Mutual Exclusivity in Pragmatic Agents

Abstract

One of the great challenges in word learning is that words are typically uttered in a context with many potential referents. Children's tendency to associate novel words with novel referents, which is taken to reflect a mutual exclusivity (ME) bias, forms a useful disambiguation mechanism. We study semantic learning in pragmatic agents—combining the Rational Speech Act model with gradient‐based learning—and explore the conditions under which such agents show an ME bias. This approach provides a framework for investigating a pragmatic account of the ME bias in humans but also for building artificial agents that display an ME bias. A series of analyses demonstrates striking parallels between our model and human word learning regarding several aspects relevant to the ME bias phenomenon: online inference, long‐term learning, and developmental effects. By testing different implementations, we find that two components, pragmatic online inference and incremental collection of evidence for one‐to‐one correspondences between words and referents, play an important role in modeling the developmental trajectory of the ME bias. Finally, we outline an extension of our model to a deep neural network architecture that can process more naturalistic visual and linguistic input. Until now, in contrast to children, deep neural networks have needed indirect access to (supposed to be novel) test inputs during training to display an ME bias. Our model is the first one to do so without using this manipulation.

The roles of neural networks in language acquisition

“…Models must have some prior heuristic bias or reasoning ability to truly develop a mutual exclusivity bias. Ohmer, König, and Franke (2020) show that by integrating the ability to reason about speakers intentions (M. C. Frank et al, 2009) into neural networks, models will develop over time a true mutual exclusivity bias when presented with novel stimuli. Importantly, they also show using an ablation study that in the absence of the ability for pragmatic reasoning, models do not develop a mutual exclusivity bias.…”

“…They help us see what kind of learning or processing of language is possible. They may focus on very specific linguistic phenomena (Ohmer et al, 2020) or capture a wide range of linguistic behavior (Huebner et al, 2021), without necessarily committing to the internal structure or representations of the model.…”

From partners to populations: A hierarchical Bayesian account of coordination and convention.

“…That is, we use the RSA equations as the linking function in the likelihood P(Dk|ϕk), representing an agent’s prediction about how a partner with meaning function ϕ k would actually behave in context (Equation 5). This use of pragmatic reasoning has been explicitly linked to principles like mutual exclusivity in word learning (Bloom, 2002; Frank et al, 2009; Gulordava et al, 2020; Ohmer et al, 2020; Smith et al, 2013). For example, upon hearing their partner use a particular utterance u to refer to an object o , a pragmatic listener can not only infer that u means o in their partner’s lexicon, but also that other utterances u ′ likely do not mean o : If they did, the speaker would have used them instead.…”

Communication games: Modeling language evolution through dyadic agent interactions

“…The choice of agents for each role, where there is one, is discussed in more detail below -but for some models, the choice of meaning is also important. Implementations may only involve a single meaning (see e.g., the Naming Game, below), choose a meaning from a uniform distribution at each interaction (e.g., Franke, 2016), or choose a meaning from, for example, a Zipfian distribution (Cuskley, Loreto, & Kirby, 2018;Ohmer, König, & Franke, 2020). Finally, how S chooses σ (getSignal()), how R interprets it (guessMeaning()), and how agents update their mappings, can involve complex weights and learning algorithms.…”