Using Ontology-based Approaches to Representing Speech Transcripts for Automated Speech Scoring

This paper presents a thesis proposal on approaches to automatically scoring non-native speech from second language tests. Current speech scoring systems assess speech by primarily using acoustic features such as fluency and pronunciation; however content features are barely involved. Motivated by this limitation, the study aims to investigate the use of content features in speech scoring systems. For content features, a central question is how speech content can be represented in appropriate means to facilitate automated speech scoring. The study proposes using ontologybased representation to perform concept level representation on speech transcripts, and furthermore the content features computed from ontology-based representation may facilitate speech scoring. One baseline and two ontology-based representations are compared in experiments. Preliminary results show that ontology-based representation slightly improves performance of one content feature for automated scoring over the baseline system.