A Stacked, Voted, Stacked Model for Named Entity Recognition

A Stacked, Voted, Stacked Model for Named Entity Recognition Dekai Wu†∗and Grace Ngai‡ and Marine Carpuat† † Human Language Technology Center HKUST Department of Computer Science University of Science and Technology Clear Water Bay, Hong Kong {dekai,marine}@cs.ust.hk ‡ Hong Kong Polytechnic University Department of Computing Kowloon Hong Kong csgngai@polyu.edu.hk Abstract This paper investigates stacking and voting methods for combining strong classifiers like boosting, SVM, and TBL, on the named-entity recognition task. We demonstrate several ef- fective approaches, culminating in a model that achieves error rate reductions on the develop- ment and test sets of 63.6% and 55.0% (En- glish) and 47.0% and 51.7% (German) over the CoNLL-2003 standard baseline respectively, and 19.7% over a strong AdaBoost baseline model from CoNLL-2002. 1 Introduction We describe multiple stacking and voting methods that effectively combine strong classifiers such as boosting, SVM, and TBL, for the named-entity recognition (NER) task. NER has emerged as an important step for many natural language applications, including machine trans- lation, information retrieval and information extraction. Much of the research in this field was pioneered in the Message Understanding Conference (MUC) (Sund- heim, 1995), which performed detailed entity extraction and identification on English documents. As a result, most current NER systems which have impressive per- formances have been specially constructed and tuned for English MUC-style documents. It is unclear how well they would perform when applied to another language. Our system was designed for the CoNLL-2003 shared task, the goal of which is to identify and classify four types of named entities: PERSON, LOCATION, ORGA- NIZATION and MISCELLANEOUS. The task specifi- cations were that two languages would be involved. We *The author would like to thank the Hong Kong Re- search Grants Council (RGC) for supporting this research in part through two research grants (RGC 6083/99E and RGC 6256/00E). were given about a month to develop our system on the first language, which was English, but only two weeks to adapt it to the surprise language, which was German. Given the goal of the shared task, we designed our system to achieve a high performance without relying too heavily on knowledge that is very specific for a par- ticular language or domain. In the spirit of language- independence, we avoided using features and information which would not be easily obtainable for almost any ma- jor language. 2 Classification Methods To carry out the stacking and voting experiments, we con- structed a number of relatively strong individual compo- nent models of the following kinds. 2.1 Boosting The main idea behind boosting algorithms is that a set of many weak classifiers can be effectively combined to yield a single strong classifier. Each weak classifier is trained sequentially, increasingly focusing more heavily on the instances that the previous classifiers found diffi- cult to classify. For the boosting framework, our system uses Ada- Boost.MH (Freund and Schapire, 1997), an n-ary classifi- cation variant of the original binary AdaBoost algorithm. It performs well on a number of natural language process- ing problems, including text …

Content truncated to 500 words