A Robust Risk Minimization based Named Entity Recognition System

A Robust Risk Minimization based Named Entity Recognition System Tong Zhang IBM T.J. Watson Research Center Yorktown Heights New York, 10598, USA tzhang@watson.ibm.com David Johnson IBM T.J. Watson Research Center Yorktown Heights New York, 10598, USA dejohns@us.ibm.com Abstract This paper describes a robust linear classifica- tion system for Named Entity Recognition. A similar system has been applied to the CoNLL text chunking shared task with state of the art performance. By using different linguistic fea- tures, we can easily adapt this system to other token-based linguistic tagging problems. The main focus of the current paper is to investigate the impact of various local linguistic features for named entity recognition on the CoNLL- 2003 (Tjong Kim Sang and De Meulder, 2003) shared task data. We show that the system per- formance can be enhanced significantly with some relative simple token-based features that are available for many languages. Although more sophisticated linguistic features will also be helpful, they provide much less improve- ment than might be expected. 1 Introduction An important research area in the field of information ex- traction is Named Entity Recognition. This topic was a central theme in the message understanding conferences (MUCs). It has become more important nowadays due to the large amount of available electronic text, which makes it necessary to build systems that can automati- cally process and extract information from text. In spite of significant work in this area, the problem itself has not been solved. Although some earlier re- ports suggested accuracy (F1-number) of machine learn- ing based systems to be in the lower 90s with relatively small amount of labeled data (for example, (Bikel et al., 1999; Mikheev et al., 1998; Sundheim, 1995)), these studies were often performed on relatively restricted do- mains. Our experience indicates that the performance of a statistically based named entity extraction system can vary significantly depending on the underlying domain. There are still open challenges to make the performance of a statistical system consistent across different types of data sources. In this paper we present a system for named entity recognition based on our earlier work on text chunking (Zhang et al., 2002). One advantage of the proposed sys- tem is that it can easily incorporate a large number of lin- guistic features. This advantage is similar to a number of other approaches, such as the maximum entropy method, which has been widely used to solve NLP problems, see (Borthwick, 1999; Ratnaparkhi, 1999) for example. The performance of our system can be significantly af- fected by the choice of available linguistic features. The main focus of this paper is to investigate the impact of some local features. Specifically we show that the system performance can be enhanced significantly with some relatively simple token-based features. More sophisti- cated linguistic features, although helpful, yield much less improvement in system performance than might be expected. We believe that this study provides useful insight into the usefulness of various available local linguistic fea- tures. Since these simple features are readily available for …

Content truncated to 500 words