that as a representation of the meaning of whole sentences,AMR has a unique advantage in semantic role labeling,and it is necessary to further strengthen the construction of AMR corpus and lay the foundation for semantic processing of Chinese sentences.
出处
《语料库语言学》
2018年第1期45-58,116,共15页
Corpus Linguistics
基金
国家社科基金项目“中文抽象语义库的建构及自动分析研究”(18BYY127)的资助
