How accurate is Converser™ really? What is the percentage of successful translations?
Measurements are complicated by Converser’s multiple options for input (speech, typing, pointing, and handwriting), and by the wide range of healthcare situations – from intake interviews to rehabilitation.Further, methodology for evaluation of speech translation, and of machine translation more generally, is a matter of ongoing debate in the research community: there can be degrees of “correctness,” and criteria can vary from “all vocabulary items and structures are perfect” to “the utterance is understandable”.Given this variability, it’s impossible to provide a scientific measurement of Converser’s overall success rate.However, it’s crucial that the system can be monitored and corrected at two levels – input and translation. Thus users can raise the accuracy as much as necessary for the sentence at hand by interactively repairing errors. The time spent on corrections can depend flexibly on the importance of the current input.
Measurements are complicated by Converser’s multiple options for input (speech, typing, pointing, and handwriting), and by the wide range of healthcare situations – from intake interviews to rehabilitation. Further, methodology for evaluation of speech translation, and of machine translation more generally, is a matter of ongoing debate in the research community: there can be degrees of “correctness,” and criteria can vary from “all vocabulary items and structures are perfect” to “the utterance is understandable”. Given this variability, it’s impossible to provide a scientific measurement of Converser’s overall success rate. However, it’s crucial that the system can be monitored and corrected at two levels – input and translation. Thus users can raise the accuracy as much as necessary for the sentence at hand by interactively repairing errors. The time spent on corrections can depend flexibly on the importance of the current input. At the input level, all text is subject to automatic s
