Abstract

This Ph.D. thesis describes the investigations we carried out in order to determine the appropriate approach to build an efficient and robust Arabic Named Entity Recognition system. Such a system would have the ability to identify and classify the Named Entities within an open-domain Arabic text.
The Named Entity Recognition (NER) task helps other Natural Language Processing approaches (e.g. Information Retrieval, Question Answering, Machine Translation, etc.) achieve a higher performance thanks to the significant information added to the text. In the literature, many research works report the adequate approaches which can be used to build an NER system for a specific language or from a language independent perspective. Yet, very few research works which investigate the task for the Arabic language have been published.
The Arabic language has a special orthography and a complex morphology which bring new challenges to the NER task to be investigated. A complete investigation of Arabic NER would report the technique which helps achieve a high performance, as well as giving a detailed error analysis and results discussion so as to make the study beneficial to the research community. This thesis work aims at satisfying this specific need. In order to achieve that goal we have:
1. Studied the different aspects of the Arabic language which are related to the
NER task;
2. Studied the state-of-art of the NER task;
3. Conducted a comparative study among the most successful Machine Learning approaches on the NER task;
4. Carried out a multi-classifier approach where each classifier deals with only one NE class and uses the appropriate Machine Learning approach and feature-set for the concerned class.

We have evaluated our experiments on different nine data-sets of different genres (newswire, broadcast news, Arabic Treebank and weblogs). Our findings point out that the multi-classifier yields the best results.