A Semi-supervised Learning Approach to Arabic Named Entity Recognition

Proceedings of Recent Advances in Natural Language Processing, pages 32–40, Hissar, Bulgaria, 7-13 September 2013. A Semi-supervised Learning Approach to Arabic Named Entity Recognition Maha Althobaiti, Udo Kruschwitz, and Massimo Poesio School of Computer Science and Electronic Engineering University of Essex Colchester, UK {mjaltha, udo, poesio}@essex.ac.uk Abstract We present ASemiNER, a semi- supervised algorithm for identifying Named Entities (NEs) in Arabic text. ASemiNER does not require annotated training data, or gazetteers. It also can be easily adapted to handle more than the three standard NE types (Person, Location, and Organisation). To our knowledge, our algorithm is the first study that intensively investigates the semi-supervised pattern-based learning approach to Arabic Named Entity Recog- nition (NER). We describe ASemiNER and compare its performance with dif- ferent supervised systems. We evaluate this algorithm by way of experiments to extract the three standard named-entity types. Ultimately, our algorithm out- performs simple supervised systems and also performs well when we evaluate its performance in order to extract three new, specialised types of NEs (Politicians, Sportspersons, and Artists). 1 Introduction Named Entities (NEs) are textual references via proper names, such as first and last names, loca- tions, and companies. Detecting NEs within un- structured text and classifying them into prede- fined categories of names is known as Named En- tity Recognition (NER) (Grishman and Sundheim, 1996). Arabic NER has been given great amount of at- tention over the past fifteen years. A number of Arabic NER systems have been developed using three approaches, which have been investigated thoroughly in the literature of NER. These ap- proaches are rule-based (Shaalan and Raza, 2007; Shaalan and Raza, 2009), Machine Learning (ML) (Benajiba et al., 2007; Benajiba and Rosso, 2007; Benajiba and Rosso, 2008; Abdul-Hamid and Dar- wish, 2010) and hybrid (Abdallah et al., 2012; Oudah and Shaalan, 2012). Over the past decade, some studies have ex- plored the possibility of solving the problem of NER with a reduced level of supervision. These studies proposed semi-supervised and unsuper- vised systems, which no longer require annotated datasets and can be easily adapted to new types (Nadeau et al., 2006; Etzioni et al., 2005; Liao and Veeramachaneni, 2009; Liu et al., 2011). This paper introduces ASemiNER, an Arabic semi-supervised NER system built under minimal supervision. Gazetteers (predefined lists of NEs) and annotated corpora are not required by ASem- iNER. That is, ASemiNER is a bootstrapping al- gorithm that takes a few examples of a particular NE type as input and iteratively induces and learns patterns, which are used to extract more examples. Extraction patterns are induced and generalised automatically from data using very general criteria that require no human intervention, and no prior knowledge of the language or the corpus domain. In addition to the fact that ASemiNER extracts and recognises the three standard NEs (Person, Loca- tion, and Organisation names), it has proven to be an adaptable system that can be easily modified to extract new NEs without the need for analysing the dataset or collecting and tagging new large …

Content truncated to 500 words