Semi-automatic generation of knowledge graph by masked language model and improved skip-gram

Abstract

The knowledge graph is used in various fields such as search, knowledge inference, and natural language understanding. In particular, recently in the field of natural language understanding, it has been argued that external knowledge is required in addition to the learned language model in order for a machine to generate a more natural language. A knowledge graph is useful for storing and expressing this external knowledge. However, when a person directly builds such a knowledge graph, it consumes a lot of time and money, and it is difficult to immediately update the knowledge. To solve this problem, we propose a method of automatically/semi-automatically building a knowledge graph using natural language processing technology. A knowledge graph is a graph consisting of relationships between words. When expressing knowledge, the most preferentially expressed structure is a taxonomy, that is, a hierarchical structure. Therefore, we first conducted a study to explore the hypernym of words to automatically build a hierarchical structure. Our proposed method is an unsupervised method that does not require large-scale training data.Then, the non-hierarchical relationship is predicted. First, the static word embedding is improved to better express the relationship between words, and then the hidden relationship between words is predicted using the improved word embedding. After extracting the relationship of words in this way, it helps the process of building a knowledge graph. We build a knowledge graph directly from data, and compare it with a pre-built ontology to check whether it can be used in practice.