This paper presents a novel lexical-similarity-based approach to automatic story segmentation in broadcast news. When measuring the connection between a pair of sentences, we take two factors into consideration, i.e. the lexical similarity and the distance between them in the text stream. Further investigation of pairwise connections between sentences is based on the technique of Laplacian Eigenmaps (LE). Talcing advantage of the LE algorithm, we construct a Euclidean space in which each sentence is mapped to a vector. The original connective strength between sentences is reflected by the Euclidean distances between the corresponding vectors in the target space of the map. Further analysis of the map leads to a straightforward criterion for optimal segmentation. Then we formalize story segmentation as a minimization problem and give a dynamic programming solution to it. Experimental results on the TDT2 corpus show that the proposed method outperforms several state-of-the-art lexical-similarity-based methods.
展开▼
