Search

Studies of meaning in human and primate communication face, in principle, similar methodological problems. In both cases, meaning is not observable directly, but must be inferred from more indirect sources, such as directly observable behavior. Recent work in probabilistic cognitive modeling of language use has therefore developed methods of inferring latent se- mantic meaning through the lens of a probabilistic model of language use. In this paper, we explore how to adapt such an approach for insightful investigations of primate communication. Towards this end, we develop a suitable probabilistic model of processes that generate communicative behavior by making use of functionally specified latent meaning representations. As a proof of concept, we apply this model to a rich, annotated data set of orangutan communicative dyadic interaction and conclude that explicit probabilistic modeling can provide additional insights for the study of animal communication pertaining to the context-dependent nature of signals and the gradual evolution of human communication systems.

Recent work shows that participants make asymmetric causal inferences from apparently symmetric correlational statements (e.g., “A is associated with B”). Can we make sense of this behavior in terms of rational language use? Experiment 1 investigates these interpretive preferences—what we call “PACE effects”—in light of theoretical and experimental pragmatics and psycholinguistics. We uncover several linguistic factors that influence them, suggesting that a pragmatic explanation is possible. Yet, since PACE effects do not show that correlational language leads to causal implicatures strong enough to influence action choice in practical decision contexts, Experiment 2 offers new evidence from an experiment that explicitly compares the effects of causal vs. correlational claims on decision-making. Our results suggest that causal inferences from correlation language are an intricate, but possibly

While the literature has mostly focused on the goal of information transfer, many linguistic phenomena only make sense in the light of further goals pursued by the agent. One such phenomenon is polite language use. In this paper, we propose a new model of polite language production. We suggest that patterns characteristic of polite language, e.g., indirectness, emerge from a tension between two goals: on the one hand, being sufficiently truthful and informative, and on the other hand, being kind to the listener. To capture these pressures, we introduce a novel model of probabilistic language production which combines a strategic choice of content selection with the usual pragmatic choice of content expression. We fit our model to empirical data from a previous experiment using a bespoke Bayesian model. We quantitatively compare our model to a previous model of politeness and discuss some ways in which our account is simpler, more general and better accounts for empirical data and theoretical considerations.

Conditionals like 'If A, then C' can be used, among others, to convey important knowledge about rules, dependencies and causal relationships. Much work has been devoted to the interpretation of conditional sentences, but much less is known about when speakers choose to use a conditional over another type of utterance in communication. To fill this gap, we consider a recently proposed computational model from probabilistic pragmatics, adapted for modeling the use of conditionals in natural language, by comparing its predictions to experimental production data from a behavioral experiment. In a novel experimental approach, we manipulate relevant causal beliefs that might influence whether utterances with conditional structure are preferred over utterances without conditional structure. This is a step towards a systematic, quantitative investigation of the situations that do or do not elicit the natural use of conditionals.

The study of emergent languages in deep multi-agent simulationshas become an important research field. While targetingdifferent objectives, most studies focus on analyzing propertiesof the emergent language—often in relation to the agents’inputs—ignoring the influence of the agents’ perceptual processes.In this work, we use communication games to investigatehow differences in perception affect emergent language.Using a conventional setup, we train two deep reinforcementlearning agents, a sender and a receiver, on a reference game.However, we systematically manipulate the agents’ perceptionby enforcing similar representations for objects with specificshared features. We find that perceptual biases of both senderand receiver influence which object features the agents’ messagesare grounded in. When uniformly enforcing the similarityof all features that are relevant for the reference game,agents perform better and the emergent protocol better capturesconceptual input properties.

We propose an extension to probabilistic pragmatic models to include a dimension that allows for the modeling of argu-mentative language use. Within our extended Rational Speech Act model, argumentative strength stands for a statisticalmeasure of observational evidence which impacts a speakers utterance choice. More concretely, our model recasts speakerutility in terms of a weight parameter which varies between being purely informative and purely argumentative. We fitthe extended RSA model to empirical data from a novel production experiment. Our initial results suggest that there isroom for argumentativity on top of informativity in formalizations of pragmatic language reasoning. Crucially, we see thatthe relationship between the two is not straightforward, as the model fails to capture instances of human behavior whichare more manipulative than expected by the suggested informativity-argumentativity trade-off. All in all, our explorationprovides us with interesting insights about this relationship.

