5. Design Recommendations

6. Related Work

Though interested in different tasks, other researchers have studied using intelligent user interfaces to speed information capture. For instance, Hermens and Schlimmer (1994) built an electronic form filler that tried to provide default values for every field on the form. Each field of the form had a decision tree to calculate a default value. Like Names++, the calculations used previously entered information to generate defaults and predictively fill in fields. Unlike Names++, the calculations themselves were constructed at run-time using a machine learning method. (Names++ does not alter predictive fillin at run-time. cf. Table 1.) They field tested their system with a single electronic form filled out several hundred times over an eight month period. They report an 87% reduction in keystrokes; loosely translating this into a speedup yields 669% speedup or approximately 3 times the 210% speedup we observed for entering a name.

Studying text prediction without field boundaries, Pomerleau (1995) built a typing completion aid. Without relying on note-taking properties, his system predicts a completion for the current word being typed (presumably in an editor). A connectionist network estimates the probability of a number of possible completions for the current word; the most likely, over some threshold, is offered to the user. Pomerleau tested his system with a pair of subjects over a two-week period and reports an increase in typing speed of 2% for English text and 13-18% for computer program code. This modest gain may be due to inefficiencies in the learning method, to lack of redundancy in the task, or to limitations in the user interface itself.

As a complement to earlier research, this paper reports the individual and collective accuracy of three user interface components. It reports user task time showing that the components significantly improve efficiency. This paper also clarifies an issue confounded in earlier work. If a learning interface is less effective than expected, is this due to an inherent limitation in the interface itself, or does its learning method perform inadequately? To answer the second question, other work compares two or more learning methods. In this paper, we hand-built the predictive fillin structures (cf. Table 1) and were able to assess the quality of the predictive fillin interface directly.

7. Conclusions


Jeffrey C. Schlimmer, schlimme@eecs.wsu.edu, 5 December 1994