Recognizing Biomedical Named Entities Using SVMs: Improving Recognition Performance with a Minimal Set of Features

Vote-Based Classifier Selection for Biomedical NER Using Genetic Algorithms

“…Four forwardparsed classifiers (e 5 , e 6 , e 9 , and e 33 ) and five backward-parsed classifiers (e 4 , e 7 , e 32 , e 34 , and e 35 ) are selected. This behavior agrees with the discussion in [4] that training SVMs with different parse directions produce systems that make errors at different boundaries. The forward-parsed classifiers are selected even though their full object F-scores are lower than many backward-parsed classifiers that are not included in the ensemble.…”

“…1, diversity can be achieved by using different model parameters and features in each ensemble member. Because of this, each classifier is trained using different settings of YamCha parameters such as dimensionality of the polynomial kernel, range of the context window and direction of parsing as used in [4]. Four different feature types which are frequently used for NER are considered in this study.…”

“…Orthographic feature includes token properties such as the existence of an upper case character or Greek letter in the token or a combination of such properties. Surface words are constructed from training data as described in [4]. The feature types mentioned include many different variations and are used in different combinations as well as in isolation to train the classifiers as illustrated in Table 1.…”

“…The wide variety of features considered for this purpose include orthographic features, morphological patterns, lexical features, semantic triggers, name alias features, gene sequences, and external resources such as gazetteers. Extensive research has been carried out for computing better feature sets so as to improve NER in this domain [3], [4]. This is generally done by trial-and-error approach where the main goal is to include discriminative features and avoid using correlated ones.…”

Classifier subset selection for biomedical named entity recognition

“…Furthermore, we studied application of various ensembling methods to a five-entity NER problem instead of the two-entity case studied by Zhou et al [9]. A novel surface word feature and two orthographic feature extraction techniques, all based on occurrence statistics of entity names that are originally proposed by the authors of this paper are also considered [11]. Experimental results conducted using JNLPBA Bio-Entity Recognition Task data [12] have shown that the proposed approach provides an F-score of 72.51%, improving the F-score achieved by the best individual classifier, the ensemble of all classifiers and three popular static classifier selection based ensembles, namely Forward Selection (FS), Backward Selection (BS) and Genetic Algorithms (GA) [13] by 2.5% and 1.3%, 0.9%, 0.9% and 0.8% respectively.…”