Benchmarking Bert-based models performances for multi-lingual Dialog Acts (DA) classification

Michel Daher Mansour, Clement Torres

Abstract: Dialog Acts (DA) classification plays an important role in chatbots and spoken dialogue systems. Such tasks are specific in two ways: First, they need to be robust to language switching within or between conversations. Second, each utterance must be understood within the context of the current dialog. In this work, we build on the rapid emergence of Deep Learning techniques applied to Natural language Processing (NLP) and the availability of pre-trained models, and propose to benchmark a series of Transformer based models on both French and English spoken (transcribed) data, with different settings to take into account the context of the dialog. Our experiments show that a baseline XLM-RobERTA can provide fairly robust performances across these two languages even if slighlty lower than mono-lingual model like CamemBERT or BERT on their own language. Also, a simple concatenation up to the last three utterances does not perform well to capture the conversation context.