Language presents a classic "chicken and egg" problem. To understand what someone means in a specific context—for instance, knowing that "I ate some of the cookies" usually implies "I didn't eat all of them"—you need to know the stable, literal meanings of the words. However, to learn those literal meanings in the first place, you often have to use contextual clues to figure out what a speaker is referring to.
Historically, researchers have modeled these two processes separately. Pragmatics explores how proficient users go beyond literal meaning using recursive reasoning, while word learning focuses on how learners map sounds to objects through statistical co-occurrence. In this paper, the authors argue that these processes are deeply intertwined. A learner’s pragmatic inference about a speaker's goal is exactly what allows them to update their internal dictionary (lexicon).
The Conventionality Assumption
The authors propose a computational model that unifies how we use language in context with how we learn word meanings. It builds on the Rational Speech Act (RSA) framework, which uses recursive Bayesian inference involving three levels of agents:
- $L_0$ (Literal Listener): Uses a basic lexicon to map words to objects.
- $S_1$ (Pragmatic Speaker): Chooses words to maximize the probability that $L_0$ picks the right object.
- $L_2$ (Pragmatic Listener): Uses Bayes' rule to "invert" $S_1$, figuring out the object based on what a rational speaker would say.
The core innovation of this paper is the Conventionality Assumption. The agents assume there is one "true," external lexicon that everyone should be using. Even if an agent is personally uncertain about a word's meaning, they assume their partner knows the "correct" version. This "social anxiety"—the drive to match a perceived external standard—prevents the model from collapsing into incoherence and provides a stable target for learning.
The authors refer to this as the Lexicographer’s Illusion: humans tend to believe words have fixed, "correct" meanings, even if we can't perfectly define them. This bias is actually a functional feature that allows us to converge on a shared language.
Rapid Convergence in Naming Games
Instead of traditional large-scale datasets, the authors used computational simulations designed to replicate well-documented empirical phenomena from developmental psychology and linguistics. They simulated "naming games" where two agents must communicate about objects using a limited set of words, updating their beliefs about the lexicon after each interaction.
In simulated 2x2 and 3x3 naming games, agents starting with zero knowledge of each other's lexicons reached high communicative success (near 1.0 probability of understanding) within just 10 to 20 dialogue turns. The model naturally favors "good" communicative systems—those that are sparse (one word per object) and low-cost (short words for common things).
The Lexicalization of Shortcuts
The model successfully replicates human inferences, such as scalar implicatures (strengthening "some" to "not all") and mutual exclusivity in children (assuming a new word refers to a novel object rather than a familiar one). Crucially, it shows how mutual exclusivity can happen via pragmatic reasoning before the child actually commits the new word to their permanent memory.
Furthermore, the model explains how linguistic shortcuts eventually become permanent parts of a language. For example, "Horn implicatures" involve using a long, costly phrase to describe an unusual event. Over time, the model demonstrates how these costly words for rare objects become baked into the literal lexicon. What starts as a clever, context-dependent inference becomes a standard definition.
Unlike previous recursive models, this framework allows a learner to observe "pragmatically strengthened" uses of words without getting confused about the word's actual literal meaning, maintaining stability in the learning process.
Limitations and Future Directions
While the model provides a robust unified theory, it has limitations:
- Complexity Scaling: The simulations used very small worlds (2-3 words/objects). Scaling this to a full human vocabulary involves significant computational hurdles in marginalizing over the space of all possible lexicons.
- Fixed Recursion: The model uses a fixed recursion depth ($n=3$). In reality, human reasoning depth might vary based on the task or the individual.
- Prior Knowledge: The model assumes agents share a prior understanding of object frequencies and word costs. Future work could explore how these priors are learned alongside the lexicon.
- Social Dynamics: The current model assumes a cooperative environment; it doesn't account for deceptive or adversarial communication.
Ultimately, by assuming that agents reason recursively about each other’s goals while striving to match a conventional shared language, this model elegantly bridges the gap between how we learn words and how we use them.
Summary updated