Using Non-Local Features to Improve Named Entity Recognition Recall

Using Non-Local Features to Improve Named Entity Recognition Recall∗ Xinnian Mao1, Wei Xu1, Yuan Dong1, 2, Saike He2, and Haila Wang1 1France Telecom R&D Center (Beijing), Beijing, 100080, P.R.China {xinnian.mao, wxu.ext, yuan.dong,haila.wang}@orange-ftgroup.com 2University of Posts and Telecommunications, Beijing, 100876, P.R.China yuandong@bupt.edu.cn; hsk000@gmail.com Abstract. Named Entity Recognition (NER) is always limited by its lower recall resulting from the asymmetric data distribution where the NONE class dominates the entity classes. This paper presents an approach that exploits non-local information to improve the NER recall. Several kinds of non-local features encoding entity token occurrence, entity boundary and entity class are explored under Conditional Random Fields (CRFs) framework. Experiments on SIGHAN 2006 MSRA (CityU) corpus indicate that non-local features can effectively enhance the recall of the state-of-the-art NER systems. Incorporating the non-local features into the NER systems using local features alone, our best system achieves a 23.56% (25.26%) relative error reduction on the recall and 17.10% (11.36%) relative error reduction on the F1 score; the improved F1 score 89.38% (90.09%) is significantly superior to the best NER system with F1 of 86.51% (89.03%) participated in the closed track. Keywords: Named Entity Recognition, Non-local Feature, Conditional Random Field 1. Introduction Named entity recognition (NER) is a subtask of information extraction that seeks to locate and classify predefined entities, such as names of persons, locations, organizations, etc. in unstructured texts. It is the fundamental step to many natural language processing applications, like Information Extraction (IE), Information Retrieval (IR) and Question Answering (QA). Most empirical approaches currently employed in NER task make decision only on local context for extract inference, which is based on the data independent assumption (Krishnan and Manning, 2006). But often this assumption does not hold because non-local dependencies are prevalent in natural language (including the NER task). How to utilize the non-local dependencies effectively is a key issue in NER task. Unfortunately, few researches have been devoted to this issue, existing works mainly focus on using the non-local information for further improving NER label consistency. There are two methods to use non-local information. One is to add additional edges to graphical model structure to represent the distant dependencies and the other is to encode the non-locality with non-local features. However, in the first approach, heuristic rules are used to find the dependencies (Bunescu and Mooney, 2004; Sutton and McCallum, 2004) or penalties for label inconsistency are required to handset ad-hoc (Finkel et al., 2005). Furthermore, high computational cost is spent for approximate inference. In order to establish the long dependencies easily and overcome the disadvantage of the approximate inference, Krishnan and Manning (2006) propose a two-stage approach using Conditional Random Fields (CRFs) with extract inference. They represent the non-locality with non-local features, and extract the non- local features from the output of the first stage CRF using local context alone; then they incorporate the non-local features into the second CRF. But the features in this approach are ∗ Copyright 2007 by Xinnian Mao, Wei Xu, Yuan Dong, Saike He, and Haila Wang 303 …

Content truncated to 500 words