3.6 Discussion
3.6.1 Effects of Contingent Gaze
3.5.3 Analysis of Gaze
None of the conditions reach the 30/70 ratio of human interaction (Argyle & Graham, 1976). Analysis of the number of participants’ gaze actions reveals that participants in the contingent nod condition produce significantly fewer gaze actions per minute in comparison to participants in thecontingent gazecondition (B=2.19, se=1.03, p=0.04), and participants in the random gaze condition (B=2.19, se=1.03, p=0.04). Conversely, participants in thecontingent nod condition sustain their gazes for longer time than participants in the contingent gaze (B=-1.73, se=0.74, p=0.03) and participants in the random gaze (B=-1.6,
se=0.74, p=0.4) conditions do (seeFigure 3.13).
Total Duration
(%ofInteraction Time)
Avg.
Duration
(PerGaze Action)
Occurences (Num
berPerMinute) 0
10 20
30 Contingent Gaze
Contingent Nod Random Gaze
* p<0.05
** p<0.01
*** p<0.00
**
*
Figure 3.13: Analysis of Gaze
3.5.4 Results Overview
All significant differences reported above are summarized inTable 3.1below.
Contingent Contingent Random
Gaze Nod Gaze
More expository utterances X
Fewer utterances per turn X
More references to previous events X
Fewer contacts to experimenter X
Fewer gaze actions X
Longer mean gaze duration X
Table 3.1: Summary of Significant Results
3.6 Discussion 45 when interacting with a robot that uses contingent gaze (Fischer, Lohan, Nehaniv, &
Lehmann, 2013; Fischer, Lohan, Saunders, et al., 2013)). Specifically, people reduce the linguistic complexity of their own production by splitting up their communication into smaller utterances, and by providing explanations for concepts through the use of expository utterances. Interestingly, participants in the current study produce expository utterances to a greater extent in thecontingent gaze condition in comparison to the other two conditions. This result confirms results previously found with the same contingency spotter (Fischer, Lohan, Nehaniv, & Lehmann, 2013;Fischer, Lohan, Saunders, et al., 2013) and shows that the results also extend to a different robot morphology. Interestingly, the current study was not able to replicate the result (Fischer, Lohan, Nehaniv, & Lehmann, 2013) that participants design their turns over several utterances, thus producing more utterances per turn. In fact, the exact opposite was the case. That is, participants in the contingent gaze condition produced fewer rather than more utterances per turn and they
assumed higher degrees of competence.
Establishing and Maintaining Common Ground
In order to better understand these seemingly contradictory results, the interactions were analyzed again. However, rather than looking at behavioral features that can be quantified, the analysis is based on a conversation analytical approach (Sacks et al., 1974). The qualitative analysis reveals two recurring themes that likely lead to the results reported above. Participants in the contingent nod condition and in the random gaze condition display an orientation to the importance that the robot is able to read the hamming markers on the Lego blocks. In addition, these participants do not plan the sentences they wish to construct ahead of time, but construct sentences incrementally one block at a time.
During this process they produce several tokens that display an intention to keeping the turn. These two phenomena are demonstrated inExample 3.1 below.
The participant holds up a Lego block in front of the robot (#1) and asks whether it can
see it (line 1). The robot reads out the word, which is understood as a demonstration of understanding and responded to by the participant by a positive assessment. The participant starts in line 5 with the discourse markerso, used here as a boundary marker for the coming talk. The participant then starts to look for candidate words among the Lego blocks, and while doing so produces a hesitation marker with a continuing intonation (line 8). This is followed by another hesitation marker (line 10) together with the last word the robot read (jo:hn), which is produced with a prolongation of the vowel and a continuing hesitation. As a result of these actions, he keeps the turn until he finds a second word block and holds it up (#2). Holding up a word block like this works as a turn yielding cue, which implicitly tells the robot to read the word block. After the robot reads out loud the two word blocks that are now visible, the participants responds with the assessment
goo:din line 13. The turn yielding cue of holding up the word block and the participant’s positive assessment of the robot’s behavior aptly display the participant’s orientation as to what is important. It also shows an understanding of the robot’s capabilities. It shows that the participant expects that the robot will respond in a certain manner to being shown a Lego block (namely, translating the code to a readable word and reading the word
01 *SP01: dear robot can you see this? #1 02 *ROB: john.
03 *SP01: correct.
04 (.)
05 *SP01: so i:'m gonna tell you something about john.
06 ((participant looks toward the stack of word blocks))
07 (2.5)
08 *SP01: e:hm,
09 (1.8)
10 *SP01: ehm jo:hn, #2
11 ((participant holds up a word block))
11 (2.9)
12 *ROB: john slices.
13 *SP01: goo:d. #3
#1 #2 #3
Example 3.1: Participant in the Random Gaze Condition
out loud). Using the next-turn-proof procedure, it becomes clear that this is what the participant expected the robot to do, as he assesses the robot’s turn positively both times.
This phenomenon is observed throughout the interaction, but also in other interactions, although primarily in participants who interacted with the robot in the contingent nod or therandom gaze conditions. Likewise, the floor managing produced by the participant naturally prolongs his turn.
