Compositionality is a key design feature of human language: the meaning of complex expressions is, for the most part, systematically constructed from the meanings of its parts and their manner of composition. This paper demonstrates that rudimentary forms of compositional communicative behavior can emerge from a variant of reinforcement learning applied to signaling games. This helps explain how compositionality could have emerged gradually: if unsophisticated agents can evolve prevalent dispositions to communicate compositional-like, there is a direct evolutionary benefit for adaptations (...) that exploit the systematicity in form-meaning mappings more rigorously. (shrink)

We reopen the investigation into the formal and conceptual relationship between bidirectional optimality theory (Blutner in J Semant 15(2):115–162, 1998 , J Semant 17(3):189–216, 2000 ) and game theory. Unlike a likeminded previous endeavor by Dekker and van Rooij (J Semant 17:217–242, 2000 ), we consider signaling games not strategic games, and seek to ground bidirectional optimization once in a model of rational step-by-step reasoning and once in a model of reinforcement learning. We give sufficient conditions for equivalence of bidirectional (...) optimality and the former, and show based on numerical simulations that bidirectional optimization may be thought of as a process of reinforcement learning with lateral inhibition. (shrink)

Signalling games are popular models for studying the evolution of meaning, but typical approaches do not incorporate vagueness as a feature of successful signalling. Complementing recent like-minded models, we describe an aggregate population-level dynamic that describes a process of imitation of successful behaviour under imprecise perception and realization of similar stimuli. Applying this new dynamic to a generalization of Lewis’s signalling games, we show that stochastic imprecision leads to vague, yet by-and-large efficient signal use, and, moreover, that it unifies evolutionary (...) outcomes and helps avoid sub-optimal categorization. The upshot of this is that we see ‘as-if’-generalization at an aggregate level, without agents actually generalizing. _1_ Introduction _2_ Background _2.1_ Sim-max games and conceptual spaces _2.2_ Vagueness in sim-max games and conceptual spaces _2.3_ Vagueness, functional pressure, and transmission biases _3_ Imprecise Imitation _3.1_ Replicator dynamic in behavioural strategies _3.2_ Noise-perturbed conditional imitation _4_ Exploring Imprecise Imitation _4.1_ Setting the stage _4.2_ Simulation set-up _4.3_ Measures of interest _4.4_ Results _5_ Discussion _5.1_ Levels of vagueness _5.2_ Evolutionary benefits of imprecision _5.3_ Related work _6_ Conclusion Appendix. (shrink)

There is substantial support for the general idea that a formalization of comprehenders' expectations about the likely next word in a sentence helps explaining data related to online sentence processing. While much research has focused on syntactic, semantic, and discourse expectations, the present event‐related potentials (ERPs) study investigates neurolinguistic correlates of pragmatic expectations, which arise when comprehenders expect a sentence to conform to Gricean Maxims of Conversation. For predicting brain responses associated with pragmatic processing, we introduce a formal model of (...) such Gricean pragmatic expectations, using an idealized incremental interpreter. We examine whether pragmatic expectancies derived from this model modulate the amplitude of the N400, a component that has been associated with predictive processing. As part of its parameterization, the model distinguishes genuine pragmatic interpreters, who expect maximally informative true utterances, from literal interpreters, who only expect truthfulness. We explore the model's non‐trivial predictions for an experimental setup which uses picture‐sentence verification with ERPs recorded at several critical positions in sentences containing the scalar implicature trigger some. We find that Gricean expectations indeed affect the N400, largely in line with the predictions of our model, but also discuss discrepancies between model predictions and observations critically. (shrink)

Franke, Dulcinati and Pouscoulous also examine a form of covert lying, by considering to what extent speakers use implicatures to deceive their addressee. The participants in their online signaling game had to describe a card, which a virtual coplayer then had to select. When the goal was to deceive rather than help the coplayer, participants produced more false descriptions (overt lies), but also more uninformative descriptions (covert lies by means of an implicature). [73].

Game theoretic pragmatics is a small but growing part of formal pragmatics, the linguistic subfield studying language use. The general logic of a game theoretic explanation of a pragmatic phenomenon is this: the conversational context is modelled as a game between speaker and hearer; an adequate solution concept then selects the to‐be‐explained behavior in the game model. For such an explanation to be convincing, both components, game model and solution concept, should be formulated and scrutinized as explicitly as possible. The (...) article demonstrates this by a concise overview of both evolutionary and non‐evolutionary approaches to game theoretic pragmatics, arguing for the use of agent‐based micro‐dynamics within evolutionary, and for the use of epistemic game theory within non‐evolutionary approaches. (shrink)

According to standard linguistic theory, the meaning of an utterance is the product of conventional semantic meaning and general pragmatic rules on language use. We investigate how such a division of labor between semantics and pragmatics could evolve under general processes of selection and learning. We present a game‐theoretic model of the competition between types of language users, each endowed with certain lexical representations and a particular pragmatic disposition to act on them. Our model traces two evolutionary forces and their (...) interaction: (i) pressure toward communicative efficiency and (ii) transmission perturbations during the acquisition of linguistic knowledge. We illustrate the model based on a case study on scalar implicatures, which suggests that the relationship between underspecified semantics and pragmatic inference is one of coevolution. (shrink)

