Abstract

We propose integrating features from lexical cohesion with elements from layout recognition to build a composite framework. We use supervised machine learning on this composite feature set to derive discourse structure on the topic level. We demonstrate a system based on this principle and use both an intrinsic evaluation as well as the task of genre classication to assess its performance. 2 Introduction A document structure tree 1 can be dened as a data structure that allows navigation of a document by sections. These trees can be hierarchically organized, having subsections of sections and may embed special items, such as gures, tables or hyperlinks. They may be used directly by an end user for document access, or indirectly through other applications. This paper describes a strategy to compute document structure using a framework that deals both with rich, semi-structured documents with layout features as well as impoverished, text stream-like documents. Our system, the Comb...

Citations