We present a novel framework for building pragmatic artificialagents with explicit and trainable semantic representations, us-ing the Rational Speech Act model. We train our agents onsupervised and unsupervised communication games and com-pare their behavior to literal agents lacking pragmatic abilities.For both types of games pragmatic but not literal agents evolvea mutual exclusivity bias. This provides a computational prag-matic account of mutual exclusivity and points out a possi-ble direction for solving the mutual exclusivity bias challengeposed by Gandhi and Lake (2019). We find that pragmaticreasoning can cause the bias either by promoting lexical con-straints during learning, or by affecting online inference. In ad-dition we show that pragmatic abilities lead to faster learningand that this advantage is even stronger when meanings to becommunicated follow a more natural distribution as describedby Zipf’s law.

This study investigates how quantifiers are used strategically to serve different argumentative goals. We report two experiments on how English speakers describe the results of school exams when being instructed to frame their descriptions either as a good or bad outcome. Experiment 1 shows that participants have clear preferences for specific quantifier combinations in this task. Experiment 2 shows that, in situations where producing descriptions that meet one's argumentative goals is difficult (i.e. framing very bad outcomes positively), participants tend to use quantifiers that are informationally weaker than other salient alternatives. Experiment 2 also shows that people have a bias to frame outcomes positively, even when the task asks them to frame them negatively. Put together, these results shed light on the question of how language users strategically explore different linguistic strategies to communicate quantity in pragmatically favorable ways, including how quantifiers are used vis-'a-vis other lexical expressions of quantity. [ABSTRACT FROM AUTHOR]

Transmission of information by means of language is a po-tentially lossy process. Especially adjunct information, suchas the graded degree of evidence, is a piece of informationthat seems prima facie likely to be distorted by reproductionnoise. To investigate this issue, we present the results of a two-step iterated narration study: first, we collected a corpus of250 crime story reproductions that were produced in parallelreproduction chains of 5 generations in depth, for 5 differentseed stories; a second separate large-scale experiment then tar-geted readers’ interpretation of these reproductions. Crucially,strength of evidence for the guilt of each story’s suspect(s)was manipulated in the initial seed stories. Across genera-tions, readers’ guilt perceptions decreased when the evidencewas originally strong, but remained stable when evidence wasoriginally weak. Analysis of linguistic measures revealed thatdissimilarity between a seed story and its reproduction, storylength, and amount of hedging language affected the readers’own guilt perception and the readers’ attribution of guilt per-ception to the author differently. The results provide evidencethat evidential information indeed influences guilt perceptionin complex ways.

Adjective ordering preferences (e.g., big brown bag vs. brownbig bag) are robustly attested in English and many unrelatedlanguages (Dixon, 1982). Scontras, Degen, and Goodman(2017) showed that adjective subjectivity is a robust predictorof ordering preferences in English: less subjective adjectivesare preferred closer to the modified noun. In a follow-up tothis empirical finding, Simoniˇc (2018) and Scontras, Degen,and Goodman (to appear) claim that pressures from success-ful reference resolution and the hierarchical structure of mod-ification explain subjectivity-based ordering preferences. Weprovide further support for this claim using large-scale sim-ulations of reference scenarios, together with an empirically-motivated adjective semantics. In the vast majority of cases,subjectivity-based adjective orderings yield a higher probabil-ity of successful reference resolution.

Words exist for referring at many levels of specificity: fromthe broadest (thing) to the most specific (Fido). What drivesthe emergence of these taxonomies of reference? Recent com-putational theories of language evolution suggest that commu-nicative demands of the environment may play a deciding role.Here, we investigate local pragmatic mechanisms of lexicaladaptation that may undergird global emergence by manipulat-ing context in a repeated reference game where pairs of partic-ipants interactively coordinate on an artificial communicationsystem. We hypothesize that pairs should converge on specificnames (e.g. Fido) when the context requires frequently mak-ing fine distinctions between entities; conversely, they shouldconverge on a more compressed system of conventions for ab-stract categories (e.g. dog) in coarser contexts, even if a finermapping would be sufficient. We show differences in the lev-els of abstraction that emerged in different environments andintroduce a statistical approach to probe the dynamics of emer-gence.

Human behavior is often remarkably flexible, showing theability to quickly adapt to the statistical peculiarities of aparticular local context. When it comes to language, previ-ous work has shown that listeners’ anticipatory interpretationsof intonational cues are adapted dynamically when cues areobserved to be stochastically unreliable. This paper reportsnovel empirical data from manual response dynamics (mouse-tracking) on how listeners adapt their predictive interpretationwhen some intonational cues are occasionally unreliable whileothers are consistently reliable. A model of rational belief dy-namics predicts that listeners adapt differently to different un-reliable intonational cues, as a function of their initial eviden-tial strength. These predictions are borne out by our data.

