How to get people to say and type what computers can understand

Abstract

This study tested whether people can be shaped to use the vocabulary and phrase structure of a program's output in creating their own inputs. Occasional computer-users interacted with four versions of an inventory program ostensibly capable of understanding natural-language inputs. The four versions differed in the vocabulary and the phrase length presented on the subjects' computer screen. Within each version, the program's outputs were worded consistently and presented repetitively in the hope that subjects would use the outputs as a model for their inputs. Although not told so in advance, one-half of the subjects were restricted to input phrases identical to those used by their respective program (shaping condition), the other half were not (modeling condition). Additionally, one-half of the subjects communicated with the program by speaking, the other half by typing. The analysis of the verbal dependent variables revealed four noteworthy findings. First, users will model the length of a program's output. Second, it is easier for people to model and to be shaped to terse, as opposed to conversational, output phrases. Third, shaping users' inputs through error messages is more successful in limiting the variability in their language than is relying on them to model the program's outputs. Fourth, mode of communication and output vocabulary do not affect the degree to which modeling or shaping occur in person-computer interactions. Comparisons of pre- and post-experimental attitudes show that both restricted and unrestricted subjects felt significantly more positive toward computers after their interactions with the natural-language system. Other performance and attitude differences as well as implications for the development of natural-language processors are discussed.