Short Utterance Dialogue Act Classification Using a Transformer Ensemble

An influx of digital assistant adoption and reliance is demonstrating the significance of reliable and robust dialogue act classification techniques. In the literature, there is an over-representation of purely lexical-based dialogue act classification methods. A weakness of this approach is the lack of context when classifying short utterances. We improve upon a purely lexical approach by incorporating a state-of-the-art acoustic model in a lexical-acoustic transformer ensemble, with improved results, when classifying dialogue acts in the MRDA corpus. Additionally, we further investigate the performance on an utterance word-count basis, showing classification accuracy increases with utterance word count. Furthermore, the performance of the lexical model increases with utterance word length and the acoustic model performance decreases with utterance word count, showing the models complement each other for different utterance lengths.