Automatic detection of plagiarized spoken responses

This paper addresses the task of automatically detecting plagiarized responses in the context of a test of spoken English proficiency for non-native speakers. A corpus of spoken responses containing plagiarized content was collected from a high-stakes assessment of English proficiency for non-native speakers, and several text-to-text similarity metrics were implemented to compare these responses to a set of materials that were identified as likely sources for the plagiarized content. Finally, a classifier was trained using these similarity metrics to predict whether a given spoken response is plagiarized or not. The classifier was evaluated on a data set containing the responses with plagiarized content and non-plagiarized control responses and achieved accuracies of 92.0% using transcriptions and 87.1% using ASR output (with a baseline accuracy of 50.0%).