A method for comparing fluency measures and its application to ITS natural language question generation

The motivation for this empirical study is to make it easier to achieve relatively more fluent dialogue. I sketch a framework for automatically assessing sentence fluency and introduce some statistical techniques that may make corpus characterization and comparisons easier. The technical approach is based on interpreting the language used in a particular situation as a sublanguage of English. Corpus-based statistical models of the English language and its putative sublanguage are constructed and two composite measures of sentence fluency are compared via four non-parametric statistical tests. With one exception, the quantitative results agree with qualitative expectations. Several additional issues are considered and directions for future work are presented.