Arabic Named Entity Recognition using Optimized Feature Sets

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 284–293, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Arabic Named Entity Recognition using Optimized Feature Sets Yassine Benajiba• Mona Diab⋄ Paolo Rosso• •Natural Language Engineering Lab., Dept. de Sistemas Inform´aticos y Computaci´on Universidad Polit´ecnica de Valencia {ybenajiba,prosso}@dsic.upv.es ⋄Center of Computational Learning Systems Columbia University mdiab@cs.columbia.edu Abstract The Named Entity Recognition (NER) task has been garnering significant attention in NLP as it helps improve the performance of many natural language processing applica- tions. In this paper, we investigate the im- pact of using different sets of features in two discriminative machine learning frameworks, namely, Support Vector Machines and Condi- tional Random Fields using Arabic data. We explore lexical, contextual and morphological features on eight standardized data-sets of dif- ferent genres. We measure the impact of the different features in isolation, rank them ac- cording to their impact for each named entity class and incrementally combine them in or- der to infer the optimal machine learning ap- proach and feature set. Our system yields a performance of Fβ=1-measure=83.5 on ACE 2003 Broadcast News data. 1 Introduction Named Entity Recognition (NER) is the process by which named entities are identified and classified in an open-domain text. NER is one of the most im- portant sub-tasks in Information Extraction. Thanks to standard evaluation test beds such as the Auto- matic Content Extraction (ACE)1, the task of NER has garnered significant attention within the natu- ral language processing (NLP) community. ACE has facilitated evaluation for different languages cre- ating standardized test sets and evaluation metrics. NER systems are typically enabling sub-tasks within 1http://www.nist.gov/speech/tests/ace/2004/doc/ace04- evalplan-v7.pdf large NLP systems. The quality of the NER sys- tem has a direct impact on the quality of the overall NLP system. Evidence abound in the literature in areas such as Question Answering, Machine Trans- lation, and Information Retrieval (Babych and Hart- ley, 2003; Ferr´andez et al., 2004; Toda and Kataoka, 2005). The most prominent NER systems approach the problem as a classification task: identifying the named entities (NE) in the text and then classify- ing them according to a set of designed features into one of a predefined set of classes (Bender et al., 2003). The number of classes differ depending on the data set. To our knowledge, to date, the ap- proach is always to model the problem with a sin- gle set of features for all the classes simultaneously. This research, diverges from this view. We recog- nize that different classes are sensitive to differing features. Hence, in this study, we aspire to discover the optimum feature set per NE class. We approach the NER task from a multi-classification perspec- tive. We create a classifier for each NE class inde- pendently based on an optimal feature set, then com- bine the different classifiers for a global NER sys- tem. For creating the different classifiers per class, we adopt two discriminative approaches: Support Vector Machines (SVM)(Vapnik, 1995), and Condi- tional Random Fields (CRF)(Lafferty et al., …

Content truncated to 500 words