A NEW DATA-DRIVEN APPROACH FOR MULTIMEDIA PRESENTATION PLANNING Amanda Stent Department of Computer Science State University of New York at Stony Brook email: stent@cs.sunysb.edu Hui Guo Department of Computer Science State University of New York at Stony Brook email: huguo@cs.sunysb.edu ABSTRACT A number of Multimedia Presentation Systems have been built for technical documentation, traffic management sys- tems, educational software and other applications. How- ever, these systems use handwritten presentation planning operators, which limits their flexibility. In this paper we describe a new data-driven approach for multimedia pre- sentation planning. We present algorithms for acquiring presentation plan operators from Web pages, and for using these operators to generate multimedia presentations. We also show how to use machine learning to adapt multimedia presentation planning to human presentation preferences. KEY WORDS Human Computer Interfaces, Multimedia Communication Systems, Collaborative Systems and Applications 1 Introduction Multimedia has proven to be an efficient and effective form of presenting information from complex data sources. In our information-rich world, multimedia presentations are all around us: on our computers, PDAs and cellphones, in our books, magazines and newspapers, and in our work- places, schools and homes. As a result, the need for Mul- timedia Presentation Systems (MMPSs), which generate multimedia presentations automatically, is growing rapidly. Multimedia presentation planning is an inherently complex process. This complexity has two basic causes: the size of the choice space for multimedia presentations, and the need for good heuristics for pruning the choice space and finding optimal paths through it [3]. Consider first the size of the choice space for mul- timedia presentations. The basic elements are pieces of content and presentation elements (e.g. tables, pictures, charts, text, speech). The planning process itself involves four tasks. Relevant content must be selected and orga- nized in accordance with some communicative goal (con- tent selection). These facts must be mapped to the presen- tation elements that will be used to present them (presenta- tion element selection). These presentation elements must be arranged temporally and spatially (presentation layout), and global and local presentation features (e.g. color, size, grouping) must be set (presentation style). To get an idea of the complexity of these tasks, consider only the task of assigning content to presentation elements. If each piece of content must be assigned to exactly one presentation el- ement, and there are c pieces of content and p presenta- tion elements, then the number of assignments of content to presentation elements is p c . If we permit redundancy (i.e. each piece of content can be assigned to one or more presentation elements) then the number of assignments be- comes (2 p - 1) * c. Many of these assignments, though theoretically possible, are infeasible or have poor presenta- tion style; however, it is difficult to automatically define a subset of feasible assignments for a particular domain and goal that is large enough to permit interesting variation. Now consider the problem of searching through this choice space for ’good’ presentations. Most people do not focus on the presentation aspects of a good presentation, because it permits them to focus on the content and on their current task. However, they can identify faults of a bad pre- sentation; they may say that it doesn’t contain the informa- tion they need or contains too much information, is unclear, is badly organized, or is unattractive. These terms are help- ful for human presentation designers, but are too vague for MMPSs, and very few people (usually graphic designers or statisticians) can explain their intuitions about presentation design clearly (c.f. [8, 9]). A presentation planning system should therefore contain methods for transforming users’ presentation preferences into quantitative metrics that elim- inate poor presentation choices and/or select good ones. This paper describes a new data-driven approach to multimedia presentation planning. Section 3 describes our algorithm for acquiring presentation operators automati- cally. Section 4 describes how we generate multimedia presentations. In Section5, we show how we can use super- vised learning to adapt multimedia presentation planning for human preferences. We sum up in Section 6. 2 Related Work Multimedia generation is usually performed in a pipeline process: content selection, presentation planning, tactical or media-specific realization, and production [1]. Exist- ing multimedia generation systems usually assign content to media during presentation planning. There are two ways of representing presentation pat- terns for multimedia presentation planners. Schemas are templates for an entire presentation [14], while presenta-