Classification of Speech Acts in Tutorial Dialog

Computer models of tutorial dialog need to segment and classify the learner’s contributions into sequences of speech acts. The responses of the computer tutor we have developed (called AutoTutor) need to be sensitive to the speech acts of the learner’s contribution during the previous turn. We developed and tested some models that classified the speech acts in naturalistic tutorial dialog into four categories: Assertions, Yes/No Questions, WH-Questions, and Frozen Expressions. The three models consist of (1) a Brill part of speech tagger, a syntactic parser, and a symbolic postprocessor, (2) a feed-forward neural network, and (3) a model based on surface linguistic features. The parsing model had the best performance; classifying speech acts with 79% accuracy. Based on the data from this study, further research will be directed toward the usefulness of a hybrid model that uses a parser to extract the main clause, which can then be classified using a neural network.