LADDER : a perceptually-based language to simplify sketch recognition user interface development

Abstract

Diagrammatic sketching is a natural modality of human-computer interaction that can be used for a variety of tasks, for example, conceptual design. Sketch recognition systems are currently being developed for many domains. However, they require signal-processing expertise if they are to handle the intricacies of each domain, and they are time-consuming to build. Our goal is to enable user interface designers and domain experts who may not have expertise in sketch recognition to be able to build these sketch systems. We created and implemented a new framework (FLUID - f acilitating user interface development) in which developers can specify a domain description indicating how domain shapes are to be recognized, displayed, and edited. This description is then automatically transformed into a sketch recognition user interface for that domain. LADDER, a language using a perceptual vocabulary based on Gestalt principles, was developed to describe how to recognize, display, and edit domain shapes. A translator and a customizable recognition system (GUILD - a generator of user interfaces using ladder descriptions) are combined with a domain description to automatically create a domain specific recognition system.

(cont.) With this new technology, by writing a domain description, developers are able to create a new sketch interface for a domain, greatly reducing the time and expertise for the task Continuing in pursuit of our goal to facilitate UI development, we noted that 1) human generated descriptions contained syntactic and conceptual errors, and that 2) it is more natural for a user to specify a shape by drawing it than by editing text. However, computer generated descriptions from a single drawn example are also flawed, as one cannot express all allowable variations in a single example. In response, we created a modification of the traditional model of active learning in which the system selectively generates its own near-miss examples and uses the human teacher as a source of labels. System generated near-misses offer a number of advantages. Human generated examples are tedious to create and may not expose problems in the current concept. It seems most effective for the near-miss examples to be generated by whichever learning participant (teacher or student) knows better where the deficiencies lie; this will allow the concepts to be more quickly and effectively refined.

(cont.) When working in a closed domain such as this one, the computer learner knows exactly which conceptual uncertainties remain, and which hypotheses need to be tested and confirmed. The system uses these labeled examples to automatically build a LADDER shape description, using a modification of the version spaces algorithm that handles interrelated constraints, and which also has the ability to learn negative and disjunctive constraints.