Social talk capabilities for dialogue systems

Small talk capabilities are an important but very challenging extension to dialogue systems. Small talk (or “social talk”) refers to a kind of conversation, which does not focus on the exchange of information, but on the negotiation of social roles and situations. Small talk is often not as limited in topics and content as so-called “task talk”, meaning conversations regarding a specific task of a dialogue system such as, e.g., providing bus schedule information. Several studies have shown that human users tend to initialize social talk in conversations with dialogue systems, especially if the dialogue system is embedded in an application that includes aspects of human personality such as embodied avatars. Moreover, studies have shown that social conversations can effectively establish an “emotional” connection between the user and the machine and create a pleasant atmosphere which is appreciated by most users. However, only few existing dialogue systems offer small talk support and nearly none systematic analysis of small talk usable for computational purposes has been proposed so far. The goal of this thesis is to provide knowledge, processes and structures that can be used by dialogue systems to satisfactorily participate in social conversations. For this purpose the thesis primarily presents, besides research in the fields of natural-language understanding and dialogue management, research on dialogue models and error handling. Regarding dialogue models, a new structured model of social talk based on a data analysis of small talk conversations is described. The functionally-motivated and content-abstract model can be used for small talk conversations on various topics. The model is based on a novel, theory-based set of social dialogue acts and is also available as computational model learned from conversation data. 4 SOCIAL-TALK CAPABILITIES FOR DIALOGUE SYSTEMS Since it cannot be guaranteed that all contents for social conversations initialized by the users of a dialogue system have been modeled, this thesis also suggests new conversation strategies for the treatment of so-called “out-of-domain” (OoD) utterances. OoD utterances are utterances which do not fall within one of the knowledge domains of the system and thus lead to errors in the input interpretation. These errors cannot be handled using the typical error strategies such as a repair, because the knowledge necessary to understand these utterances is missing. The new strategies are based on information from human-human communication extracted from various sources. The presented research is technologically encapsulated in a software toolkit. The toolkit provides software extensions to dialogue systems that enable social talk. For evaluation the tools are integrated into a conversational agent application: a barkeeper agent in a virtual world. Two evaluations, an overall usability evaluation of the agent and an evaluation of the two main tools, indicate a clear improvement in the users’ perception of the agent when the tools are activated, especially in the areas of naturalness, natural-language understanding and conversation flow. The fun the users had while using the application seems to be strongly related to the system’s social talk abilities.