Poorly formed input and miscommunication in natural-language keyboard dialogue: An exploratory study

摘要

Subjects (n = 21) completed a series of tasks involving data-base query using a simulated natural-language interface yielding a corpus of 476 user inputs. Ill-formedness and miscommunication were studied using a variety of taxonomies proposed in previous research. Misspelling ran at 14% (of user input), extra-grammaticality, largely consisting of ellipsis, at 40%, and misunderstanding, much of which remained undetected, at 20%. The study verified a number of expectations of the “naturalistic” approach to interface design, in particular the correlation of low-level ill-formedness with miscommunication at planning level, and the existence of strong temporal effects in the data according as the dialogue proceeded from beginning to end. The implications of the findings for interface design are discussed.