Named Entity Recognition through Classifier Combination

Named Entity Recognition through Classifier Combination Radu Florian and Abe Ittycheriah and Hongyan Jing and Tong Zhang IBM T.J. Watson Research Center 1101 Kitchawan Rd, Yorktown Heights, NY 10598, USA {raduf,abei,hjing,tzhang}@us.ibm.com Abstract This paper presents a classifier-combination experimental framework for named entity recognition in which four diverse classi- fiers (robust linear classifier, maximum en- tropy, transformation-based learning, and hid- den Markov model) are combined under differ- ent conditions. When no gazetteer or other ad- ditional training resources are used, the com- bined system attains a performance of 91.6F on the English development data; integrat- ing name, location and person gazetteers, and named entity systems trained on additional, more general, data reduces the F-measure error by a factor of 15 to 21% on the English data. 1 Introduction This paper investigates the combination of a set of di- verse statistical named entity classifiers, including a rule-based classifier – the transformation-based learning classifier (Brill, 1995; Florian and Ngai, 2001, hence- forth fnTBL) with the forward-backward extension de- scribed in Florian (2002a), a hidden Markov model clas- sifier (henceforth HMM), similar to the one described in Bikel et al. (1999), a robust risk minimization classi- fier, based on a regularized winnow method (Zhang et al., 2002) (henceforth RRM) and a maximum entropy clas- sifier (Darroch and Ratcliff, 1972; Berger et al., 1996; Borthwick, 1999) (henceforth MaxEnt). This particular set of classifiers is diverse across multiple dimensions, making it suitable for combination: • fnTBL is a discriminant classifier – it bases its clas- sification decision only on the few most discriminant features active on an example – while HMM, RRM and MaxEnt are agglomerative classifiers – their de- cision is based on the combination of all features ac- tive for the particular example. • In dealing with the data sparseness problem, fnTBL, MaxEnt and RRM investigate and integrate in their decision arbitrary feature types, while HMM is de- pendent on a prespecified back-off path. • The search methods employed by each classifier are different: the HMM, MaxEnt and RRM classifiers construct a model for each example and then rely on a sequence search such as the Viterbi algorithm (Viterbi, 1967) to identify the best overall sequence, while fnTBL starts with most frequent classification (usually per token), and then dynamically models the interaction between classifications, effectively performing the search at training time. • The classifiers also differ in their output: fnTBL and RRM return a single classification per exam- ple1, while the MaxEnt and HMM classifiers return a probability distribution. The remainder of the paper is organized as follows: Sec- tion 2 describes the features used by the classifiers, Sec- tion 3 briefly describes the algorithms used by each clas- sifier, and Section 4 analyzes in detail the results obtained by each classifier and their combination. 2 The Classification Method and Features Used All algorithms described in this paper identify the named entities in the text by labeling each word with a tag corresponding to its position relative to a named entity: whether it starts/continues/ends …

Content truncated to 500 words