Top-down cohesion segmentation in summarization

Abstract

The paper proposes a new method of linear text segmentation based on lexical cohesion of a text. Namely, first a single chain of disambiguated words in a text is established, then the rips of this single chain are considered as boundaries for the segments of the cohesion text structure (Cohesion TextTiling or CTT). The summaries of arbitrarily length are obtained by extraction using three different methods applied to the obtained segments. The informativeness of the obtained summaries is compared with the informativeness of the pair summaries of the same length obtained using an earlier method of logical segmentation by text entailment (Logical TextTiling or LTT). Some experiments about CTT and LTT methods are carried out for four "classical" texts in summarization literature showing that the quality of the summarization using cohesion segmentation (CTT) is better than the quality using logical segmentation (LTT).