Topic Segmentation: Application of Mathematical Morphology to Textual Data

Abstract

Mathematical Morphology (MM) offers a generic theoretical framework for data processing and analysis. Nevertheless, it remains essentially used in the context of image analysis and processing, and the attempts to use MM on other kinds of data are still quite rare. We believe MM can provide relevant solutions for data analysis and processing in a far broader range of application fields. To illustrate, we focus here on textual data and we show how morphological operators (here the morphological segmentation using watershed transform) may be applied on these data. We thus provide an original MM-based solution to the thematic segmentation problem, which is a typical problem in the fields of natural language processing and information retrieval (IR).

More precisely, we consider here TV broadcasts through their transcription obtained by automatic speech recognition. To perform topic segmentation, we compute the similarity between successive segments using a technique called vectorization which has recently been introduced in the IR field. We then apply a gradient operator to build a topographic surface to be segmented using the watershed transform. This new topic segmentation technique is evaluated on two corpora of TV broadcasts on which it outperforms other existing approaches. Despite using very common morphological operators (i.e., the standard Watershed Transform), we thus show the potential interest of MM to be applied on non-image data.