关键词: 意图识别, 客体导向性意图, 社会性意图, 成本最小化, μ抑制

Abstract:

The recognition of the intentions of observed actions or behaviors is an important social function in the human visual system. Previous research has invested much effort into understanding how the human vision system recognizes object-directed intentions of actions (i.e., actions are implemented to approach physical objects without influencing others). However, actions are also directed to social entities or agents to impact others, which is defined as social intentions (or social interaction intentions).
This study aimed to investigate how vision systems recognize social intentions. For the purpose of demonstrating that individuals involved in social interaction are rational and should maximize the utility of actions overall, this paper proposes the hypothesis that when the costs of Agent A helping Agent B to achieve the goal state are less than the costs of Agent B acting alone to achieve this goal state (i.e., cue of minimizing cost), these two agents are recognized with social intentions.
To test the above hypothesis, we manipulated the theory of minimizing costs by presenting cartoonized animations that depict how two agents move and influence each other. Specifically, a movable Agent A placed an apple in front of Agent B, who is always stationary. In front of Agent B, a fence was either set in place or not, to operationalize the costs of Agent B achieving the target (i.e., the apple). In this case, when a fence was placed in front of Agent B, his path of achieving the target was blocked; accordingly, when Agent A pushed the apple to the front of Agent B and helped Agent B achieve the target, the costs of Agent B (i.e., path) to achieve the target were less than when Agent B achieved the target alone. However, when there was no fence in front of Agent B, the costs of Agent B achieving the target alone were less than the costs of Agent A helping Agent B to achieve the target. In brief, only when there was a fence in front of Agent B, the actions of Agent A in placing the apple in front of Agent B aligns with the information of minimizing costs, and can be recognized as a social intention; when there was no fence in front of Agent B, the object-directed intention should be attributed, as Agent A approached the target of the apple. To identify the recognized intentions of actions, we measured μ suppression (electroencephalogram oscillations within the 8~13 Hz range in the sensorimotor regions; namely, C3 and C4 channels) related to action understanding. It was suggested that the functional grouping of two individuals in social intentions should induce greater suppression of the representation of individual object-directed actions. To strictly control for the possible low-level differences, the action of Agent A, placing the apple in front of Agent B (i.e., transferring action) was paired with the action of Agent A, placing the apple in front of a stone (i.e., disposing action), which was typically recognized as an object-directed intention, whether the fence was present or not. Each action lasted two seconds, and participants were asked to count the fillers (i.e., incomplete actions) when watching the actions presented on the screen.
In Experiment 1, when the fence was present in front of Agent B, the transferring action (M = -17.3% relative to the baseline) induced more μ suppression than the disposing action (M = -8.5%). More importantly, the occipital α with the same frequency band as μ was not modulated by the action type, but this component was suggested to be functional with attentional mechanisms. These results were further confirmed by cluster-based permutation tests without selecting the channels of interest. In Experiment 2, to test whether the effect in Experiment 1 was dependent upon the information of minimizing costs, the fence was removed and accordingly, the critical information was absent. We found that the difference in μ suppression between transferring and disposing actions was insignificant when the fence was not present.
To further test the hypothesis proposed in this study, we used a behavioral indicator (i.e., measuring the sensitivity of changes). We manipulated the information of minimizing costs, as in Experiments 1 and 2, but participants engaged in a change detection task. In this task, a set of identical actions were memorized in sequence and participants were required to detect whether anything changed in the test animation compared to those previously memorized. It has been suggested that chunking results in more efficient processing of the configuration (e.g., encoding of interactants’ identity), but involves a cost for the individual parts within it, resulting in a memory confusion effect. Hence, if Agent A is perceived as having a social intention toward Agent B, they should be chunked in memory. Accordingly, participants would be less likely to detect changes within the interaction (i.e., the roles of Agents A and B in an interaction were swapped during the test; defined as role swap), but would be more likely to detect changes in pair composition (i.e., the recipient in an interaction was replaced by the recipient from another interaction, defined as structure change) relative to kinematically identical non-social transferring actions.
It was found that in Experiment 3a, when the fence was placed in front of Agent B in the role swap condition, participants were more sensitive to such change in the disposing action (M = 1.97, SE = 0.25) than to the transferring action (M =1.38, SE = 0.24); by contrast, in the structure change condition, the sensitivity of detecting such change in the transferring action (M = 2.04, SE = 0.21) was higher than that of the disposing action (M = 1.51, SE = 0.23). In Experiment 3b, when there was no fence in front of Agent B, participants were even more sensitive to the role swap change than the structure change, but it was not influenced by the action type.
It has been widely suggested that the disposing action is attributed to an object-directed intention (i.e., regardless of whether the fence was present or not) and the recognized social intention should induce greater suppression and higher sensitivity for a structure change and lower sensitivity for a role swap change than the recognized object-directed intention. Hence, we concluded that the results, in which the transferring action induced more μ suppression and higher sensitivity for a structure change and lower sensitivity for a role swap change than the disposing action (i.e., when the fence was present), were attributed to the fact that the transferring action was recognized as having a social intention. However, this recognition depends on the information of minimizing costs; otherwise, the difference in μ suppression and different sensitivities of changes between transferring and disposing actions would be observed as well, when the fence was not present. Hence, this study provides solid evidence that when the costs of Agent A helping Agent B to achieve the goal state are less than the costs of Agent B acting alone to achieve this goal state (i.e., minimizing costs), they are recognized with social intentions.

Key words: intention recognition, object-directed intention, social intention, minimizing costs, μ suppression