In this paper we examine topic segmentationof narrative documents, which arecharacterized by long passages of textwith few headings. We first present resultssuggesting that previous topic segmentationapproaches are not appropriate fornarrative text. We then present a featurebasedmethod that combines features fromdiverse sources as well as learned features.Applied to narrative books and encyclopediaarticles, our method shows results thatare significantly better than previous segmentationapproaches. An analysis of individualfeatures is also provided and thebenefit of generalization using outside resourcesis shown.
展开▼