Improving Interaction Quality Recognition Using Error Correction

Determining the quality of an ongoing interaction in the field of Spoken Dialogue Systems is a hard task. While existing methods employing automatic estimation already achieve reasonable results, still there is a lot of room for improvement. Hence, we aim at tackling the task by estimating the error of the applied statistical classification algorithms in a two-stage approach. Correcting the hypotheses using the estimated model error increases performance by up to 4.1 % relative improvement in Unweighted Average Recall.