Analysis and implementation of methods for the text categorization

Text Categorization (TC) is the automatic classification of text documents under pre-defined categories, or classes. Popular TC approaches map categories into symbolic labels and use a training set of documents, previously labeled by human experts, to build a classifier which enables the automatic TC of unlabeled documents. Suitable TC methods come from the field of data mining and information retrieval, however the following issues remain unsolved. First, the classifier performance depends heavily on hand-labeled documents that are the only source of knowledge for learning the classifier. Being a labor-intensive and time consuming activity, the manual attribution of documents to categories is extremely costly. This creates a serious limitations when a set of manual labeled data is not available, as it happens in most cases. Second, even a moderately sized text collection often has tens of thousands of terms in that making the classification cost prohibitive for learning algorithms that do not scale well to large problem sizes. Most important, TC should be based on the text content rather than on a set of hand-labeled documents whose categorization depends on the subjective judgment of a human classifier. This thesis aims at facing the above issues by proposing innovative approaches which leverage techniques from data mining and information retrieval. To face problems about both the high dimensionality of the text collection and the large number of terms in a single text, the thesis proposes a hybrid model for term selection which combines and takes advantage of both filter and wrapper approaches. In detail, the proposed model uses a filter to rank the list of terms present in documents to ensure that useful terms are unlikely to be screened out. Next, to limit classification problems due to the correlation among terms, this ranked list is refined by a wrapper that uses a Genetic Algorithm (GA) to retaining the most informative and discriminative terms. Experimental results compare well with some of the top-performing learning algorithms for TC and seems to confirm the effectiveness of the proposed model. To face the issues about the lack and the subjectivity of manually labeled datasets, the basic idea is to use an ontology-based approach which does not depend on the existence of a training set and relies solely on a set of concepts within a given domain and the relationships between concepts. In this regard, the thesis proposes a text categorization approach that applies WordNet for selecting the correct sense of words in a document, and utilizes domain names in WordNet Domains for classification purposes. Experiments show that the proposed approach performs well in classifying a large corpus of documents. This thesis contributes to the area of data mining and information retrieval. Specifically, it introduces and evaluates novel techniques to the field of text categorization. The primary objective of this thesis is to test the hypothesis that: text categorization requires and benefits from techniques designed to exploit document content. hybrid methods from data mining and information retrieval can better support problems about high dimensionality that is the main aspect of large document collections. in absence of manually annotated documents, WordNet domain abstraction can be used that is both useful and general enough to categorize any documents collection. As a final remark, it is important to acknowledge that much of the inspiration and motivation for this work derived from the vision of the future of text categorization processes which are related to specific application domains such as the business area and the industrial sectors, just to cite a few. In the end, it is this vision that provided the guiding framework. However, it is equally important to understand that many of the results and techniques developed in this thesis are not limited to text categorization. For example, the evaluation of disambiguation methods is interesting in its own right and is likely to be relevant to other application fields.