Generating a Table-of-Contents

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 544–551, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Generating a Table-of-Contents S.R.K. Branavan, Pawan Deshpande and Regina Barzilay Massachusetts Institute of Technology {branavan, pawand, regina}@csail.mit.edu Abstract This paper presents a method for the auto- matic generation of a table-of-contents. This type of summary could serve as an effec- tive navigation tool for accessing informa- tion in long texts, such as books. To gen- erate a coherent table-of-contents, we need to capture both global dependencies across different titles in the table and local con- straints within sections. Our algorithm ef- fectively handles these complex dependen- cies by factoring the model into local and global components, and incrementally con- structing the model’s output. The results of automatic evaluation and manual assessment confirm the benefits of this design: our sys- tem is consistently ranked higher than non- hierarchical baselines. 1 Introduction Current research in summarization focuses on pro- cessing short articles, primarily in the news domain. While in practice the existing summarization meth- ods are not limited to this material, they are not universal: texts in many domains and genres can- not be summarized using these techniques. A par- ticularly significant challenge is the summarization of longer texts, such as books. The requirement for high compression rates and the increased need for the preservation of contextual dependencies be- tween summary sentences places summarization of such texts beyond the scope of current methods. In this paper, we investigate the automatic gener- ation of tables-of-contents, a type of indicative sum- mary particularly suited for accessing information in long texts. A typical table-of-contents lists topics described in the source text and provides informa- tion about their location in the text. The hierarchical organization of information in the table further re- fines information access by specifying the relations between different topics and providing rich contex- tual information during browsing. Commonly found in books, tables-of-contents can also facilitate access to other types of texts. For instance, this type of summary could serve as an effective navigation tool for understanding a long, unstructured transcript for an academic lecture or a meeting. Given a text, our goal is to generate a tree wherein a node represents a segment of text and a title that summarizes its content. This process involves two tasks: the hierarchical segmentation of the text, and the generation of informative titles for each segment. The first task can be addressed by using the hier- archical structure readily available in the text (e.g., chapters, sections and subsections) or by employ- ing existing topic segmentation algorithms (Hearst, 1994). In this paper, we take the former approach. As for the second task, a naive approach would be to employ existing methods of title generation to each segment, and combine the results into a tree struc- ture. However, the latter approach cannot guarantee that the generated table-of-contents forms a coher- ent representation of the entire text. Since titles of different segments are generated in isolation, …

Content truncated to 500 words