Digestor: Device-Independent Access to the World-Wide Web

Timothy W. Bickmore and Bill N. Schilit

Main points:

• depricates Unwired Planet manual reauthoring, multiple-device authoring (such as style sheets), client-side navigation (i.e. select portion of larger page to view at any given time, like our segmentation, or magic lenses).
• operate on syntatic vs. semantic (structure vs. content) document, use elision vs. transformation (summarization vs. decimation)
• they manually reauthored some pages to get a feel for what's important
  • need to keep some images
  • headers (H1-H6) often used incorrectly, used for style not structure
  • "semantic elision" by removing frame sidebars works well
  • pages can be heuristically categorized, category dictates strategy
  • whitespace reduction works really well
• Digestor: bag of tricks plus categorization to pick appropriate trick based on particular page and also the target device
• tricks:
  • outlining, first sentence extraction, image reduction by const. factor, font size change
  • "heuristic planner" to search possible transformation space, screen area as measure of merit of a heurisitic (!!!)
  • sensible [pre|post|]conditions on heuristics for aesthetics
• implementation
  • uses proxy, builts syntax tree from HTML, retrieves inlines, then launches heuristic planner to iteratively search transf. space
  • implemented in Java