Improving Arabic Light Stemming in Information Retrieval Systems

Abstract

Information retrieval refers to the retrieval of textual documents such as newsprint and magazine articles or Web documents. Due to extensive research in the IR field, there are many retrieval techniques that have been developed for Arabic language. The main objective of this research to improve Arabic information retrieval by enhancing light stemming and preprocessing stage and to contribute to the open source community, also establish a guideline for Arabic normalization and stop-word removal. To achieve these objectives, we create a GUI toolkit that implements preprocessing stage that is necessary for information retrieval. One of these steps is normalizing, which we improved and introduced a set of rules to be standardized and improved by other researchers. The next preprocessing step we improved is stop-word removal, we introduced two different stop-word lists, the first one is intensive stop-word list for reducing the size of the index and ambiguous words, and the other is light stop-word list for better results with recall in information retrieval applications. We improved light stemming by update a suffix rule, and introduce the use of Arabized words, 100 words manually collected, these words should not follow the stemming rules since they came to Arabic language from other languages, and show how this improve results compared to two popular stemming algorithms like Khoja and Larkey stemmers. The proposed toolkit was integrated into a popular IR platform known as Terrier IR platform. We implemented Arabic language support into the Terrier IR platform. We used TF-IDF scoring model from Terrier IR platform. We tested our results using OSAC datasets. We used java programming language and Terrier IR platform for the proposed systems. The infrastructure we used consisted of CORE I7 CPU ran speed at 3.4 GHZ and 8 GB RAM.