基于事件语义角色的常识知识获取  

Commonsense Knowledge Acquisition Via Semantic Roles in Events

作　　者：王亚 曹存根[1] 王石[1] WANG Ya;CAO Cungen;WANG Shi(Key Laboratory of Intelligent Information Processing,Institute of Computer Technology,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院计算技术研究所智能信息处理重点实验室,北京100190 [2]中国科学院大学,北京100049

出　　处：《中文信息学报》2023年第6期77-88,共12页Journal of Chinese Information Processing

基　　金：国家重点研究与发展计划(2017YFC1700300,2017YFB1002300);国家自然科学基金(61702234)。

摘　　要：常识知识获取是知识获取的瓶颈问题,该文提出了利用事件语义角色进行常识知识获取的方法。首先获取独立于句子(事件)的语义角色相关的常识知识(例如,施事有意志)和语义角色在句子中涉及的常识知识(例如,施事有嘴是施事喝受事的必要条件),该文将这些常识知识看作常识知识获取模板;接着使用一个神经网络模型对大规模Web文本进行语义角色的自动标注;最后利用标注结果中的文本内容替换常识知识获取模板中的语义角色来进一步获取常识知识。实验结果表明,该方法可以获取大量高精度的常识知识。此外,该文通过组合事件谓词(简称谓词)和标注结果中语义角色对应的文本内容生成一个谓词的下位谓词短语或事件,进而自动获取上下位关系常识。The commonsense knowledge acquisition is known as a bottleneck issue.This paper proposes an approach to acquire commonsense knowledge via semantic roles in events.Firstly,we collect commonsense knowledge about semantic roles independent of sentences(events)(e.g.agents act of their own volition)and semantic roles involved in sentences(e.g.an agent must have a mouth to drink).This kind of commonsense knowledge is regarded as commonsense knowledge acquisition templates.We then use a neural network model to automatically label semantic roles in large-scale Web corpus.Finally,the semantic roles in a commonsense knowledge acquisition template are replaced by the text contents in the annotation results to further enrich commonsense knowledge.Experimental results show that this approach is effective for acquiring large amounts of high-quality commonsense knowledge.In addition,we can automatically obtain commonsense knowledge about isa relations by generating hyponyms(i.e.phrases or events)of an event predicate(predicate for short)by combining this predicate with text contents corresponding to semantic roles.

关 键 词：语义角色 句模框架构建 常识知识获取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]