People infer communicative action through an expectation for efficient communication

Abstract

Humans often communicate using body movements like winks, waves, and nods. However, it is unclear how we identify when someone's physical actions are communicative. Given people's propensity to interpret each other's behavior as aimed to produce changes in the world, we hypothesize that people expect communicative actions to efficiently reveal that they lack an external goal. Using computational models of goal inference, we predict that movements that are unlikely to be produced when acting towards the world and, in particular, repetitive ought to be seen as communicative. We find support for our account across a variety of paradigms, including graded acceptability tasks, forced-choice tasks, indirect prompts, and open-ended explanation tasks, in both market-integrated and non-market-integrated communities. Our work shows that the recognition of communicative action is grounded in an inferential process that stems from fundamental computations shared across different forms of action interpretation.

Fig. 1. Conceptual illustrations of trajectories used in Studies 1–3 and the Explanation Control.

See https://osf.io/ehb48/ for full stimuli set and videos. The trajectories used were a subset of paths built by sequentially chaining combinations of sixteen primitive movements. Paths were sorted into eight movement classes that captured high-level properties of the movements: (A) maximally efficient paths (n = 2 paths), (B) paths that retrace themselves back to their origin (n = 3), (C) paths that move toward multiple quadrants (n = 3), (D) paths that move toward only one quadrant (n = 3), (E) paths that partially retrace themselves (n = 3), (F) paths that intersect themselves, but do not start and end in the same position (n = 3), (G) paths that have repeated components that form a pattern (n = 3), and (H) paths that do not retrace themselves, but start and end in the same position (n = 3). a Example paths used in Study 1. b Example paths from the unbordered condition in Study 3. c Example paths from the bordered condition in Study 3.

Fig. 2. Results from Studies 1–3 and the Explanation Control, combining original studies and their pre-registered replications.

a Average communicativeness judgments (y-axis) as a function of the path’s rarity (x-axis) in Study 1 (n = 60), operationalized here as deviation from the shortest path (see “Methods”). b Average communicativeness judgments in Study 2 (y-axis) as a function of number of path repetitions (x-axis; n = 60). Vertical lines show bootstrapped 95% confidence intervals. Dots represent individual judgments. c Study 3 results as a function of rarity and condition. Each point represents a path with the average communicativeness in the bordered condition (x-axis) and the unbordered condition (y-axis). Smaller circles represent rarer paths. Points above the diagonal line indicate a higher communicativeness rating in the unbordered condition relative to the bordered condition. The difference across conditions was larger for paths with higher rarity in a mixed-effects model (β_{rarity:condition} = 1.14, p = 0.002, n = 30; replication: β_{rarity:condition} = 0.92, p = 0.015, n = 30), further suggesting that these paths were no longer seen as communicative because environmental constraints in the bordered condition removed the paths’ perceived rarity. d Average reported difficulty of generating a world-directed explanation as a function of the path’s rarity (x-axis) in our Explanation Control (n = 60). If Study 1 judgments were driven by a pure inability to consider world-directed explanations, then the difficulty of explanation should show a strong positive correlation with rarity, but this was not the case.

Fig. 3. Results from a computational model that determines communicativeness when the movement reveals that it lacks a world-directed goal.

Critically, this model has a more nuanced concept of rarity, as it can infer world-directed goals for inefficient movements (such as an agent zig-zagging toward an object). a Model predictions showing the belief that each path from Study 1 was world-directed as a function of time. Each line represents one of the 23 videos from Study 1 (partially occluded due to over-plotting), with frame number on the x-axis and model prediction on the y-axis. Lower model predictions indicate that the movement looked less world-directed (in log-space; see Eq. 1). Each line’s shading indicates the average communicativeness rating received for that video in Study 1. b Final model predictions (x-axis) against participant judgments (y-axis) from Study 1. To make scales comparable, we transformed model ratings into communicative inferences through a linear regression predicting participant judgments based on the model’s final output (see SI for details). Each point represents a path’s model prediction (x-axis; negative log of probability of a world-directed goal given the full trajectory—equivalent to the predictions on frame 12 of panel a)—against average communicativeness ratings from Study 1 (y-axis). c Disagreement between model predictions and participant judgments as a function of the movement class (see Fig. 1 for movement class labels). Positive numbers indicate that the model saw the movement as more communicative than participants and negative numbers indicate that the model saw the movement as less communicative than participants.

Fig. 4. Schematic of body movements used in Studies 4–7.

See https://osf.io/wxdka/ for videos. a–d Basic movements, rare movements, repetitive movements, and rare + repetitive movements. The actor’s eyes were never visible and the face was always directed at the camera, preventing participants from relying on facial information. All videos were normed to ensure that they exhibited rarity and repetition. See SI for norming data on videos.

Fig. 5. Results from Study 4a, Study 4b, and the two controls.

Each bar shows the average judgment and vertical lines show bootstrapped 95% confidence intervals. Dots represent individual judgments. a Average communicativeness judgments from US participants for Low Punctuality (Study 4a; n = 30) and Natural Punctuality (Study 4b, n = 30) demonstrators broken down by video type. b Results from the Weirdness Control and Familiarity Control studies. Participants’ beliefs about a movement’s weirdness (Weirdness Control; n = 30) or familiarity (Familiarity Control; n = 60) failed to explain our results.

Fig. 6. Results from Study 5.

Left to right: Average preference for the rare or repetitive video as communicative with US, Tsimane’, and San Borja participants. San Borja is the Spanish-speaking market-integrated town that is closest to the Tsimane’ communities, which served as a geographic control. Each bar shows the average percentage of endorsements for rare or repetitive movement when directly contrasted with a basic one. U.S. (n = 40) and San Borja (n = 40) participants completed the task with the low punctuality demonstrator videos. Tsimane’ participant results are combined across both low punctuality (n = 120) and natural punctuality videos (n = 60). Vertical lines show bootstrapped 95% confidence intervals.

Fig. 7. Results from Studies 6 and 7.

Vertical lines represent 95% bootstrapped confidence intervals. a Average percentage of trials in Study 6 where US (n = 59) and Tsimane’ (n = 59) participants judged that the actor was alone or with someone as a function of the type of video they watched. Participants were more likely to judge that the actor was alone when watching a basic movement, and more likely to infer the presence of another agent when watching a rare + repetitive movement. b Average percentage of world-directed and communicative explanations that US (n = 100) and Tsimane’ (n = 32) participants produced as a function of video type (Study 7). Results are normalized after excluding explanations coded as descriptive or other (see SI). Participants were more likely to produce a world-directed explanation when watching a basic movement, and significantly more likely to produce a communicative one when watching a rare + repetitive one (this result does not depend on the pre-registered exclusion of descriptive and other explanations; see SI).