Arabic Mention Detection: Toward Better Unit of Analysis

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 709–712, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Arabic Mention Detection: Toward Better Unit of Analysis Yassine Benajiba Center for Computational Learning Systems Columbia University ybenajiba@ccls.columbia.edu Imed Zitouni IBM T. J. Watson Research Center izitouni@us.ibm.com Abstract We investigate in this paper the adequate unit of analysis for Arabic Mention Detection. We experiment different segmentation schemes with various feature-sets. Results show that when limited resources are available, models built on morphologically segmented data out- perform other models by up to 4F points. On the other hand, when more resources extracted from morphologically segmented data become available, models built with Arabic TreeBank style segmentation yield to better results. We also show additional improvement by combin- ing different segmentation schemes. 1 Introduction This paper addresses an important and basic task of information extraction: Mention Detection (MD)1: the identification and classification of textual refer- ences to objects/abstractions (i.e., mentions). These mentions can be either named (e.g. Mohammed, John), nominal (city, president) or pronominal (e.g. he, she). For instance, in the sentence “President Obama said he will visit ...” there are three men- tions: President, Obama and he. This is similar to the Named Entity Recognition (NER) task with the additional twist of also identifying nominal and pronominal mentions. We formulate the mention de- tection problem as a classification problem, by as- signing to each token in the text a label, indicating whether it starts a specific mention, is inside a spe- cific mention, or is outside all mentions. The se- lection of the unit of analysis is an important step toward a better classification. When processing lan- guages, such as English, using the word itself as the 1We adopt here the ACE nomenclature: http://www.nist.gov/speech/tests/ace/index.html unit of analysis (after separating punctuations) leads to a good performance (Florian et al., 2004). For other languages, such as Chinese, character is con- sidered as the adequate unit of analysis (Jing et al., 2003). In this paper, we investigate different seg- mentation schemes in order to define the best unit of analysis for Arabic MD. Arabic adopts a very com- plex morphology, i.e. each word is composed of zero or more prefixes, one stem and zero or more suffixes. Consequently, the Arabic data is sparser than other languages, such as English, and it is necessary to “segment” the words into several units of analysis in order to achieve a good performance. (Zitouni et al., 2005) used Arabic morphologically segmented data and claimed to have very competi- tive results in ACE 2003 and ACE 2004 data. On the other hand, (Benajiba et al., 2008) report good re- sults for Arabic NER on ACE 2003, 2004 and 2005 data using Arabic TreeBank (ATB) segmentation. In all published works, authors do not mention a spe- cific motivation for the segmentation scheme they have adopted. Only for the Machine Translation task, (Habash and Sadat, 2006) report several results using different Arabic segmentation schemes. They …

Content truncated to 500 words