Interpersonal Circumplex Measurements and Interpretation
As a reminder, our goal is to develop a generative model of attitude variations for a virtual agent. We will evaluate our model by the perception of the generated attitudes through a questionnaire. To find the most relevant adjectives that characterize the perception of attitude, we perform a literature review on the usage of IPC measurements. This study will help us fill the evaluation requirements.
The IPC has recently become a popular model for assessing interpersonal dispositions such as interpersonal problem (e.g., problems related to assaulting others) [Alden et al., 1990], value (how interpersonal experiences, such as expressing herself openly, are important to a person?) [Locke, 2000], self-efficacy (interpersonal actions a person believes she can express) and traits (e.g., firm) [Wiggins, 1995]. All the IPC measures are based on the same theory: there is a particular location within the circumplex space for each inter-personal disposition. Most IPC inventories split the IPC into eight octants or scales that are alphabetically labeled counterclockwise: P A, BC, DE, F G, HI, JK, LM and N O (see Figure 2.1). Each octant can be represented by a set of characteristic adjectives, e.g., dominant and assertive for P A octant.
To build an IPC inventory, psychologists started by building the questionnaires describ-ing the measured interpersonal dispositions, for example, by analyzing psychotherapy interviews. Then, using statistical analyses such as Principal Component Analysis, par-ticipant answers were clustered and displayed on the IPC. The works of Locke, Adamic, Acton, and Revelle provided overviews of interpersonal circumplex measures or invento-ries [Locke and Adamic, 2012, Acton and Revelle, 2014]. As they reported, the Inter-personal Check List (ICL), proposed by Leary, was the first ICP inventory [Leary, 1957]. Based on Sullivan’s interpersonal theory of personality [Sullivan, 1953] on one hand, and observing interactions among psychotherapy group members on the other hand, Leary constructed a circumplex model that represented interpersonal traits (cf. Figure 2.3). As we can see, Leary classified 16 interpersonal behaviors on the interpersonal circle. Each of the 16 behaviors is evaluated by multi-level measures: (1) reflexes are illustrated in internal circle and indicated by alphabetical letters (A to P). (2) The center ring indicates the behaviors provoked by persons adapting the interpersonal behaviors. For example, a person who uses the reflex P tends to provoke others to respect. (3) The next circle illustrates extreme reflexes like compulsive and dominant. Finally, the circle perimeter is divided into eight interpersonal behaviors (e.g., managerial-autocratic). The ICL model has been widely used in psychological and socio-psychological research [Clark, T. L., & Taulbee, 1981]. However, researchers reported that ICL did not adequately fit the circum-plex model. It presents significant measurement gaps between the four quadrants of the circumplex [Kiesler, 1996, Wiggins et al., 1988, Locke, 2000]. Wiggins et al. [Wiggins, 1979] proposed the Interpersonal Adjective Scales (IAS) to address these limitations. An interpersonal adjective is defined as “a pattern of dyadic interactions that has rel-atively clearcut social (status) and emotional (love) consequences for both participants (self and other)” [Wiggins, 1979] (p. 398). Based on this definition, 800 terms were identified as interpersonal. For simplifying the rating and the interpretation, the 800 terms have been reduced to 128 adjectives and then to 64 in the final version of the IAS (IAS-R). To validate the IAS model, participants rated how accurately each adjective describes them on a 8-point scale. The methodology used to build IAS served as a basis for the develop-ment of other IPC measures like the Inventory of Interpersonal Problems (IIP). Moreover, IAS is now the standard measure of interpersonal traits [Locke and Adamic, 2012]. Horowitz et al. studied a large sample of psychotherapy interviews for reporting the most frequent interpersonal problems [Horowitz, 1997]. Based on this study, the Inven-tory of Interpersonal Problems (IIP) was developed. It consists of 64 items that assess interpersonal excesses and deficiencies. participants rated how distressed they have been for each problem on a 5-point scale. The items are divided into two sections: “It is hard for me…” and “I am too much…”. IIP was used to identify the relationship between inter-personal problems and psychopathology and psychotherapy [Ruiz et al., 2010]. Locke reported that the IIP can help guide therapeutic interventions for interpersonal prob-lems [Locke and Adamic, 2012]. For example, the interpersonal problems assessed by the IIP are related to the types of interpersonal expectations that are readily targeted by therapeutic interventions. For example, dominant people expect others to be critical whereas friendly people expect others to be dismissive.
Self-efficacy is how confident a person is able to perform some action [Rogelberg, 2017]. The Circumplex Scales of Interpersonal Efficacy (CSIE) assess a person’s confi-dence that she can successfully perform behaviors [Locke and Sadler, 2007]. Answers range from 0 (not at all confident) to 10 (absolutely confident).
