Conversational agents and negative lessons from behaviourism

Abstract

© 2019, Springer Nature Switzerland AG. This chapter addresses the question of whether it is enough to extract from data the knowledge needed to implement socially believable conversational agents. Contrary to the popular views, the answer is negative. In this respect, the chapter points to some shortcomings of fully data-driven approaches to dialogue management, including the lack of external criteria for the selection of dialogue corpora, and the misconception of dialogue structure and dialogue context. To point to these shortcomings is not to undervalue data-driven approaches, but to emphasize the message that big data provide only a partial account of human-machine dialogue, and thus must not remain wedded to small linguistic theory, as it is currently the case.