The effects of restricted vocabulary size on voice interactive discourse structure

Advances in the design of Natural Language Understanding (NLU) interfaces and speech recognition technology have provided the basis for developing domain-specific, voice interactive Natural Language Dialogue Systems. Current speech recognition constraints limit the vocabulary size that these systems employ. To date, the effects of specific vocabulary restrictions on task performance and dialogue structure with systems utilizing speech input have not been fully explored. This study attempted to ameliorate this deficiency. To examine the effects of restricted vocabularies on task performance and discourse structure, a three phase study was employed. Phase 0 (pilot) was conducted to obtain unrestricted baseline dialogues elicited from a verbal interactive repair task. Phase 1 was designed to assess the initial effects of using restricted vocabularies (300 words, 200 words, and 125 words) derived from the baseline data. Phase 2 was completed to evaluate the effects of practice on the task and discourse structure. Users of all three restricted vocabularies in Phase 1 successfully completed their tasks. Discourse results indicated significant differences between the unrestricted and restricted vocabulary groups, but not among the restricted groups themselves. These included: longer task completion time, fewer words used, shorter average utterance length, fewer pronouns, and a higher percentage of simple, goal directed utterances for the restricted groups. Expected similarities were noted when comparing the results of Phase 2 with Phase 1. Phase 2 also revealed comparable practice effects between the restricted (125 word) and unrestricted vocabulary groups. These included: decreased task completion time, decreased total word usage, fewer clarification subdialogues, and an increase in goal directed utterances. Results from this study, when compared to previous research, suggest that vocabulary restrictions produce similar discourse and task effects for both speech and keyboard input. Additionally, the observed practice effects indicate that as task familiarity increases, subjects' utterances become more precise and goal directed regardless of vocabulary restrictions. These results provide support for proponents of NLU interfaces for intermittent users of moderately structured tasks.