Rational Speech Acts Research Articles

The pragmatics of free choice <em>any</em>

In this paper, I examine previous claims by Menéndez-Benito (2005; 2010) and Dayal (2013) that You may read any book entails Every book may be read on its own. In light of several key counterexamples, including those raised by Chierchia (2013) and Szabolcsi (2019), I argue that this is not an entailment at all, but a particularly robust implicature that arises in the pragmatics. In order to derive this implicature, I combine Szabolcsi’s semantics for universal free choice items (FCIs) with a formal pragmatic model in the Rational Speech Act (RSA) framework. I update Champollion et al.’s (2019) RSA model of free choice with technical innovations from Franke & Bergen (2020) to model how a listener interprets an utterance containing the universal FCI any given a range of possible exhaustified parses. This pragmatic model predicts the robustness of the implicature observed by Menéndez-Benito and Dayal, and further predicts that it is more stable across varied contexts than other implicatures arising from the same utterance.

Using contrastive inferences to learn about new words and categories

In the face of unfamiliar language or objects, description is one cue people can use to learn about both. Beyond narrowing potential referents to those that match a descriptor (e.g., “tall”), people could infer that a described object is one that contrasts with other relevant objects of the same type (e.g., “the tall cup” contrasts with another, shorter cup). This contrast may be in relation to other objects present in the environment (this cup is tall among present cups) or to the referent’s category (this cup is tall for a cup in general). In three experiments, we investigate whether people use such contrastive inferences from description to learn new word–referent mappings and learn about new categories’ feature distributions. People use contrastive inferences to guide their referent choice, though size – and not color – adjectives prompt them to consistently choose the contrastive target over alternatives (Experiment 1). People also use color and size description to infer that a novel object is atypical of its category (Experiments 2 and 3): utterances like “the blue toma” prompt people to infer that tomas are less likely to be blue in general. However, these two inferences do not trade off substantially: people infer a described referent is atypical even when the descriptor was necessary to establish reference. We model these experiments in the Rational Speech Act (RSA) framework and find that it predicts both of these inferences. Overall, people are able to use contrastive inferences from description to resolve reference and make inferences about a novel object’s category, letting them learn more about new things than literal meaning alone allows.

She adapts to her student: An expert pragmatic speaker tailoring her referring expressions to the Layman listener.

Communication is a dynamic process through which interlocutors adapt to each other. In the development of conversational agents, this core aspect has been put aside for several years since the main challenge was to obtain conversational neural models able to produce utterances and dialogues that at least at the surface level are human-like. Now that this milestone has been achieved, the importance of paying attention to the dynamic and adaptive interactive aspects of language has been advocated in several position papers. In this paper, we focus on how a Speaker adapts to an interlocutor with different background knowledge. Our models undergo a pre-training phase, through which they acquire grounded knowledge by learning to describe an image, and an adaptive phase through which a Speaker and a Listener play a repeated reference game. Using a similar setting, previous studies focus on how conversational models create new conventions; we are interested, instead, in studying whether the Speaker learns from the Listener's mistakes to adapt to his background knowledge. We evaluate models based on Rational Speech Act (RSA), a likelihood loss, and a combination of the two. We show that RSA could indeed work as a backbone to drive the Speaker toward the Listener: in the combined model, apart from the improved Listener's accuracy, the language generated by the Speaker features the changes that signal adaptation to the Listener's background knowledge. Specifically, captions to unknown object categories contain more adjectives and less direct reference to the unknown objects.

The Rational Speech Act Framework

The past decade has seen the rapid development of a new approach to pragmatics that attempts to integrate insights from formal and experimental semantics and pragmatics, psycholinguistics, and computational cognitive science in the study of meaning: probabilistic pragmatics. The most influential probabilistic approach to pragmatics is the Rational Speech Act (RSA) framework. In this review, I demonstrate the basic mechanics and commitments of RSA as well as some of its standard extensions, highlighting the key features that have led to its success in accounting for a wide variety of pragmatic phenomena. Fundamentally, it treats language as probabilistic, informativeness as gradient, alternatives as context-dependent, and subjective prior beliefs (world knowledge) as a crucial facet of interpretation. It also provides an integrated account of the link between production and interpretation. I highlight key challenges for RSA, which include scalability, the treatment of the boundedness of cognition, and the incremental and compositional nature of language.

Are Narrow Focus Exhaustivity Inferences Bayesian Inferences?

