Generation and evaluation of user tailored responses in multimodal dialogue

Abstract

When people engage in conversation, they tailor their utterances to their conversational partners, whether these partners are other humans or computational systems. This tailoring, or adaptation to the partner takes place in all facets of human language use, and is based on a mental model or a user model of the conversational partner. Such adaptation has been shown to improve listeners' comprehension, their satisfaction with an interactive system, the efficiency with which they execute conversational tasks, and the likelihood of achieving higher level goals such as changing the listener's beliefs and attitudes. We focus on one aspect of adaptation, namely the tailoring of the content of dialogue system utterances for the higher level processes of persuasion, argumentation and advice-giving. Our hypothesis is that algorithms that adapt content for these processes, according to a user model, will improve the usability, efficiency, and effectiveness of dialogue systems. We describe a multimodal dialogue system and algorithms for adaptive content selection based on multi-attribute decision theory. We demonstrate experimentally the improved efficacy of system responses through the use of user models to both tailor the content of system utterances and to manipulate their conciseness.

User preferences can drive facial expressions: evaluating an embodied conversational agent in a recommender dialogue system

“…In a similar study evaluating the output of the MATCH restaurant recommendation system (Walker et al 2004), participants preferred recommendations that were tailored to their own user preferences over those that were tailored to the preferences of some other user. In an overhearer-style user evaluation of the FLIGHTS system-in which the subjects judged the appropriateness of simulated interactions for a hypothetical target user- Demberg and Moore (2006) found that the UMSR approach was preferred over the SR approach; as noted above, the subsequent Winterboer and Moore (2007) study confirmed that the UMSR system also performed better in a task-based evaluation.…”

“…AMVF-based user models have been used in a variety of generation systems and domains, including a series of recommender-style systems: GEA (Carenini and Moore 2006), MATCH (Walker et al 2004), and FLIGHTS ). All of these systems generate descriptions and comparisons of options-houses, restaurants, and airline flights, respectively-that are tailored to the user's likes and dislikes.…”

“…2.1 (e.g., (Walker et al 2004;Demberg and Moore 2006)), the primary hypothesis for this experiment was as follows:…”

“…Several recent studies have shown that a system that tailors its output to the characteristics of the target audience can produce descriptions that are preferred by users over nontailored descriptions, and that can even have a greater impact on user behaviour (Walker et al 2004;Carenini and Moore 2006;Demberg and Moore 2006).…”

The synthetic teammate project

“…The language generation and dialog management components were developed to capture the dynamic nature of human language production, following earlier approaches involving dynamic dialog constraints (Ericsson 2004), accommodation (Matessa 2000), and adaptive content selection (Walker et al 2004). The focus of the language generation component is on selecting from a set of possible utterances, akin to overgeneration-and-ranking approaches (Varges 2006).…”