An automatic feedback system based on confidence deviations of prediction and detection models for english phrase break

Abstract

This paper presents a method to construct a feedback provision model for English phrase breaks by utilizing confidence deviations of prediction and detection models. The proposed method consists of prediction, detection and feedback provision models. The prediction and detection models adopted conditional random fields classifiers performed on the Boston University radio news corpus, and achieved accuracies of 90.15% and 91.62%, respectively. The feedback provision model determines three types of feedbacks for each disjunction using the differences between the prediction and detection confidences. In a validation experiment for the feedback provision, the proposed method demonstrated a Pearson’s correlation coefficient of 0.74 between the feedback provision model’s scores and human fluency assessments.