KoBERTSEG: 한국어 BERT를 이용한 Local Context 기반 주제 분리 방법론

Abstract

Topic segmentation refers to the work of separating a document consisting of several topics into unit documents, such as paragraphs, with one single topic. Topic segmentation has been considered as one of main preprocessing step prior to performing natural language processing tasks, such as document summary or document classification. This paper proposes a Korean BERT-based news article segmentation method aiming at separating a single news article, in which multiple subjects exist, into news segments, each of which contains a single subject. The proposed model has the advantage of being able to capture a wider range of semantic relationships compared to existing topic segmentation studies by borrowing a structure proposed for document summarization. Experimental results on a Korean news article dataset show that the proposed method outperform the benchmark models for topic segmentation. In addition, we also show that the proposed method can be used for practical news clip summarization task, supporting the possibility of implementing the application service based on Korean topic segmentation model.