An automatic normalized cut topic segmentation approach

Abstract

This paper presents an automatic topic segmentation approach based on subwords normalized cut (Ncut) for Chinese broadcast news, since the classical Ncut has a limitation that the number of segments has to be set as a prior. We abstract a text into a weighted undirected graph, where the nodes correspond to sentences and the weights of edges describe inter-sentence lexical similarities at Chinese subwords level, thus the segmentation task is formalized as a graph-partitioning problem under the Ncut criterion. In order to break through the limitation, we proposed a text dotplotting inspired method, which can evaluate the segmentation results and select the optimal number of segments automatically. Lastly, we put the whole approach into a machine learning framework, learning the best arguments on train set. Our method achieved relative improvement of 3% over non-automatic subwords Ncut, also the previous best method.