Unsupervised Methods of Topical Text Segmentation for Polish

This paper describes a study on performance of existing unsupervised algorithms of text documents topical segmentation when applied to Polish plain text documents. For performance measurement five existing topical segmentation algorithms were selected, three different Polish test collections were created and seven approaches to text preprocessing were implemented. Based on quantitative results (Pk and WindowDiff metrics) use of specific algorithm was recommended and impact of pre-processing strategies was assessed. Thanks to use of standardized metrics and application of previously described methodology for test collection development, comparative results for Polish and English were also obtained.