M-CNER: A Corpus for Chinese Named Entity Recognition in Multi-Domains

M-CNER: A Corpus for Chinese Named Entity Recognition in Multi-Domains Qi Lu, YaoSheng Yang, Zhenghua Li, Wenliang Chen, Min Zhang School of Computer Science and Technology Collaborative Innovation Center of Novel Software Technology and Industrialization Soochow University, Suzhou 215006, China luqibhf@qq.com, ysyang@stu.suda.edu.cn, { zhli13, wlchen, mingzhang}@suda.edu.cn Abstract In this paper, we present a new corpus for Chinese Named Entity Recognition (NER) from three domains : human-computer interaction, social media, and e-commerce. The annotation procedure is conducted in two rounds. In the first round, one sentence is annotated by more than one persons independently. In the second round, the experts discuss the sentences for which the annotators do not make agreements. Finally, we obtain a corpus which have five data sets in three domains. We further evaluate three popular models on the newly created data sets. The experimental results show that the system based on Bi-LSTM-CRF performs the best among the comparison systems on all the data sets. The corpus can be used for further studies in research community. Keywords: Named Entity Recognition; Chinese Data Set; Information Extraction 1. Introduction In recent years, there has been significant progress on the task of Named-Entity Recognition (NER) by using sequence labeling models in the settings of supervised learning, such as CRF and LSTM-CRF (Lafferty et al., 2001; Huang, Xu, and Kai 2015). NER is one of the most important natural language processing (NLP) tasks. Its performance highly affects further applications, such as relation extraction (Bunescu and Mooney, 2005), and question answering (Jurafsky and Martin, 2009). As with the setting of supervised learning, building NER systems needs a massive amount of labeled training data which are often annotated by humans. However, for most languages, large-scale labeled datasets are only readily available in some domains, for example the news domain. For other domains like social media and dialog texts, there is a lack of such data sets. The NER systems trained on the news domain often perform worse in other domains. It is a reasonable solution to create human- annotated data in new domains to improve the performance of NER system. In this paper, we present a new corpus for Chinese Named Entity Recognition in multi-domains, named M- CNER. We create several data sets in three domains: human-computer interaction, social media, and e- commerce, which are often used in real applications. We require the annotators to label some predefined entities. In the annotation procedure, the annotators label the sentences independently in the first round. One sentence is labeled by more than one persons. In the second round, experts check the entities which have disagreement among the annotators. The detailed settings are described in Section 2. Most traditional high performance sequence labeling models for NER are linear statistical models, including Hidden Markov Models (HMM) and Conditional Random Fields (CRF) (Ratinov and Roth, 2009; Passos et al., 2014; Luo et al., 2015). CRF has been widely used for this task for the last decades, but in the most recent years, non-linear neural networks have become popular for …

Content truncated to 500 words