This paper explores a fundamental question in linguistics and cognitive science: when we listen to someone, do we use a "one-size-fits-all" model of human communication, or do we tailor our interpretations based on who is speaking? By comparing how adults interpret messages attributed to other adults versus 4-year-old children, the authors demonstrate that our "pragmatic" expectations—our ability to read between the lines—are deeply tied to our beliefs about the speaker’s reasoning sophistication.
TL;DR
Listeners are significantly less likely to draw pragmatic inferences (reading into what a speaker didn't say) when they believe the speaker is a 4-year-old child rather than an adult. This suggests that human language processing involves building speaker-specific models of rationality rather than assuming a universal standard of cooperativity.
What Problem Does This Paper Address?
In classical linguistics, the "Cooperative Principle" suggests that we expect speakers to be informative and rational. For example, if you ask, "Do you have siblings?" and someone says, "I have a brother," you infer they only have one brother. If they had two, saying "I have a brother" would be under-informative (a violation of the Gricean Maxim of Quantity).
However, do we hold a 4-year-old to the same standard? Previous research in the Rational Speech Act (RSA) framework often assumes a generic, perfectly rational speaker. This paper addresses the gap between these idealized models and real-world social cognition: do listeners adjust their "pragmatic" interpretations based on the perceived cognitive limits of their conversation partner?
Methods
The researchers used a reference game paradigm. Participants were shown three objects (e.g., a red square, a red triangle, and a green circle) and received a message (e.g., "red") intended to identify one of them.
- Unambiguous Trials: The message only fits one object.
- Critical (Implicature) Trials: The message is literally true for two objects (the "target" and the "competitor"). However, the competitor has a unique label available (e.g., "triangle"), while the target does not (e.g., "square" was an "inexpressible" message). A pragmatic listener should infer the speaker meant the target, reasoning that if they meant the competitor, they would have used the more specific word "triangle."
The key manipulation was the speaker identity. Participants were told the messages came from either another adult or a 4-year-old child. The researchers also collected open-ended explanations of the participants' reasoning to categorize their strategies.
Datasets
The study involved two experiments with native English speakers recruited via Prolific:
- Experiment 1 (N=79): Used a "scientist communicating with an alien" cover story for the child condition.
- Experiment 2 (N=160): A replication that removed the alien story and added photos of a child to make the speaker identity more salient and comparable to the adult condition.
Key Insights and Ideas
- The "Child Discount": Listeners give children a "pragmatic pass." When the speaker was a child, participants were much more likely to interpret the message literally (guessing 50/50 between the two red objects) rather than drawing a pragmatic inference.
- Belief-Driven RSA: The authors propose a modified RSA model where the listener’s interpretation is a mixture of two possibilities: the speaker is a "literal" speaker ($S_0$) or a "pragmatic" speaker ($S_1$). This is weighted by $\lambda$, representing the listener's belief in the speaker's rationality.
- Individual Variability: Not everyone adjusts their model. Some participants acted as "literal listeners" regardless of the speaker, suggesting they either didn't engage in pragmatic reasoning themselves or didn't think the task required it.
- Salience of Identity: The effect was robust but required the speaker's characteristics to be salient (e.g., through photos or specific cover stories) to trigger the perspective-taking mechanism.
Key Results
- Higher Pragmatic Ratings for Adults: In Experiment 1, the average probability assigned to the "correct" pragmatic target was 70.8% for adult speakers vs. 57.3% for child speakers.
- Replication in Experiment 2: The effect persisted with 70.7% (Adult) vs. 62.2% (Child). While the gap narrowed slightly without the "alien" confound, the statistical evidence for the speaker effect remained strong.
- Qualitative Evidence: Strategy annotations showed that participants explicitly mentioned doubting a child's ability to "think logically" or "make complex decisions" as a reason for their lower target ratings.
Limitations and Future Work
The authors acknowledge that the "inexpressible features" (e.g., telling participants they couldn't use the word "square") is an artificial constraint of the reference game. In natural language, we rarely have such hard limits on vocabulary.
Future research could explore:
- Bottom-up learning: Can listeners update their model of a speaker's rationality just by observing their behavior, without being told their age?
- Naturalistic Stimuli: Using audio or video of children and adults to see if paralinguistic cues (like tone of voice) further modulate these pragmatic inferences.
- Other Identities: Testing effects of language proficiency (non-native speakers) or clinical conditions (e.g., speakers with impairments) on pragmatic interpretation.
Summary updated