Table 2.1 gives a summary of some IPC measures and table 2.2 indicates some repre-sentative adjectives for each octant of the IPC. For each measure, we indicate: the number of items used to measure a specific interpersonal disposition, the question and its answer scale, as well as the number and kind of participants who answered the questionnaire in the initial study.
Scoring and Interpreting the IPC Measurements
After choosing a suitable IPC inventory depending at the task at hand, the next step is to score and interpret the answers of participants. One commune approach for analyzing such data is the circular profile. This profile presents each person’s scores on the eight oc-tants of the circumplex (cf. Figure 2.4). For computing this profile from any IPC inventory, Locke follows three steps [Locke, 2012]:
1. Compute the general factor score by averaging the eight octant scores.
2. Ipsatize octant scores by subtracting the general factor score from each octant score.
3. Plot the ipsatized scores on the IPC ranging from the lowest value to the highest value. To interpret the circular profiles, Gurtman explains: “circular profiles tend to rise to a peak value and then decline. The peak clearly indicates the predominant trend in the profile and suggests the individual’s predominant interpersonal style or typology” [Gurtman, 2009b]. Figure 2.4 plots the circular profiles of two persons P1 and P2 who answered the CSIE inventory. For the profile P1, the peak is in the lower-right region which suggests a friendly-yielding behavior, whereas for P2 the peak is in the lower-left quadrant suggesting a hostile-yielding style. Based on the circular profile, we can also compare the behaviors of both participants P1 and P2: they are similar in efficacy for being dominant (P A) and yielding (HI). On the opposite, participant P 1 is more friendly (LM) than distant (DE), unlike participant P 2.
Leary introduced another approach called vector scoring by summarizing the circular profiles with a single point on the circumplex [Leary, 1957]: the vertical coordinate gives the perceived dominance based on Equation 2.1 whereas the horizontal coordinate char-acterizes the friendliness based on Equation 2.2, by combining the ipsatized octant scores as indicated in [Wiggins, 1979]. For example, the vertical coordinate (DOM) represents the weight of the octant scores according to their directions compared to the dimension of dominance. Thus, we sum the scores of P A, BC and N O (vary in the same direction as dominance) and we subtract the scores HI, F G and JK (vary in the opposite direction of dominance). The values of DOM and F R define a vector in the IPC space whose angle can be calculated by Equation 2.3 and length by Equation 2.5. The angle is adjusted as indicated in Equation 2.4. The vector angle indicates the predominant interpersonal be-havior [Wiggins et al., 1988, Gurtman and Balakrishnan, 1998, Gurtman, 2009a, Locke and Adamic, 2012].
DOM = 0:03 (PA HI) + 0:02 (NO + BC FG JK) (2.1).
F R = 0:03 (LM DE) + 0:02 (NO BC FG + JK) (2.2).
Multimodal Expressions of Social Attitude
Nonverbal behavior is an important component in human interaction. It can participate to the regulation of interaction (e.g., nodding may indicate an agreement with the speaker), it can complete and structure the speech (e.g., raising eyebrows can accentuate an ele-ment of the speech ) [Ekman and Friesen, 1969, Argyle, 1988, Cosnier, 1997]. Non-verbal behavior also contributes to the expression of emotions and attitudes [Argyle, 1988]. In our work, we are interested in the expression of interpersonal attitude through non-verbal behavior. The relationship between non-verbal behaviors and attitudes has been widely investigated in psychology and sociology [Gifford, 1991, Mehrabian, 1969]. In the fol-lowing, we summarize the most significant findings on the relation between non-verbal behaviors and interpersonal attitudes.
Gestures: McNeill categorized gestures into two main categories: communicative gestures and adaptor gestures [McNeill, 1992]. Gestures can bear information about the speaker’s attitude. For example, adaptor gestures that consist in touching oneself or manipulating objects are mainly related to submissive attitude but can also be as-sociated with hostility in some cases [Burgoon, J. K. and Le Poire, 1999]. Touching her interlocutor can be a sign of friendliness and of dominance depending on the type of the touch [Carney et al., 2005, Burgoon et al., 1984]. Frequency and expres-sivity of gestures, like amplitude and intensity, directly influence the perception of an attitude. Also, performing large gestures may be a sign of dominance. Dominant people are also generally characterized by gesturing more compared to submissive people.
Postures: when two interacting persons adapt unconsciously their postures one to another, we can predict a change in their interpersonal attitudes [Richmond, V. & McCroskey, 2000]. Lafrance noted that postural mirroring (adopting the same pos-ture as one’s interlocutor) can be a sign of friendliness [Lafrance, 1982]. Leaning towards and taking a closer position to her interlocutor can be perceived as a sign of submission, whereas reverse behaviors, such as leaning backwards, could express dominance [Carney et al., 2005, Burgoon et al., 1984, Burgoon, J. K. and Le Poire, 1999]. Adopting a posture by occupying a large space, in the same way as large ges-tures, are signs of dominance [Carney et al., 2005, Burgoon et al., 1984, Burgoon, J. K. and Le Poire, 1999, Gifford and Hine, 1994]. For example, dominant people extend their legs more than submissive people do [Gifford, 1991].