However, participants in the contingent gaze condition behave quite differently under similar circumstances. Rather than displaying an orientation to the robot’s ability to read the word blocks, participants in thecontingent gaze condition make other aspects of their communication with the robot relevant. Specifically, while they hold up word blocks as was seen inExample 3.1, they do not wait for the robot to read the word out loud, nor do they produce positive assessments of the robot’s ability. Instead, they spend time explaining the grammatical concepts that each of the word blocks represents, thus making an implicit reference to the introductory phase, in which the concepts were first introduced to the robot. This phenomenon is demonstrated inExample 3.2.
The participant starts with the discourse markerso, which, as in the previous interaction, works here as a boundary marker for the next action. Thus,second onerefers to the second sentence the participant is going to construct (having just completed the first one). By sayingsecond one, the participant shows here that she considers previous as actions as common ground between them. This can be seen as evidence that he assigns epistemic competence to the robot. As also seen inExample 3.2 above, the participant holds up the word block while saying john, but already after the micro pause in line 1 she moves
3.6 Discussion 47
01 *SP01: so second one (.) john. #1
02 ((participant attaches the word block to the platform))
03 (1.8)
04 *SP01: john.
05 *ROB: john.
06 *SP01: mm hm that's the subject of a sentence. #2
07 (0.5)
08 *SP01: then (.) eats.
09 (0.4)
10 *ROB: john eats.
11 *SP01: yes (.) to eat is a verb. #3
#1 #2 #3
Example 3.2: Participant in the Contingent Gaze Condition
the word to the platform. After she attaches the word block to the platform (line 2) she repeats the word, john (line 4). The robot reads out the word in line 5, but this is responded to only by a minimal response, which is followed by an exposition of the word’s grammatical position in the sentence she is forming. She does this by tapping on the word block ‘john’, while already holding up the next word block (#2). The robot then reads out the two words now visible. Although this is responded to by the participant with a positive assessment (yes), it is within the same utterance followed up by another exposition of the grammatical position the word has in the sentence being formed. Thus, the participant here does not perform any work on the floor management, since it is not necessary. While she displays an orientation to the robot’s ability to read the word blocks as a conditional relevant next action (as seen in lines 6 and 11), it is treated here as a cue to teach the robot something new, rather than to confirm that the robot is working within specified parameters, which is the case forExample 3.1. Again, this phenomenon is recurring not only in this particular interaction, but also in other interactions, in which participants interact with the robot in thecontingent gaze condition. These two interaction formats are summarized inFigure 3.14.
The quantitative analysis suggested that participants reduce the linguistic complexity in their communication with the robot for some measures (expository utterances), while the opposite were true for other measures (number of utterances per turn). However, the qualitative analysis presented above suggests that the linguistic features investigated are used by participants in very different ways. Fischer, Lohan, Saunders, et al. (2013), Lohan et al. (2011)argue that contingent gaze allows people access to the robot’s competences, and as a result thereof they adjust their communication accordingly. For participants in
Holds up block
Reads word
Waits for robot to read word P
P
P
Produces assessment P
Response Non-response R
P Participant R Robot Holds up block Reads word P
Explains word function P
R Random gaze/ Contingent nod
Conditions Contingent gaze
Condition
Response Non-response
Figure 3.14: Interaction Formats
the random gaze and contingent nod conditions of the current study the gaze was not meaningful. Some evidence of this is found in participants’ need to confirm what words are in use, and to asses the robot’s performance in that confirmation. However, for participants in the contingent gaze condition the gaze was indeed meaningful and displays not only an awareness towards verbal actions (which is equal for all three conditions), but also for non-verbal actions. Therefore, participants did not need to confirm what the robot knows, but could move on to start teaching the robot. This conclusion is additionally supported by the number of references participants make to previous events. Here, participants in thecontingent gaze condition not only implicitly refer to the introductory phase as seen in Example 3.2, but also make significantly more explicit references to previous actions and events. These references are realized in utterances such as“this is similar to...”, and “as I said before”. That is, participants in thecontingent gaze condition establish early on in the interaction a common ground (Clark, 2002) through which the interaction unfolds and builds upon. Part of this common ground is everything that has occurred up to this point in the interaction, an assumption that participants in the other condition are likely not to make.
Contingency or Timing?
The results so far point in the direction that contingent gaze gives the impression of a responsive and situationally aware agent with which one can interact with in socially meaningful ways. As such this is a confirmation of previous work, which reach similar conclusions (e.g. Fischer, Lohan, Saunders, et al. (2013), Fischer et al. (2012), Lohan et al. (2011), Skantze et al. (2014)). In addition, it demonstrates that results obtained in
3.6 Discussion 49 previous work are stable even when adopted to a different robot morphology.
Furthermore, one of the aims of this study was to investigate the differences between contingency and timing. That is, the study addresses whether the positive aspects of contingent gaze, reported here and by others, can be attributed to the gaze action being contingent, or whether the gaze be replaced by any other contingent action, as indicated by the results reported by Mehlmann et al. (2014). The current study suggests that gaze alone (as for example therandom gaze condition) does not bring forth the positive aspects discussed above, either. Likewise, replacing the gaze action with another contingent modality (as for example the contingent nod condition) does also not bring forth the positive aspect of contingent gaze either. Thus, contingent gaze is a social behavior that people recognize and which displays an awareness to human behavior in social interaction.
This display directly influences how the interaction unfolds, what assumptions people make about their interaction partner, and which linguistic resources they put to use. However, people may be completely unaware of this influence, as suggested by the lack of differences between conditions for the subjective ratings. In conclusion, contingency is not merely a matter of producing ‘some action’ at the right time, but is an interaction governed by timing, conditional relevance, and modality.