Understanding Dialogue for Human Communication

Dialogue is a peculiar activity of humans and a crucial characteristic of human cognition. It is not surprising that dialogue have been investigated, under different perspectives, by linguists, cognitive scientists, and philosophers and, in the last decades, by computer scientists. This chapter shows the progress achieved in computational linguistics to design formal models of dialogues and exploit them in human-machine systems. We highlight that collaborative dialogues follow sequences of turns characterized by speech acts and that they show an internal coherence based on conversational goals. Analysis carried on dialogue collections reveals the importance of modeling mixed-initiative schema, various types of subdialogues, and grounding among interlocutors, as they help to achieve the speakers’ communicative goals. On the computational side, both knowledge-driven and machine learning technologies are nowadays used to model a pipeline of dialogue components, particularly for task-oriented situations, including automatic speech recognition, utterance understanding, dialogue state tracking, dialogue policy making, and response generation. In recent years, research on dialogue systems has moved toward the so-called conversational AI, which takes advantage of the power of neural architectures to induce models from annotated dialogues. Neural models have achieved state-of-the-art performance, and end-to-end solutions are now proposed in place of traditional dialogue pipelines. However, we argue that current models are applied to relatively narrow tasks and still scratch the surface of capturing human collaborative dialogues’ effectiveness and cognitive abilities.