A Self-learning Template Approach for Recognizing Named Entities from Web Text

International Joint Conference on Natural Language Processing, pages 1139–1143, Nagoya, Japan, 14-18 October 2013. A Self-learning Template Approach for Recognizing Named Entities from Web Text Qian Liu†‡, Bingyang Liu†‡, Dayong Wu‡, Yue Liu‡, Xueqi Cheng‡ †University of Chinese Academy of Sciences, Beijing, China ‡Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China {liuqian, liubingyang}@software.ict.ac.cn {wudayong, liuyue, cxq}@ict.ac.cn Abstract Recognizing Chinese Named entities from the Web is challenging, due to the lack of labeled data and differences between Chi- nese and English. We propose a semi- supervised approach which leverages seed entities and the large unlabeled data to learn templates. Some high-quality tem- plates are generated iteratively to extrac- t new named entities based on the model of quality metrics. Experimental results show that our approach significantly out- performs the baseline method and it is ro- bust against the changes of the Web. 1 Introduction Compared to English Named Entity Recognition (NER), Chinese NER is more difficult. For exam- ple, although capitalization plays an very impor- tant role in English NER, this language-specific feature is useless in Chinese NER. Moreover, the lack of space between words in Chinese often makes the work based on word segmentation (Sun et al., 2002) failure. Especially, there are a lot of new words and ambiguous words in the Web tex- t, which increase the error of word segmentation. The loss of precision propagates to NER. Currently, NER is mainly based on supervised models. Such models perform well in single do- main (Wang, 2009; Du et al., 2010). Unfortunate- ly, the data of the Web is open-domain and always changing. The performance of them degrades bad- ly, since the distribution of the Web data is differ- ent from that of the training data. For instance, the average F1 score of the Stanford NER, which is trained on the CoNLL03 shared task data set and achieves state-of-the-art performance on that task, drops from 90.8% to 45.8% on tweets (Liu et al., 2011). Despite a high-quality training data set, which has the same distribution as the Web data and covers all kinds of domains, can improve the performance of NER, as far as we know, there are no such labeled data. Furthermore, named entities change over the time, especially person names and company names. (Tsuchiya et al., 2009) shows that 20%∼30% of named entity types are replaced with new ones every year in Mainichi Newspaper articles. Such change is even more obvious in the Web data and leads to the unreliable labeled data. Constantly annotating new data is time-consuming and expensive. In this paper, we propose a semi-supervised ap- proach that uses self-learning templates to solve above problems. Instead of annotating a massive amount of data, we leverage a small number of named entities and the large unlabeled data to dis- cover new named entities that can not be identified using the training data. Experiments show that our approach raises the F1 from 75.9% to 88.6% on the Chinese Web data without retraining the exist- ing model, …

Content truncated to 500 words