A Hybrid Approach to Features Representation for Fine-grained Arabic Named Entity Recognition

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 984–995, Dublin, Ireland, August 23-29 2014. A Hybrid Approach to Features Representation for Fine-grained Arabic Named Entity Recognition Fahd Alotaibi School of Computer Science University of Birmingham, UK fsa081@cs.bham.ac.uk Faculty of Computing King Abdulaziz University, KSA fsalotaibi@kau.edu.sa Mark Lee School of Computer Science University of Birmingham, UK m.g.lee@cs.bham.ac.uk Abstract Despite considerable research on the topic of Arabic Named Entity Recognition (NER), almost all efforts focus on a traditional set of semantic classes, features and token representations. In this work, we advance previous research in a systematic manner and devise a novel method to represent these features, relying on a dependency-based structure to capture further evidence within the sentence. Moreover, the work also describes an evaluation of the method involving the capture of global features and employing the clustering of unannotated textual data. To meet this set of goals, we conducted a series of evaluations to evaluate different aspects that demonstrate great improvement when compared with the baseline model. 1 Introduction Traditionally, the focus of Arabic NER has been on a very limited number of semantic classes, i.e. PERSON, ORGANISATION and LOCATION, utilising the newswire domain such as those described by Benajiba and Rosso (2008), Benajiba et al. (2010) and Abdul-Hamid and Darwish (2010) . This limits higher-level applications (such as question answering) from extracting in-depth knowledge and working on a relatively open domains. This paper addresses the issue of a fine-grained NER of 50 classes for Arabic and presents a com- prehensive set of experiments that evaluate innovative means of representing the features set. Thus, the contribution of this paper falls into different categories with unique outcomes, as follows: 1. A novel approach to representing the features is used, relying on dependency representation. This representation overcomes the drawback of current window-based representations of features. 2. The representation of global evidence involves clustering unannotated textual data, employing hier- atical clusters (Brown et al., 1992). 3. Due to the fact that there is no comparable work to use as a comparison in the task of Arabic fine- grained NER, a baseline model was developed, based on Conditional Random Fields (CRF), using the best features, as established and used elsewhere in the literature. 4. Development of publically available gold-standard fine-grained NER corpora1 from two different genres, i.e. Newswire and Wikipedia. Each contribution is discussed in more detail during in the remainder of this paper. 2 Arabic Fine-grained Named Entity Corpora The majority of Arabic NER approaches are supervised, ensuring that the machine learns from an an- notated corpus and aims to predict unseen text. This approach requires a reasonable bank of labelled data. This section examines the availability of such an annotated dataset at the fine-grained level, and the creation of gold-standard corpora. This work is licensed under a Creative Commons Attribution 4.0 International Licence. 1Available at: http://sourceforge.net/projects/arabic-named-entity-corpora/ and Mirror at: http://fsalotaibi.kau.edu.sa/Pages-Arabic-NE-Corpora.aspx 984 2.1 Available Corpora One of the earliest corpora publically released was ANERcorp, developed by Benajiba et …

Content truncated to 500 words