It is one of the prerequisites for artificial intelligence to understand and possess human values to infer their values from their behavior. However, in NLP related fields, the current research mainly focuses on the judgment of the values or morality of the text, rarely inferring their values from the subject's behavior, and also lacks corresponding data resources. This paper first constructs the Chinese core value-behavior frame. It is based on China's socialist core values and is divided into two parts: 1) category system. There are 8 categories of core values, which are further subdivided into 19 categories of bi-directional values and corresponding to 38 types of behaviors; 2) Factor system. There are 7 types of factors. Then, text sentences containing subject behavior are extracted from the corpus and manually labeled according to the system. Then, a fine-grained Chinese value-behavior knowledge base containing 6994 behavior sentences and their corresponding fine-grained values and directions, and 34965 elements is constructed. Finally, this paper puts forward the tasks of value category classification, direction detection and joint discrimination. Experimental results show that the method based on the pretraining language model performs well in judging the direction of values, and has great room for improvement in fine-grained value category classification and multi-label value category classification.
价值观计算；人工智能伦理；价值-行|体系；价值-行为知识库
刘鹂远、张三乐、于东、薄琳
展开 >
北京语言大学信息科学学院国家语言资源监测与研究平面媒体中心北京市海淀区学院路15号，100083
价值观计算；人工智能伦理；价值-行|体系；价值-行为知识库
Chinese national conference on computational linguistic
Nanchang(CN)
The 21st Chinese national conference on computational linguistic
NETL
