Finding Appropriate Subset of Votes Per Classifier Using Multiobjective Optimization: Application to Named Entity Recognition

Finding Appropriate Subset of Votes Per Classifier Using Multiobjective Optimization: Application to Named Entity Recognition Asif Ekbal 1∗, Sriparna Saha 1∗ and Md. Hasanuzzaman2 1 Heidelberg University, 69120 Heidelberg, Germany Email:asif.ekbal@gmail.com, sriparna.saha@gmail.com 2 West Bengal Industrial Development Corporation, Kolkata, India Email: hasanuzzaman.im@gmail.com ∗First two authors are the joint first authors. Abstract. In this paper, we report a multiobjective optimization (MOO) based technique to select the appropriate subset of votes per classifier in an ensemble system. We hypothesize that the reliability of prediction of each classifier differs among the various output classes. Thus, it is necessary to find out the subset of classes for which any particular classifier is most suitable. Rather than optimizing a single measure of classification quality, we simultaneously optimize two different measures of classification quality using the search capability of MOO. We use our proposed technique to solve the problem of Named Entity Recognition (NER). Maximum Entropy (ME) model is used as a base to build a number of classifiers depending upon the various representations of the contextual, orthographic word-level and semantically motivated features. Evaluation results with a resource constrained language like Bengali yield the recall, precision and F-measure values of 87.98%, 93.00%, and 90.42%, respectively. Experimental results suggest that the use of semantic feature can significantly improve the overall system performance. Results also reveal that the classifier ensemble identified by the proposed MOO based approach performs better in comparison to the individual classifiers, two different baseline ensembles and the classifier ensemble identified by a single objective genetic algorithm (GA) based approach. 1 Introduction Named Entity Recognition (NER) is an important pipelined module in many Natural Language Processing (NLP) application areas that include machine translation, information retrieval, in- formation extraction, question-answering, automatic summarization etc. Machine learning ap- proaches are popularly being used for NER due to their flexible adaptation to new domains and languages. Most of the existing works in NER cover the languages such as English, European lan- guages and some of the Asian languages like Chinese, Japanese and Korean. India is a multilingual country with great linguistic and cultural diversities. In India, there are 22 official languages that are inherited from almost all the existing linguistic families in the world. However, the works related to NER in Indian languages have started to emerge only very recently. Named Entity (NE) identification in Indian languages in general and Bengali in particular is more difficult and chal- lenging compared to others due to facts such as: (i). missing of capitalization information, (ii). appearance of NEs in the dictionary with some other specific meanings, (iii). free word order nature of the languages, (iv). resource-constrained environment, i.e. non-availability of corpora, annotated corpora, name dictionaries, good morphological analyzers, part of speech (POS) taggers etc. Some of the recent works related to Bengali NER can be found in (Ekbal and Bandyopad- hyay, 2009b; Ekbal and Bandyopadhyay, 2009a; Ekbal and Bandyopadhyay, 2008b). Other works related to Indian language NER are reported in the proceedings of the IJCNLP-08 Workshop on NER for South and …

Content truncated to 500 words