Evaluation of topic segmentation algorithms on Arabic texts

Abstract

In this paper, we are interested in the topic segmentation of Arabic texts. For this aim, we evaluate two based lexical cohesion algorithms: MinCutSeg and BayesSeg by using the Pk and WindowDiff metrics. To assess how well each algorithm works, each was applied on three datasets with longer texts from two different domains: transcribed multi-party conversations and written texts. After adaptation to the Arabic language, the test results show significant differences in performance depending on the types of documents.