Head direction and movement: communicative functions of head movements are also varied. When listening, head movement can be a backchannel indicating an agreement, disagreement or understanding [Heylen et al., 2008]. Head direction and movement can also be relevant signals in predicting attitude. A bowed head can be a sign of submission, a head tilt of friendliness whereas a raised head may express dominance [Gifford, 1991, Debras and Cienki, 2012, Stivers, 2008]. On the other hand, a head shake can correlate to different attitudes: dominance [Gifford and Hine, 1994, Carney et al., 2005, Hall et al., 2005] and friendliness [Burgoon, J. K. and Le Poire, 1999, Gifford, 1991], depending on the context. Gaze: gaze is a crucial element in measuring social dimensions such as engage-ment [Kendon, 1967, Abele, 1986] and attitude [Argyle, M., Dean, 1965, Duncan, St. jr., Fiske, 1977, Burgoon et al., 1984, Hall et al., 2005]. Mutual gaze is a sign of dominance and friendliness whereas gaze shift is perceived as a sign of submission, while direct gaze is a sign of dominance. Generally, dominant people gaze more at their interlocutors than submissive ones [Hall et al., 2005].
Facial expression: the role of facial expression and their impact on the perception of attitude have also been studied [Knutson, 1996, Tiedens et al., 2000, Carney et al., 2005]: joyful expressions are associated with friendliness and dominance, fearful and sadness expressions with submission, while anger and disgust expres-sions are linked to hostility and dominance. Smile has been reported as the typical signal of friendliness [Keating and al, 1981] but could also express dominance in some situations [Hall et al., 2005]. Finally, Keating et al. studied the influence of eyebrow movements on attitude perception and showed that, generally, a frown eyebrow is perceived as expressing dominance while a raised eyebrow expresses submission [Keating and al, 1981].
Table of contents :
1.1 Context and Research Issues
1.3 Manuscript Organization
2 Theoretical Background
2.1 Attitude Definition
2.2 Attitude Representation
2.3 Interpersonal Circumplex Measurements and Interpretation
2.3.2 Scoring and Interpreting the IPC Measurements
2.4 Multimodal Expressions of Social Attitude
2.5 Non-verbal Behavior Interpretation
3 State of the Art on Attitude Modeling for Virtual Agents
3.1 Attitude Modeling for Embodied Conversational Agents
3.1.1 ECA’s Behavior Expressing Attitude
3.1.2 Attitude Dynamics over Time
3.1.3 Attitude Generation Models
3.2 Sequence-Based Multimodal Behavior Modeling
4 Sequence Mining: State of the Art and Our Algorithm
4.1 Non-Temporal Algorithms
4.2 Temporal Algorithms
4.3 Temporal Sequence Mining Algorithms
4.4 HCApriori Algorithm
4.4.3 Evaluation and Results
4.5 Pattern Quality Assessment
5 Sequence-Based Attitude Variation Modeling
5.1 Extraction of Relevant Patterns Expressing Attitude Variations
5.1.1 Building Sequence Databases Representing Attitude Variations
5.1.2 Pattern Extraction
5.2 Evaluation of the Extracted Patterns
5.2.1 Experimental Design
6 Attitude Planner
6.2 Sequential Attitude Planner Model
6.2.1 Intention Sequence Generation
6.2.2 Attitude Sequence Selection
6.2.3 Intention Sequence Enrichment
6.2.4 Signal Replacement
6.3.1 Experimental Design
7 Generative Model of Agent’s Behaviors in Human-Agent Interaction
7.1 Related works
7.3 Neural Networks and LSTM
7.3.1 Neural Networks: Overview
7.3.2 Recurrent Neural Networks: LSTM
7.4 LSTM Model
7.4.1 Prediction Model
7.5.1 EyesWeb: User’s Behavior Analysis
7.5.2 Flipper: Dialog Management
7.5.4 BehaviorPrediction: Agent’s Behavior Prediction
7.6.1 Independent Variables
8 Engagement Modeling in Dyadic Interactions
8.1 Related Works
8.1.1 Engagement-Related Behaviors
8.1.2 Engagement Prediction
8.2 LSTM Model for Engagement Prediction
9.1.1 Attitude Variation Modeling
9.1.2 Adapting Agent’s Behavior According to the User’s Behavior
9.1.3 Engagement Prediction
9.2 Limits and Perspectives
A Results of the First Study
B Results of the Second Study
C Engagement Prediction