Proactive Spoken Dialogue Interaction in Multi-Party Environments

As it was predicted already in 1968 by Stanley Kubrick (1928-1999) and Arthur C. Clark (1917–2008) in the science ction movie 2001 – A Space Odyssey [Kubrick, 1968] the future has arrived. Computers are by now playing a prominent role in our everyday lives. Over the past decades they have evolved from big, monstrous mainly industrial machines to small mobile and extremely powerful devices that are in one way or another used by presumably every human being in the developed world. The quote by the ‘supercomputer’ HAL 9000 from Kubrick's movie shows that the computer is equipped with human-like qualities. It possesses natural language capabilities for both, understanding and speaking, the ability of logical reasoning and proactive behaviour, just to name a few character traits. The human characters of the movie are quoted in the movie to describe the computer as a sixth member of their space ship crew.

Dialogue involving two persons has for a long time been a popular topic of research. Dialogue systems involving two participants (the system and the user) have by now been _rmly established in everyday life, especially in the _eld of call centre applications. Multi-party dialogue, i.e. conversation between more than two participants, is on the other side a rather novel _eld in research. It started most likely with the reclassi_cation of the at that time conventional conversational roles of speaker and hearer (e.g. [Searle, 1969, Austin, 1962]). Clark and Carlson (1982) modi_ed Searle's speech act theory (1969) to enable multi-party interaction which was the start to multi-faceted research in multiparty dialogue and thus also opening the way for multi-party dialogue systems which have just started to emerge within the last few years.

Adequate dialogue data is needed to investigate the multi-party interaction with a computer system as an independent dialogue partner in the conversation with several humans. A look is taken at existing multi-party corpora in order to see if any of these could be used to investigate our research questions. The presented collection is limited to corpora that make use of multiple modalities. Although a wide variety of corpora is available, as far as the authors are aware, there is no existent collection of data that stresses the designated features important for our research. Thus, we perform the data collection presented in this chapter. We deploy the Wizard-of-Oz technique as introduced in Section 2.1 to simulate the envisaged system in the example domain of restaurant selection (refer to Section 1.3 for a description of the system) to obtain realistic interaction data which is then used to assist the development of the system (Chapter 4) and for evaluation (Chapter 5). The WOZ recording setup and procedure are presented in Section 3.2 followed by a detailed description of the software tool we have developed to support the recordings. The tool is easily adaptable to other domains and requirements and publicly available for other developers. The collected data results in the PIT corpus which is finally presented in Section 3.3.

In this section, the evaluation of our dialogue system is presented. Established methods for evaluating spoken language dialogue systems differentiate between subjective and objective methods as described in Section 2.2. The aim of the evaluation of our dialogue system is to appraise the user acceptance and rating of this novel sort of interactive system. Thus, the main focus is put on subjective evaluation for which data was obtained through the questionnaires filled out by the participants prior and subsequent to the data recordings. Usability evaluation is performed using two established methods (AttrakDiff [Hassenzahl et al., 2003] and a modified version of SASSI [Hone and Graham, 2000]). Evaluation is performed over the different recording sessions to detect the improvement of the system as well as comparing the different setups with and without avatar using the data of the Session III dialogues. A technical self-assessment of the participants was further conducted in order to validate the comparison of the di_erent recording sessions. The results of the evaluation are presented in Section 5.1.

The present book features a novel sort of spoken language dialogue system that acts as an independent dialgoue partner in the interaction with two users. As the field of application of dialogue systems has started to rapidly change and expand, established dialogue systems are exposed to new challenges some of which have been taken up in the scope of this work. The book started with a short general introduction on spoken language dialogue systems, followed by a description of current trends and related work in dialogue systems research. Chapter 1 was concluded with a description of the said dialogue system. Chapter 2 provided a basis for the topics covered in the subsequent chapters. Thus, fundamentals on relevant subjects of corpus development, (multi-party) human-human and human-computer interaction, and evaluation were presented. A detailed description on dialogue management focuses on the information state update approach, further taking a look at recent efforts on multi-party dialogue management. Chapter 3 listed a selection of up-to-date multi-party dialogue corpora. As none of those corpora exhibits the characteristics we aim at investigating in order to obtain interaction models suitable for the development of our dialogue management, the PIT corpus of multi-party dialogues was recorded in an extensive Wizard-of-Oz environment as presented in the remainder of the chapter. Chapter 4 presented the multi-party dialogue management that enables proactive interaction in the conversation. Finally, the novel sort of dialogue system was evaluated in terms of usability and user acceptance as presented in Chapter 5. The contributions of this work are described in the following in detail.