Investigating Automatic Alignment Methods for Slide Generation from Academic Papers

Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL), pages 111–119, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Investigating Automatic Alignment Methods for Slide Generation from Academic Papers Brandon Beamer and Roxana Girju Department of Linguistics University of Illinois Urbana, IL {bbeamer,girju}@illinos.edu Abstract In this paper we investigate the task of auto- matic generation of slide presentations from academic papers, focusing initially on slide to paper alignment. We compare and eval- uate four different alignment systems which utilize various combinations of methods used widely in other alignment and question an- swering approaches, such as TF-IDF term weighting and query expansion. Our best aligner achieves an accuracy of 75% and our findings show that for this application, av- erage TF-IDF scoring performs more poorly than a simpler method based on the number of matched terms, and query expansion degrades aligner performance. 1 Introduction Automatic generation of slide presentations is a task the Computational Linguistics community has not yet pursued in much depth. A robust system capable of generating slide presentations from papers would save the author much tedium when organizing her presentations. In this paper we investigate this task from a novel perspective. While others have devel- oped interesting approaches to slide generation from documents by modeling the problem in a unique way (Utiyama and Hasida, 1999; Shibata and Kuro- hashi, 2005), the aim of the research this paper initi- ates is to discover how humans create slide presen- tations, focusing more specifically on academic pa- pers. Thus we take a corpus-based approach to the problem, and as a first step focus on the task of au- tomatically aligning slide presentations to academic papers. We built a corpus of 296 slide-paper pairs and im- plemented four slide to paper aligners which utilize popular information retrieval methods such as TF- IDF term weighting and query expansion. In this paper we show that, in this application, TF-IDF term weighting is inferior to a simpler scoring mecha- nism based only on the number of matched terms and query expansion degrades aligner performance. Our best aligner achieves an accuracy of 75%. 2 Related Work Automatic slide generation from documents is a thus far under-investigated topic. Utiyama and Hasida (1999) generate slides from GDA1 (global document annotation) tagged documents. They detect topics within the documents by analyzing GDA corefer- ence links, modeled each slide as a topic and item- ized elaborations (which were also tagged with the GDA tag set). Shibata and Kurohashi (2005) convert Japanese documents to slide representation by pars- ing their discourse structures and representing the resulting tree in an outline format. While (Utiyama and Hasida, 1999) and (Shibata and Kurohashi, 2005) generate slides from documents by modeling the task in creative ways, we aim to learn something deeper regarding how humans actually go about the task. Creating a corpus of slide/paper pairs will en- able us to study the intricacies involved in how real humans approach this task. Our current focus is slide to paper (region) align- ment, …

Content truncated to 500 words