Abstract
Abstract
In listening assessment, a domain definition inference of temporal fluency features remains under-researched, posing a challenge for test developers in creating authentic listening texts. To address this issue, this paper reports two studies. Study I examined interrelations of temporal fluency features (e.g., articulation rate) in lecture settings. Lecture passages in TOEFL iBT and IELTS were compared to real-world data (Fluency Corpus of Academic English Lectures [FCAEL]). Results showed both similarities (e.g., positive association between articulation rate and end-clause silent pause frequency), and dissimilarities (e.g., correlations between articulation rate and mid-clause filled pause frequency and between repetition frequency and mid-clause filled pause frequency) among the three. Since correlations do not necessarily indicate that disfluencies co-occur, Study II investigated co-occurrences of two disfluency features that indicated high correlations in Study I (i.e., repetitions and filled pauses) in FCAEL and TOEFL. IELTS was excluded because of zero observations of repetitions. Results from generalized-linear mixed-effects models showed proportional similarities between the two, while paired-sample t-tests found significantly lower frequency of both co-occurrences and non-co-occurrences in TOEFL compared to FCAEL. The study proposes thresholds of co-occurrences of repetitions and filled-pauses in academic English lecture texts. The study also argues that variability should be considered construct-relevant, instead of noise.