Efficient arabic text classification using feature selection techniques and genetic algorithm

The significant growth of the number of Internet users in recent years has been accompanied by a very large increase in the number of data being shared at every moment, specifically text data. This large increase in the number of text data makes it necessary for researchers to try to develop text classification (TC) techniques to be able to deal with this huge amount of data and classify it into pre-defined categories. Arabic is one of the languages that has not received enough TC studies, as dealing with Arabic texts is much more difficult than dealing with texts written in other languages such as English even for Arabic speakers because of its highly derivative nature, which makes it a great challenge for researchers. This thesis aims to establish an effective and integrated Arabic TC system. Web scraping techniques will be used to create a new Arabic dataset that includes many versions. Different pre-processing techniques will be applied to each version before evaluating all versions and selecting the best version to ensure that high quality data is used with the classification system. An improved approach based on the combination of common feature selection (FS) techniques, filter FS method and wrapper FS method, in one approach will be used to make the most of the advantages of the two methods with four classifiers, 21 term weighting (TW) schemes, and genetic algorithm (GA) to select the best features from the selected data version. Filter FS method techniques use the best features found using TW schemes to classify text using classifiers. When all the classifiers and TW schemes have been evaluated, the best classifier and the best 5 TW schemes are selected, which will help to identify a carefully selected set of high-quality features to use with the GA to find the best subset of features. In wrapper FS, the GA performs operations similar to real biological operations, such as crossover and mutations, on the best features selected using best TW schemes many times, where the features are changed in each generation to choose the best features and reduce the lowest quality features to get to the best subset of features. The results show that the proposed approach, hybrid FS- GA approach, can outperform the best results obtained in the filter FS method, a common FS method, using a subset of features with a size of less than 10% of the total feature size and it can also select up to 1% of the original feature size with great efficiency. It is worth noting that these good results are not only related to the Arabic, as the proposed classification system has proven very effective with other languages such as English and Turkish. This thesis presents a new Arabic dataset of high quality and makes it available, in all its versions, to all researchers and those interested in the Arabic TC, and it contributes to provide detailed information about the performance of each TC and TW scheme, and runtime needed to obtain the best results to aid in future research aimed at developing Arabic TC techniques.