Semi-supervised Sequence Labeling for Named Entity Extraction based on Tri-Training: Case Study on Chinese Person Name Extraction

Abstract

Named entity extraction is a fundamental task for many knowledge engineering applications. Existing studies rely on annotated training data, which is quite expensive when used to obtain large data sets, limiting the effectiveness of recognition. In this research, we propose an automatic labeling procedure to prepare training data from structured resources which contain known named entities. While this automatically labeled training data may contain noise, a self-testing procedure may be used as a follow-up to remove low-confidence annotation and increase the extraction performance with less training data. In addition to the preparation of labeled training data, we also employed semi-supervised learning to utilize large unlabeled training data. By modifying tri-training for sequence labeling and deriving the proper initialization, we can further improve entity extraction. In the task of Chinese personal name extraction with 364,685 sentences (8,672 news articles) and 54,449 (11,856 distinct) person names, an F-measure of 90.4% can be achieved.