A System for Understanding and Reformulating Tables

Tables are an important means for communicating information in written media, and understanding such tables is a challenging problem in document layout analysis. Possible applications include extracting information for populating databases which can later be manipulated or queried, and reformulating existing tables so that they can be presented in a medium different from their original target (e.g., on a much smaller screen, or via a spoken language interface). For example, in a system that reads email over the phone, it would be inconvenient to be forced to listen to the entire body of a large table being read sequentially. On the other hand, many tables are formatted in a way that makes it quite natural to consider querying them via simple relational-database-type commands. Clearly, being able to recognize and exploit such structure in a document would be extremely useful. In this paper, we describe a prototype for a complete, end-to-end system that takes raw ASCII text as input, spots and parses any number of tables that it might contain, and generates a simple interactive man-machine dialog allowing a user to access the table data via a spoken language interface. This type of application framework is both compelling and tremendously challenging, as it spans the fields of document analysis, user interface design and development, text-to-speech synthesis, speech recognition, and dialog systems. An overview of our system is shown in Figure 1. While many of the components are still the subject of active research, preliminary indications are promising.