Theme Extraction in Bulgarian: Experiments in Supervised and Unsupervised Settings

In this work we address the task of automatic keyword and keyphrase extraction from unstructured text, and suit it to the need of a popular Bulgarian media for induction of ’themes’. Themes are defined as text snippets that summa-rize the essence of an article, facilitating the itemization and retrieval of collections of textual resources and document clustering. We evaluate the performance of several generic methods for keyword and keyphrase extraction on a corpus of articles in Bulgarian, as to the best of our knowledge no such study has been conducted in the past. The meth-ods that we discuss rely on widely accepted information re-trieval and machine learning techniques and are language-independent. We also consider the effect of a stemmer com-ponent on the keyphrase extraction accuracy. The satisfacto-ry performance of our models in spite of the limited linguistic knowledge incorporated in them recommends our models a