딥러닝 기반 한국어 개체명 인식의 평가와 오류 분석 연구

Abstract

Named entity recognition is a natural language processing task that recognizes and classifies named entities in an unstructured text. The targets of NER are not limited to typical proper names for persons, locations and organizations, but also date, time and quantity expressions and can be further expanded to names of events, animals, plants, materials and other encyclopedic entities. A real-world NER system is also expected to be tuned to process domain-specific terminologies. In this study, the researchers built and tested a BERT based Korean NER system and proposed methods for evaluation and error analysis. The study trained the system with 140K word NER corpus and evaluated with 60K test. Error types are proposed to be categorized into four classes: detection, boundary, segmentation, and labelling. Error rates are found to vary greatly from 1% to 30% between entity labels, which are grouped into the most accurate time and quantity expressions, relatively accurate proper names, and highly erroneous terminologies. We expect that the error analysis will provide insights for finding a better way of data collection and post-processing correction.