Language use and interpretation is heavily contingent on context. But human interlocutors need not always agree what the actual context is. In game theoretic approaches to language use and interpretation, interlocutors’ beliefs about the context are the players’ beliefs about the game that they are playing. Together this entails that we need to consider cases in which interlocutors have different subjective conceptualizations of the game they are in. This paper therefore extends iterated best response reasoning, as an established model for (...) pragmatic reasoning, to games with unawareness. This extension not only leads to more plausible context models for many communicative situations, but also to improved predictions for otherwise problematic cases and an extension of the scope of pragmatic phenomena that can be captured by game theoretic analysis. (shrink)

In a recent contribution in this journal, Sascia Pavan proposed a new game theoretic approach to explain generalized conversational implicatures in terms of general principles of rational behavior. His approach is based on refining Nash equilibrium by a procedure called iterated admissibility. I would like to strengthen Pavan’s case by sketching an epistemic interpretation of iterated admissibility, so as to further our understanding of why iterated admissibility might be a good approximation of pragmatic reasoning. But the explicit epistemic view taken (...) here also points to some shortcomings of his approach in comparison to rivaling accounts. (shrink)

A vexing puzzle about vagueness, rationality, and evolution runs, in crude abbreviation, as follows: vague language use is demonstrably suboptimal if the goal is efficient, precise and cooperative information transmission; hence rational deliberation or evolutionary selection should, under this assumed goal, eradicate vagueness from language use. Since vagueness is pervasive and entrenched in all human languages, something has to give. In this paper, we focus on this problem in the context of signaling games. We provide an overview of a number (...) of proposed ways in which vagueness may come into the picture in formal models of rational and evolutionary signaling. Most argue that vague signal use is simply the best we can get, given certain factors. Despite the plausibility of the proposals, we argue that a deeper understanding of the benefits of vagueness needs a more ecological perspective, namely one that goes beyond the local optimization of signaling strategies in a homogeneous population. As an example of one possible way to expand upon our current models, we propose two variants of a novel multi-population dynamic of imprecise imitation where, under certain conditions, populations with vague language use dominate over populations with precise language use. (shrink)

Intonation plays an integral role in comprehending spoken language. Listeners can rapidly integrate intonational information to predictively map a given pitch accent onto the speaker's likely referential intentions. We use mouse tracking to investigate two questions: (a) how listeners draw predictive inferences based on information from intonation? and (b) how listeners adapt their online interpretation of intonational cues when these are reliable or unreliable? We formulate a novel Bayesian model of rational predictive cue integration and explore predictions derived under a (...) concrete linking hypothesis relating a quantitative notion of evidential strength of a cue to the moment in time, relative to the unfolding speech signal, at which mouse trajectories turn towards the eventually selected option. In order to capture rational belief updates after concrete observations of a speaker's behavior, we formulate and explore an extension of this model that includes the listener's hierarchical beliefs about the speaker's likely production behavior. Our results are compatible with the assumption that listeners rapidly and rationally integrate all available intonational information, that they expect reliable intonational information initially, and that they adapt these initial expectations gradually during exposition to unreliable input. All materials, data, and scripts can be retrieved here: https://osf.io/dnbuk/. (shrink)

The language we use over the course of conversation changes as we establish common ground and learn what our partner finds meaningful. Here we draw upon recent advances in natural language processing to provide a finer‐grained characterization of the dynamics of this learning process. We release an open corpus (>15,000 utterances) of extended dyadic interactions in a classic repeated reference game task where pairs of participants had to coordinate on how to refer to initially difficult‐to‐describe tangram stimuli. We find that (...) different pairs discover a wide variety of idiosyncratic but efficient and stable solutions to the problem of reference. Furthermore, these conventions are shaped by the communicative context: words that are more discriminative in the initial context (i.e., that are used for one target more than others) are more likely to persist through the final repetition. Finally, we find systematic structure in how a speaker's referring expressions become more efficient over time: Syntactic units drop out in clusters following positive feedback from the listener, eventually leaving short labels containing open‐class parts of speech. These findings provide a higher resolution look at the quantitative dynamics of ad hoc convention formation and support further development of computational models of learning in communication. (shrink)

Computational models have shown that purely statistical knowledge about words’ linguistic contexts is sufficient to learn many properties of words, including syntactic and semantic category. For example, models can infer that “postman” and “mailman” are semantically similar because they have quantitatively similar patterns of association with other words. In contrast to these computational results, artificial language learning experiments suggest that distributional statistics alone do not facilitate learning of linguistic categories. However, experiments in this paradigm expose participants to entirely novel words, (...) whereas real language learners encounter input that contains some known words that are semantically organized. In three experiments, we show that the presence of familiar semantic reference points facilitates distributional learning and this effect crucially depends both on the presence of known words and the adherence of these known words to some semantic organization. (shrink)

