Improving Multilingual Named Entity Recognition with Wikipedia Entity Type Mapping

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1275–1284, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Improving Multilingual Named Entity Recognition with Wikipedia Entity Type Mapping Jian Ni and Radu Florian IBM T. J. Watson Research Center 1101 Kitchawan Road, Yorktown Heights, NY 10598, USA {nij, raduf}@us.ibm.com Abstract The state-of-the-art named entity recognition (NER) systems are statistical machine learn- ing models that have strong generalization ca- pability (i.e., can recognize unseen entities that do not appear in training data) based on lexical and contextual information. How- ever, such a model could still make mis- takes if its features favor a wrong entity type. In this paper, we utilize Wikipedia as an open knowledge base to improve multilin- gual NER systems. Central to our approach is the construction of high-accuracy, high- coverage multilingual Wikipedia entity type mappings. These mappings are built from weakly annotated data and can be extended to new languages with no human annotation or language-dependent knowledge involved. Based on these mappings, we develop several approaches to improve an NER system. We evaluate the performance of the approaches via experiments on NER systems trained for 6 languages. Experimental results show that the proposed approaches are effective in improv- ing the accuracy of such systems on unseen entities, especially when a system is applied to a new domain or it is trained with little train- ing data (up to 18.3 F1 score improvement). 1 Introduction Named entity recognition (NER) is an important NLP task that automatically detects entities in text and classifies them into pre-defined entity types such as persons, organizations, geopolitical entities, lo- cations, events, etc. NER is a fundamental compo- nent of many information extraction and knowledge discovery applications, including relation extraction, entity linking, question answering and data mining. The state-of-the-art NER systems are usually sta- tistical machine learning models that are trained with human-annotated data. Popular models in- clude maximum entropy Markov models (MEMM) (McCallum et al., 2000), conditional random fields (CRF) (Lafferty et al., 2001) and neural networks (Collobert et al., 2011; Lample et al., 2016). Such models have strong generalization capability to rec- ognize unseen entities1 based on lexical and contex- tual information (features). However, a model could still make mistakes if its features favor a wrong en- tity type, which happens more frequently for unseen entities as we have observed in our experiments. Wikipedia is an open-access, free-content Inter- net encyclopedia, which has become the de facto on-line source for general reference. A Wikipedia page about an entity normally includes both struc- tured information and unstructured text information, and such information can be used to help determine the entity type of the referred entity. So far there are two classes of approaches that exploit Wikipedia to improve NER. The first class of approaches use Wikipedia to generate features for NER systems, e.g., (Kazama and Torisawa, 2007; Ratinov and Roth, 2009; Radford et al., 2015). Kazama and Torisawa (2007) try to find the Wikipedia entity for each …

Content truncated to 500 words