Named Entity Recognition for Chinese Social Media with Jointly Trained Embeddings

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 548–554, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Named Entity Recognition for Chinese Social Media with Jointly Trained Embeddings Nanyun Peng and Mark Dredze Human Language Technology Center of Excellence Center for Language and Speech Processing Johns Hopkins University, Baltimore, MD, 21218 npeng1@jhu.edu, mdredze@cs.jhu.edu Abstract We consider the task of named entity recognition for Chinese social media. The long line of work in Chinese NER has fo- cused on formal domains, and NER for social media has been largely restricted to English. We present a new corpus of Weibo messages annotated for both name and nominal mentions. Additionally, we evaluate three types of neural embeddings for representing Chinese text. Finally, we propose a joint training objective for the embeddings that makes use of both (NER) labeled and unlabeled raw text. Our meth- ods yield a 9% improvement over a state- of-the-art baseline. 1 Introduction Named entity recognition (NER), and more gen- erally the task of mention detection1, is an essen- tial component of information extraction technolo- gies: the first step before tasks such as relation extraction (Bunescu and Mooney, 2005) and en- tity linking (Dredze et al., 2010; Ratinov et al., 2011). A long line of work has focused on NER in both formal and informal domains (Collins and Singer, 1999; McCallum and Li, 2003; Nadeau and Sekine, 2007; Jin and Chen, 2008; He et al., 2012a), with recent efforts turning towards social media (Finin et al., 2010; Liu et al., 2011; Ritter et al., 2011; Fromreide et al., 2014; Li et al., 2012; Liu et al., 2012). While NER has included work on several languages, work on social media NER has largely focused on English language data.2 We consider NER on Chinese social media from the popular Sina Weibo service, both because of 1Since we consider name and nominals, our work is closer to mention detection. For simplicity, we use the term NER. 2Etter et al. (2013) considered Spanish Twitter, which is quite similar to English from the standpoint of building mod- els and features. the popularity of the service (comparable in size to Twitter and previously used in NLP research (Ling et al., 2013)) and the challenges faced in processing Chinese language data. One approach is to utilize lexical embeddings to improve NER systems (Collobert and Weston, 2008; Turian et al., 2010; Passos et al., 2014), including for Twit- ter (Cherry and Guo, 2015). However, the use of embeddings for Chinese remains a challenge. Unlike most languages, we cannot easily assign an embedding to each Chinese word without au- tomated segmentation, which may be unreliable, especially when we want to model informal text.3 For this reason, state-of-the-art NER systems for Chinese do not tag words; they instead tag charac- ters directly (Mao et al., 2008). While work has explored different embeddings for Chinese (Liu et al., 2014; Sun et al., 2014; Qiu et al., 2014; Chen et al., 2015), their inclusion in …

Content truncated to 500 words