1 Third Year Report

The aim of my PhD research is focused on Text Mining, one major research school in Knowledge Discovery in Databases (KDD), and in particular language-independent Documentbase Pre-processing (DPP) for classification / categorisation of documents, noted as Text Classification (TC), using novel algorithms for the identification of hidden patterns, rules, regularities and/or trends within these documents. Significant techniques in Data Mining, the classical research school in KDD that parallels to Text Mining, are involved to support this research, especially when dealing with a very large documentbase, such as Classification Rule Mining (CRM), Association Rule Mining (ARM), etc. One possible way to understand the framework of TC is to split it as DPP plus CRM. When applying Classification Association Rule Mining (CARM), a well established intersection between CRM and ARM, in TC, (1) the large volume of textual data (i.e., a given documentbase usually consists of more than 10,000 documents, where each document contains hundreds of words) can be handled, and (2) it is possible to deal with a small number of noisy data (i.e., a number of misspelling and/or cross-language words in documents). Based on (2), this approach can be identified as language-insensitive or language-independent in some angles, while acknowledging that such alternative approaches exist. With regards to the CARM based approach, developing a TC oriented Language-independent DPP (TC-LI-DPP) approach will consequently result an “almighty” TC approach, which have a general applicability regardless of the language(s) in which the documentbase to be classified are presented. In this report (a report that details what have been done within the third year of my PhD research), a number of language-independent DPP techniques, to support single-label Nclass TC, are described and compared. The discussion focuses on the vector space / “bag-of-*” model while acknowledging that alternative approaches to languageindependent DPP exist. A simple but effective statistical key / significant word identification approach is proposed which in turn is coupled with a number of phrase identification mechanisms. The emphasis in all cases is on language-independence so that the techniques described have a general applicability regardless of the language(s) in which the documentbase to be mined are presented. PDF created with pdfFactory Pro trial version www.software-partners.co.uk