AtGentive IST-4-027529-STP - Attentive Agents for Collaborative LearnersKey contributions of the report include: 1) a multi-staged conceptualisation ofattention to disambiguate controversial findings in the ECA literature; 2) a method forthe study of executive attention in human-agent interaction; 2) the evaluation ofColette’s non-verbal communication; 3) experimental results showing that embodiedagents do indeed capture user attention, and can disrupt performance under certaincircumstances; 4) an experimental validation of the persona effect.2. AttentionAttention serves as a set of mechanisms which regulate cognitive processes andfeelings. Recent advances in neuroimaging techniques have supported the existenceof different cognitive networks relating to specific aspects of attention (Posner andRothbart, 2007). Three different networks were identified which supports differenttypes of tasks: alerting, orienting and executive attention.• Alerting is the achievement and maintenance of a state of arousal, orsensitivity to incoming stimuli.• Orienting involves the selection of information from a source of incomingstimuli.• Executive attention involves mechanisms capable to monitor and resolveconflicts among incoming stimuli (physical objects and events, thoughts, andfeelings).These three tasks can be conceived as separate steps which lead to informationprocessing. Alerting stays at the button level: it refers to arousal (the subject is readyto receive information). Orienting and executive attention are involved at differentstages of the selection of information. Cognitive processes happening at the level ofexecutive attention regulate the contents of working memory (Engle, 2002). Executiveattention is the ability to maintain or suppress information in working memory,focussing to relevant parts of the perceptual field, while ignoring tasks irrelevantstimuli. Hence, executive attention is involved in a variety of higher-cognitive tasksunderlying intelligence, such as reading and listening, learning, and self-regulation ofpositive and negative affects.This distinction of attention as separate networks devoted to specific tasks isimportant and may help to clarify some of the contradictory results reported in the HCIliterature.2.1 Attention & ECAIn recent years, increasing effort has been devoted to the study of the distribution ofthe user attention to different elements of the computer interface during taskexecution (Roda and Thomas, 2006). ECA’s are special interface elements, as theiranthropomorphic appearance can induce social attributions and biases. The humanface is an extraordinary stimulus. Research in psychology has demonstrated anextremely efficient detection of facial expressions, with a particular relevance to threatand fear (Hansen and Hansen, 1988). There is evidence that affective facialexpressions are automatically processed and can interfere with other tasks (Stenberget al., 1998). If the emotion conveyed by a face does not match the emotional valenceof a verbal message, understanding is delayed or even impeded. Significant for thedesign of ECA’s is the finding that emotional expressions in a face can be perceivedoutside the focus of attention and tend to guide focal attention to the location of theDel <strong>4.4</strong>: AtGentive Final Evaluation Report – Appendix A page 2
AtGentive IST-4-027529-STP - Attentive Agents for Collaborative Learnersface (Eastwood et al., 2001). Evidence of the importance of consistency betweenverbal messages and facial animations in human-agent interaction is reported by(Berry et al., 2005). Inconsistency strongly decreased memory for verbal information.These findings are of utmost importance for the design of ECA’s, as the status of theart in graphical rendering of emotions and their synchronisation with timing andcontent of the verbal message cannot still guarantee a perfect match between the twoinformation channels. In case of conflict, ignoring the agents’ facial expression andfocussing on the verbal message may be complex. Hence, we may expect that agentinteraction in suboptimal conditions can hamper communication.Although attention is a common dependent variable in many evaluations of ECA’s,conclusive evidence is still missing. Several factors can be held responsible for sucha lack of agreement (Clark and Choi, 2005, Dehn and van Mulken, 2000, Gulz, 2004).Firstly, empirical research on ECA’s suffers from a generalised lack of methodologicalrigour, affecting operational definitions of core constructs, and, as a consequence,methods and procedures for measuring them. Secondly, most evidence consists ofresults from single ecological studies, whereas scientific generalisation would requirea series of experimental studies. Generalisation in PA research is further complicatedby the large variance introduced by testing different learning environments, users,and agent instantiations. Finally, these evaluations tend to address a large set ofdependent variables at once, including performance indicators (e.g., learningoutcomes), cognitive processes (e.g., attention allocation, memory, problem-solving),motivational and attitudinal measures (e.g., willingness to use, satisfaction). Althoughecological studies have potentials in addressing social and motivational variables,they lack the control required by the study of cognitive processes.2.1.1 Operational shortcomingsAt the heart of the persona effect lays the assumption that non verbal cues conveyedby an embodied agent have the potential to guide the user attention towardsimportant elements of the task at hand. Analysing this effect within theconceptualisation of attention as a multi-staged process implies that embodied agentshave the capability to increase alertness, support attention orientation, and that, indoing so, they do not add any demands to the executive control of attention (e.g.,there is no distraction induced by the increase in stimulation). We believe that thismultilayered framework of analysis can help interpret some of the inconsistencies inthe empirical research on the role of attention in agent-human interaction.In human-agent research, attention has been defined in many different ways. Forinstance, it has been associated to time spent performing a primary task, such asplaying cards (Takeuchi and Naito, 1995) or filling in a questionnaire (Sproull et al.,1996) while interacting with a virtual face as compared to a control condition (i.e., noface). Although both studies reported longer response time in the face condition, theirauthors interpreted this effect in opposite ways. Takeuchi and Naito (1995)associated the longer time to distraction. They claimed that the user attention wasdetracted from the primary task because it focused on interpreting the facialexpressions of their virtual opponent. Conversely, Sproull and colleagues (1996)explained the longer time taken by users when answering psychological tests in theface condition as a measure of attentiveness. They concluded that people werepaying more attention to the primary task, when the questionnaire was presented bya character than when it was presented by textual display. This conclusion wasderived by the association of performance time with increased arousal, whichaccording to the authors fostered self-reflection, thus slowing down the activity. Analternative interpretation could be that both studies implicitly dealt with executiveDel <strong>4.4</strong>: AtGentive Final Evaluation Report – Appendix A page 3