Instance Sampling for Identification of Arabic Pleonastic Pronouns M. Abdul-Mageed 1 Instance Sampling for Automatic Identification of Arabic Pleonastic Pronouns

The term anaphora describes backward reference to items previously occurring in a text (see e.g., Mitkov, 2002). The pointing back item is called an anaphor and the item to which it refers is called its antecedent. The identification of an anaphor’s antecedent is termed anaphora resolution and is considered one of the most difficult tasks in natural language processing (NLP) since it relies on both linguistic and world knowledge. Resolving an anaphor to its antecedent is crucial in many NLP applications such as text summarization, machine translation, information extraction and question answering systems. I employ the term pleonastic pronouns to refer to cases of pronouns that do not introduce a new referent in discourse. In the literature, the terms non-referential, nonanaphoric, and expletive have been used to label these types of pronouns. Ability to identify non-referential pronouns before attempting an anaphora resolution task is significant since the system would not have to attempt resolving such pronouns and hence would end up with fewer errors. In addition, as (Boyd, Gegg-Harrison, and Byron, 2005) maintain in their study of non-referential it in English, this task of detecting non-referential pronouns could be incorporated into a part-of-speech tagger or parser, or treated as an initial step in semantic interpretation. The number of non-referential pronouns is sometimes non-trivial and hence it is significant to devote enough attention to its identification. In this paper I apply a memory-based learning (Daelemans and van den Bosch, 2005) method for identifying non-referential pronouns in an annotated sub-segment of the Penn Arabic Treebank, v3. I follow (Evans, 2001) in defining the task as a binary classification and investigate what instance sampling methods give the best results. Finally, I investigate what features are most important for the classification task. I acquire 89% of successful classification. Instance Sampling for Identification of Arabic Pleonastic Pronouns M. Abdul-Mageed 2