Academia.eduAcademia.edu

Figure 2 – uploaded by K Jokinen

See full PDFdownloadDownload figure
Fig. 2. Feedback and gaze in the Estonian data. The attributes on the y axis are from the top: Emphasis, Elicit, Non-accept, Accept, CPU, CP and No-feedback. occurs when Turn Hold. Looking at Turn and Gaze, Turn End goes together with Gaze Interlocutor (the speaker looks at the interlocutor to ask for feedback when releasing the turn), and Gaze Down goes rather systematically with Turn Hold (by not looking at the interlocutor, the speaker makes it harder for them to interrupt).

Figure 2 Feedback and gaze in the Estonian data. The attributes on the y axis are from the top: Emphasis, Elicit, Non-accept, Accept, CPU, CP and No-feedback. occurs when Turn Hold. Looking at Turn and Gaze, Turn End goes together with Gaze Interlocutor (the speaker looks at the interlocutor to ask for feedback when releasing the turn), and Gaze Down goes rather systematically with Turn Hold (by not looking at the interlocutor, the speaker makes it harder for them to interrupt).

Related Figures (7)

Table 1. Functional annotation features and semiotic types Table 2. Annotation features for facial displays
Table 1. Functional annotation features and semiotic types Table 2. Annotation features for facial displays
additional attribute is related to whether the feedback is one of acceptance or refusal. A third option is to add a value for Emotion. Turn management is coded by the three general features Gain, End and Hold. An additional dimension con- cerns whether the turn changes in agreement between the speakers or not, thus giving rise to different values for Gain and End. The semiotic categories are based on Peirce (1931). They correspond roughly to pointing gestures (Indexical Deictic); beats (Indexical Non-deictic); gestures expressing a semantic feature by similarity, including metaphoric gestures (Iconic); and emblems (Symbolic). Table 2 shows the features available in MUMIN to describe shape and dynamics of facial displays (hand gestures and body posture features are also part of the scheme but were not considered for this experiment). The categories are inten- tionally coarse-grained, in that we only want them to be specific enough to be able to distinguish and characterise non-verbal expressions that have a feedback, turn management and sequencing function.
additional attribute is related to whether the feedback is one of acceptance or refusal. A third option is to add a value for Emotion. Turn management is coded by the three general features Gain, End and Hold. An additional dimension con- cerns whether the turn changes in agreement between the speakers or not, thus giving rise to different values for Gain and End. The semiotic categories are based on Peirce (1931). They correspond roughly to pointing gestures (Indexical Deictic); beats (Indexical Non-deictic); gestures expressing a semantic feature by similarity, including metaphoric gestures (Iconic); and emblems (Symbolic). Table 2 shows the features available in MUMIN to describe shape and dynamics of facial displays (hand gestures and body posture features are also part of the scheme but were not considered for this experiment). The categories are inten- tionally coarse-grained, in that we only want them to be specific enough to be able to distinguish and characterise non-verbal expressions that have a feedback, turn management and sequencing function.
Table 3. Input statistics
Table 3. Input statistics
Table 4. Simplified feedback attributes
Table 4. Simplified feedback attributes
Table 5. The classification accuracy of communicative functions on the Estonian anc Danish data.
Table 5. The classification accuracy of communicative functions on the Estonian anc Danish data.
To start with feedback, the following patterns emerge in the Danish data between Gaze and Feedback: gaze to Interlocutor seems to coincide with Ac- cept while gaze to Side, Other or Down corresponds to No feedback. Also Head and Feedback pattern together: nods are accompanied by Accept, while side- turns and side-tilts have No feedback. Head is usually Jerk or leaning Forward if Feedback is Elicit. It must be remembered that the data only show gesture behaviour in two subjects: therefore, these patterns may only concern these two specific individuals. However, both the gaze patterns and the head movements also confirm our intuitions about the way in which feedback is expressed. Fig. 1. Feedback clusters in the Danish data. The attributes on the y axis are from the top: Emphasis, Elicit, Non-accept, Accept, CPU, CP and No-feedback.
To start with feedback, the following patterns emerge in the Danish data between Gaze and Feedback: gaze to Interlocutor seems to coincide with Ac- cept while gaze to Side, Other or Down corresponds to No feedback. Also Head and Feedback pattern together: nods are accompanied by Accept, while side- turns and side-tilts have No feedback. Head is usually Jerk or leaning Forward if Feedback is Elicit. It must be remembered that the data only show gesture behaviour in two subjects: therefore, these patterns may only concern these two specific individuals. However, both the gaze patterns and the head movements also confirm our intuitions about the way in which feedback is expressed. Fig. 1. Feedback clusters in the Danish data. The attributes on the y axis are from the top: Emphasis, Elicit, Non-accept, Accept, CPU, CP and No-feedback.
Table 6. 10-fold cross validation for facial displays
Table 6. 10-fold cross validation for facial displays
Related topics:
Computer ScienceArtificial IntelligenceMachine Learning
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved