Feature Extraction in Subject Classification of Text Documents in Polish

摘要

In this work we evaluate two different methods for deriving features for a subject classification of text documents. The first method uses the standard Bag-of-Words (BoW) approach, which represents the documents with vectors of frequencies of selected terms appearing in the documents. This method heavily relies on the natural language processing (NLP) tools to properly preprocess text in the grammar- and inflection-conscious way. The second approach is based on the word-embedding technique recently proposed by Mikolov and does not require any NLP preprocessing. In this method the words are represented as vectors in continuous space and this representation of words is used to construct the feature vectors of the documents. We evaluate these fundamentally different approaches in the task of classification of Polish language Wikipedia articles with 34 subject areas. Our study suggests that the word-embedding based features seem to outperform the standard NLP-based features providing sufficiently large training dataset is available.