Text categorization on Reuters corpus

1 Task The task of text categorization can be described as follows: given a set of documents, we want to assign to each document one or more text categories or no category. In this term project, we want categorize documents from the well-known Reuters-21578 corpus which is a collection of 21578 articles published on Reuters in 1987. We have chosen only three most frequent text categories as the target categories: Mergers/Acquisitions (ACQ) Earnings and Earnings Forecasts (EARN) Money/Foreign Exchange (MONEY-FX) Since this categories may overlap, we have at total 8 target labels for each subset of these categories. We want to train a classifier that will assign a target label to a given document. In the text categorization task, we want to compare two different approaches. Both approaches are based on building an ensemble of binary classifiers. In the first approach, we will use a simple majority voting method to assign a target label. In the second approach, we will use the outputs of this ensample as the input for another classifier. Our ensemble of classifiers will consist of 8 classifiers: for each text category (ACQ, EARN, MONEY-FX) we will train 3 binary classifiers with different machine learning methods:  Support Vector Machine (SVM)  Random Forest (RF)  Naïve Bayes (NB) The output of each of these binary classifiers will be TRUE or FALSE – assigning or not assigning a given label. In the majority voting approach, if two of classifiers will output TRUE, we will assign the particular text category to an input document. The target label will be determined as the composition of these assigned text categories. In the second approach, we will train a decision tree that will use these eight binary outputs. The decision tree will directly assign the target label. In the text categorization task, the first step is to transform text documents into a set of features suitable for classifier learning methods. Convenient transformation could be representing the document as a set of words, ignoring their order in the text. Thus we will split the text documents into the words and then each feature will correspond to occurrence of a particular term in the text.