Automatically Developing a Fine-grained Arabic Named Entity Corpus and Gazetteer by utilizing Wikipedia

International Joint Conference on Natural Language Processing, pages 392–400, Nagoya, Japan, 14-18 October 2013. Automatically Developing a Fine-grained Arabic Named Entity Corpus and Gazetteer by utilizing Wikipedia Fahd Alotaibi School of Computer Science University of Birmingham, UK fsa081@cs.bham.ac.uk Mark Lee School of Computer Science University of Birmingham, UK m.g.lee@cs.bham.ac.uk Abstract This paper presents a methodology to ex- ploit the potential of Arabic Wikipedia to as- sist in the automatic development of a large Fine-grained Named Entity (NE) corpus and gazetteer. The corner stone of this approach is efficient classification of Wikipedia articles to target NE classes. The resources developed were thoroughly evaluated to ensure reliability and a high quality. Results show the developed gazetteer boosts the performance of the NE classifier on a news-wire domain by at least 2 points F-measure. Moreover, by combining a learning NE classifier with the developed cor- pus the score achieved is a high F-measure of 85.18%. The developed resources overcome the limitations of traditional Arabic NE tasks by more fine-grained analysis and providing a beneficial route for further studies. 1 Introduction Previous efforts that have been made to develop an Ara- bic NER either focused on traditional NE classes (Be- najiba et al., 2010) or sought to expand only one class at a time (Shaalan and Raza, 2007). Applications such as Question Answering (QA) receive more benefits when a fine-grained NER is developed. This is true when we consider that, the majority of factoid questions are about named entities (Noguera et al., 2005). Having a finer NER, results in the possibility of extracting more semantic knowledge from the context. For example, if we consider the following sentence: (ﻞﺋﺎﺳﻭﺕﺎﻛﺮﺷﺮﺒﻛﺃﻲﻫﻲﻧﺰﻳدﺖﻟﺍﻭﺔﻛﺮﺷ ﻢﻟﺎﻌﻟﺍﻲﻓﻪﻴﻓﺮﺘﻟﺍﻭﻡﻼﻋﻹﺍ/ˇsrkℏ wAlt dyzny hy Okbr ˇsrkAt wsAˆyl AlIςlAm wAltrfyh fy AlςAlm/ ‘Walt Dis- ney is the largest media company in the entertainment world’)1 We would have more semantic information if we could tag (ﻲﻧﺰﻳدﺖﻟﺍﻭ/wAlt dyzny/ ‘Walt Disney’) 1Throughout this paper, Arabic words are represented in three variants: (Arabic word /HSB transliteration scheme (Habash et al., 2007) / “English translation”) as [ORG-ENTERTAINMENT] rather than just [ORG]. This deeper semantics is very helpful when answering factoid question like “What is the largest entertainment company?” Supervised machine learning technologies have been successfully adopted for several natural language tasks, including NER. These technologies require a reason- able portion of data to be accessible in the training phase, containing a number of positive and negative examples to learn from and to circumvent the problem of data sparseness. Traditional methods for compiling such data involve recruiting individuals to annotate a certain corpus manually. This is tedious work, as well as costly and time consuming. Moreover, manually an- notating a large portion of a relatively open domain cor- pus beyond a news-wire and across various genres is not easy for an individual to achieve. Therefore, when developing a reasonable fine- grained NE corpus two questions should be answered. First, what proper fine-grained semantic classes should be established? Second, how to develop a reasonable sized fine-grained NE corpus at minimum cost? This work answers …

Content truncated to 500 words