Previous accounts of how people develop expertise have focused on how deliberate practice transforms the cognitive and perceptual representations and processes that give rise to expertise. However, the likelihood of developing expertise with a particular tool may also depend on the degree to which that tool fits pre‐existing perceptual and cognitive abilities. The present studies explored whether the abacus—a descendent of the first human computing devices—may have evolved to exploit general biases in human visual attention, or whether developing expertise with (...) the abacus requires learning special strategies for allocating visual attention to the abacus. To address this question, we administered a series of visual search tasks to abacus experts and subjects who had little to no abacus experience, in which search targets and distractors were overlaid atop abacus “beads.” Across three studies, we found that both experts and naïve subjects were faster to detect targets in semantically relevant components of the abacus, suggesting that abacus training is not required to exhibit attentional biases toward these components of the abacus. This finding suggests that the attentional biases that scaffold numerical processing of the abacus may emerge from general properties of visual attention that are exploited by the design of the abacus itself. (shrink)

Although the language we encounter is typically embedded in rich discourse contexts, many existing models of processing focus largely on phenomena that occur sentence-internally. Similarly, most work on children's language learning does not consider how information can accumulate as a discourse progresses. Research in pragmatics, however, points to ways in which each subsequent utterance provides new opportunities for listeners to infer speaker meaning. Such inferences allow the listener to build up a representation of the speakers' intended topic and more generally (...) to identify relationships, structures, and messages that extend across multiple utterances. We address this issue by analyzing a video corpus of child–caregiver interactions. We use topic continuity as an index of discourse structure, examining how caregivers introduce and discuss objects across utterances. For the analysis, utterances are grouped into topical discourse sequences using three annotation strategies: raw annotations of speakers' referents, the output of a model that groups utterances based on those annotations, and the judgments of human coders. We analyze how the lexical, syntactic, and social properties of caregiver–child interaction change over the course of a sequence of topically related utterances. Our findings suggest that many cues used to signal topicality in adult discourse are also available in child-directed speech. (shrink)

Children tend to produce words earlier when they are connected to a variety of other words along the phonological and semantic dimensions. Though these semantic and phonological connectivity effects have been extensively documented, little is known about their underlying developmental mechanism. One possibility is that learning is driven by lexical network growth where highly connected words in the child's early lexicon enable learning of similar words. Another possibility is that learning is driven by highly connected words in the external learning (...) environment, instead of highly connected words in the early internal lexicon. The present study tests both scenarios systematically in both the phonological and semantic domains across 10 languages. We show that phonological and semantic connectivity in the learning environment drives growth in both production‐ and comprehension‐based vocabularies, even controlling for word frequency and length. This pattern of findings suggests a word learning process where children harness their statistical learning abilities to detect and learn highly connected words in the learning environment. (shrink)

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. (shrink)

What makes a word easy to learn? Early‐learned words are frequent and tend to name concrete referents. But words typically do not occur in isolation. Some words are predictable from their contexts; others are less so. Here, we investigate whether predictability relates to when children start producing different words (age of acquisition; AoA). We operationalized predictability in terms of a word's surprisal in child‐directed speech, computed using n‐gram and long‐short‐term‐memory (LSTM) language models. Predictability derived from LSTMs was generally a better (...) predictor than predictability derived from n‐gram models. Across five languages, average surprisal was positively correlated with the AoA of predicates and function words but not nouns. Controlling for concreteness and word frequency, more predictable predicates and function words were learned earlier. Differences in predictability between languages were associated with cross‐linguistic differences in AoA: the same word (when it was a predicate) was produced earlier in languages where the word was more predictable. (shrink)

The linguistic input children receive across early childhood plays a crucial role in shaping their knowledge about the world. To study this input, researchers have begun applying distributional semantic models to large corpora of child‐directed speech, extracting various patterns of word use/co‐occurrence. Previous work using these models has not measured how these patterns may change throughout development, however. In this work, we leverage natural language processing methods—originally developed to study historical language change—to compare caregivers' use of words when talking to (...) younger versus older children. Some words' usage changed more than others; this variability could be predicted based on the word's properties at both the individual and category levels. These findings suggest that caregivers' changing patterns of word use may play a role in scaffolding children's acquisition of conceptual structure in early development. (shrink)

Standard applications of evolutionary game theory look at a single game and focus on the evolution of behavior for that game alone. Instead, this article uses tools from evolutionary game theory to study the competition between choice mechanisms in a rich and variable multigame environment. A choice mechanism is a way of subjectively representing a decision situation, paired with a method for choosing an act based on this subjective representation. We demonstrate the usefulness of this approach by a case study (...) that shows how subjective representations in terms of regret that differ from the actual fitness can be evolutionarily advantageous. (shrink)