Subsequence kernels-based Arabic text classification

Abstract

Kernel methods have known huge success in machine learning. This success is mainly due to their flexibility to deal with high dimensionality of the feature space of complex data such as graphs, trees or textual data. In the field of text classification (TC) their performances have supplanted traditional algorithms. For textual data, different kernels were introduced (P-spectrum, All-Sub-sequences, Gap-Weighted Subsequences kernel, ...) to improve the performance of TC systems. In this paper, we carried out a system for Arabic TC which supports aspects of order and co-occurrence of words within a text. Transducers, specific automata, are used to represent documents. Such representation allows an efficient implementation of subsequence kernel. An empirical study is conducted to evaluate the ATC system on the large SPA corpus. Results show an improvement of the classification in terms of precision.