Language provides multiple ways of conveying the opposite:A person not happy can be unhappy, sad, or perhaps neither,just not happy. Rather than being redundant, we hypothesizethat uncertainty about the meaning of negation markers allowslisteners to derive fine-grained distinctions among these vari-ous alternatives. We formalize this hypothesis in a probabilis-tic model of gradable adjectives (e.g., happy), and use this toaddress an outstanding puzzle: how to interpret double nega-tions (e.g., not unhappy). Our model makes surprising addi-tional predictions about a putative difference between morpho-logical antonyms (unhappy) and negated positives (not happy):Listeners should judge unhappy as more sad than not happyonly when confronted with alternatives in context; when inter-preted in isolation, we predict no difference in understanding.Two behavioral experiments confirm consistent orderings ofinterpretations that interact with the presentational context inthe way predicted. These findings support the hypothesis thatlisteners represent uncertainty even about the most logical ele-ments of language.

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.

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/

Complex uncertainty expressions such as probably likely andcertainly possible naturally occur in everyday conversations.However, they received much less attention in the literaturethan simple ones. We propose a probabilistic model of the useand interpretation of complex uncertainty expressions based onthe assumption that their predominant function is to communi-cate factual information about the world, and that further layersof uncertainty are pragmatically inferred. We collected empir-ical data on the use and interpretation of these expressions anduse it for detailed model criticism.

We investigate the influence of the adverb surprisingly on themeaning of the quantity words few and many, which them-selves have been associated with a reading expressing sur-prise. To learn about the meaning contribution of “surprise”,we compare surprisingly with the intensifier incredibly anda compared to phrase explicitly marking surprise. Based onan empirical measure of subjects’ expectations about everydayevents, a Bayesian model uses data from a sentence judgmenttask to infer likely levels of surprise associated with the differ-ent constructions of interest.

Two factors seem to play a major role in the cultural evolutionof language. On the one hand, there is functional pressure to-wards efficient transfer of information. On the other hand, lan-guages have to be learned repeatedly and will therefore showtraces of systematic stochastic disturbances of the transmissionof linguistic knowledge. While a lot of attention has been paidto the effects of cognitive learning biases on the transmissionof language, there is reason to expect that the class of possiblyrelevant transmission perturbations is much larger. This papertherefore explores some potential effects of transmission noisedue to errors in the observation of states of the world. We lookat three case studies on (i) vagueness, (ii) meaning deflation,and (iii) underspecified lexical meaning. These case studiessuggest that transmission perturbations other than learning bi-ases might help explain attested patterns in the cultural evolu-tion of language and that perturbations due to perceptual noisemay even produce effects very similar to learning biases.

Natural languages exhibit properties that are difficult to explainfrom a purely functional perspective. One of these properties isthe systematic lack of upper-bounds in the literal meaning ofscalar expressions. This investigation addresses the develop-ment and selection of such semantics from a space of possiblealternatives. To do so we put forward a model that integratesBayesian learning into the replicator-mutator dynamics com-monly used in evolutionary game theory. We argue this syn-thesis to provide a suitable and general model to analyze thedynamics involved in the use and transmission of language.Our results shed light on the semantics-pragmatics divide andshow how a learning bias in tandem with functional pressuremay prevent the lexicalization of pragmatic inferences.

People’s beliefs about everyday events are both of theoreti-cal interest in their own right and an important ingredient inmodel building—especially in Bayesian cognitive models ofphenomena such as logical reasoning, future predictions, andlanguage use. Here, we explore several recently used methodsfor measuring subjective beliefs about unidimensional contigu-ous properties, such as the likely price of a new watch. Asa first step towards a way of assessing and comparing beliefelicitation methods, we use hierarchical Bayesian modeling forinferring likely population-level beliefs as the central tendencyof participants’ individual-level beliefs. Three different depen-dent measures are considered: (i) slider ratings of (relative)likelihood of intervals of values, (ii) a give-a-number task, and(iii) choice of the more likely of two intervals of values. Ourresults suggest that using averaged normalized slider ratingsfor binned quantities is a practical and fairly good approxima-tor of inferred population-level beliefs.

Possibility and probability expressions, like possibly or prob-ably, are frequently assumed to communicate that the proba-bility of a proposition is above a certain threshold. Most pre-vious empirical research on these expressions has focused oncases of known objective chance: if the true objective proba-bility is given, would a speaker use possibly, probably or oneof their kin? Here, we investigate the use of probability expres-sions when speakers have subjective uncertainty about objec-tive chance, i.e., higher-order uncertainty. Experimental datasuggest that speakers’ choices of a probability expression is acomplex function of their state of higher-order uncertainty. Weformulate a computational probabilistic model of pragmaticspeaker behavior that explains the experimental data.