Bootstrapping Large-scale Named Entities using URL-Text Hybrid Patterns

International Joint Conference on Natural Language Processing, pages 293–301, Nagoya, Japan, 14-18 October 2013. Bootstrapping Large-scale Named Entities using URL-Text Hybrid Patterns Chao Zhang Shiqi Zhao Haifeng Wang Baidu Inc No. 10, Shangdi 10th Street, Haidian District, Beijing 100085, China {zhangchao01,zhaoshiqi,wanghaifeng}@baidu.com Abstract Automatically mining named entities (NE) is an important but challenging task, pattern-based and bootstrapping strategy is the most widely accepted solution. In this paper, we propose a novel method for NE mining using web document titles. In addition to the traditional text patterns, we propose to use url-text hybrid patterns that introduce url criterion to better pinpoint high-quality NEs. We also design a mul- ticlass collaborative learning mechanism in bootstrapping, in which different pat- terns and different classes work together to determine better patterns and NE in- stances. Experimental results show that the precision of NEs mined with the pro- posed method is 0.96 and 0.94 on Chinese and English corpora, respectively. Com- parison result also shows that the proposed method significantly outperforms a rep- resentative method that mines NEs from large-scale query logs. 1 Introduction The task of named entity mining (NEM) aims to mine named entities (NE) of given categories from raw data. NEM is essential in many application- s. For example, NEM can generate NE gazetteer- s necessary for the task of named entity recog- nition (NER) (Cohen and Sarawagi, 2004; Kaza- ma and Torisawa, 2008; Talukdar et al., 2006). It can also help improve the search results in web search (Pas¸ca, 2004), and increase the coverage of knowledge graphs. Extensive research has been conducted on NEM, in which pattern-based methods are the most popular. Handcrafted or automatically learnt patterns are usually used to extract NE instances from various corpora, such as web documents, search engine’s retrieved snippets, and query logs. Bootstrapping strategy is often applied to gener- ate more patterns and instances iteratively so as to improve the coverage of the system. The method we propose also belongs to the fam- ily of pattern-based NEM. However, our method is a departure from the previous ones. It makes con- tributions from the following aspects: First, we design url-text hybrid patterns instead of the tra- ditional text patterns. We take url criterion into account, so as to measure the quality of the source webpages. Second, we propose Multiclass Collab- orative Learning (MCL) mechanism, which glob- ally scores and ranks the patterns and NE instances within mutiple classes in bootstrapping. We evaluate our method in two languages, i.e., Chinese and English, so as to demonstrate the language-independent nature of the method. We mine NEs using the system for five categories in both languages, including star, film, TV play, song, and PC game. Experimental results show that the average precision of the extracted NEs is 96% in Chinese and 94% in English. Meanwhile, the average coverage computed against a bench- mark repository is 61% and 55% for the two lan- guages. Comparative experiments further show that our method significantly outperforms a rep- resentative conventional method. 2 Related Work …

Content truncated to 500 words