Machine Learning Approaches to Improving

This thesis reports the investigations into the task of phone-level pronunciation error detection, the performance of which is heavily affected by the imbalanced distribution of the classes in a manually annotated data set of non-native English (Read Aloud responses from the TOEFL Junior Pilot assessment). In order to address problems caused by this extreme class imbalance, two machine learning approaches, cost-sensitive learning and oversampling, are explored to improve the classification performance. Specifically, approaches which assigned weights inversely proportional to class frequencies and synthetic minority over-sampling technique (SMOTE) were applied to a range of classifiers using feature sets that included information about the acoustic signal, the linguistic properties of the utterance, and word identity. Empirical experiments demonstrate that both balancing approaches lead to a substantial performance improvement (in terms of f1 score) over the baseline on this extremely imbalanced data set. In addition, this thesis also discusses which features are the most important and which classifiers are most effective for the task of identifying phone-level pronunciation errors in non-native speech.