In successful communication, the literal meaning of linguistic utterances is often enriched by pragmatic inferences. Part of the pragmatic reasoning underlying such inferences has been successfully modeled as Bayesian goal recognition in the Rational Speech Act (RSA) framework. In this paper, we try to model the interpretation of question-answer sequences with narrow focus in the answer in the RSA framework, thereby exploring the effects of domain size and prior probabilities on interpretation. Should narrow focus exhaustivity inferences be actually based on Bayesian inference involving prior probabilities of states, RSA models should predict a dependency of exhaustivity on these factors. We present experimental data that suggest that interlocutors do not act according to the predictions of the RSA model and that exhaustivity is in fact approximately constant across different domain sizes and priors. The results constitute a conceptual challenge for Bayesian accounts of the underlying pragmatic inferences.

Numeral terms and the predictive potential of Bayesian updating

AbstractIn the experiment described in the paper Noah Goodman & Andreas Stuhlmüller. 2013. Knowledge and im-plicature: Modeling language understanding as social cognition.Topics in Cognitive Science5(1). 173–184, empirical support was provided for the predictive power of the Rational Speech Act (RSA) model concerning the interpretation of utterances employing numerals in uncertainty contexts. The RSA predicts a Bayesian interdependence between beliefs about the probability distribution of the occurrence of an event prior to receiving information and the updated probability distribution after receiving information. In this paper we analyze whether the RSA is a descriptive or a normative model. We present the results of two analogous experiments carried out in Polish. The first experiment does not replicate the original empirical results. We find that this is due to different answers on the prior probability distribution. However, the model predicts the different results on the basis of different collected priors: Bayesian updating predicts human reasoning. By contrast, the second experiment, where the answers on the prior probability distribution are as predicted, is a replication of the original results. In light of these results we conclude that the RSA is a robust, descriptive model, however, the experimental assumptions pertaining to the experimental setting adopted by Goodman and Stuhlmüller are normative.

Limits to the Rational Production of Discourse Connectives.

Rational accounts of language use such as the uniform information density hypothesis, which asserts that speakers distribute information uniformly across their utterances, and the rational speech act (RSA) model, which suggests that speakers optimize the formulation of their message by reasoning about what the comprehender would understand, have been hypothesized to account for a wide range of language use phenomena. We here specifically focus on the production of discourse connectives. While there is some prior work indicating that discourse connective production may be governed by RSA, that work uses a strongly gamified experimental setting. In this study, we aim to explore whether speakers reason about the interpretation of their conversational partner also in more realistic settings. We thereby systematically vary the task setup to tease apart effects of task instructions and effects of the speaker explicitly seeing the interpretation alternatives for the listener. Our results show that the RSA-predicted effect of connective choice based on reasoning about the listener is only found in the original setting where explicit interpretation alternatives of the listener are available for the speaker. The effect disappears when the speaker has to reason about listener interpretations. We furthermore find that rational effects are amplified by the gamified task setting, indicating that meta-reasoning about the specific task may play an important role and potentially limit the generalizability of the found effects to more naturalistic every-day language use.

Corpus, experimental and modeling investigations of cross-linguistic differences in pronoun resolution preferences

We investigate the impact of syntactic alternatives on pronoun resolution in ambiguous constructions in English and French. Previous research detected language-specific preferences in pronoun resolution in utterances of the type “The postman met the streetsweeper before he went home”. These preferences have been attributed to the interaction of information structural and syntactic constraints inducing a subject bias on the one hand, and Gricean reasoning processes taking into account alternative syntactic constructions on the other hand. A corpus study of four English and French corpora shows that an alternative construction which takes a subject antecedent (“The postman met the streetsweeper before going home”) is much less frequent in spoken English than French. A Rational Speech Act (RSA) model with corpus frequencies integrated as language-specific costs on the use of each construction makes empirical predictions for pronoun resolution preferences in French and English for sentences with “avant”/“before” which have been tested before but also for sentences with “après”/“after” which have not been tested so far. New experimental data show a very good fit of the model predictions for pronoun resolution preferences in English as well as for the differences in antecedent choices between French and English. However, experimental data showing differences in antecedent choices between French sentences with “après” and “avant” deviate from model predictions, indicating that more factors need to be taken into account. The combination of Bayesian modeling, corpus analyses and experimental data shows that RSA models can make relevant and falsifiable predictions for cross-linguistic variation in processing.

Reevaluating pragmatic reasoning in language games.

The results of a highly influential study that tested the predictions of the Rational Speech Act (RSA) model suggest that (a) listeners use pragmatic reasoning in one-shot web-based referential communication games despite the artificial, highly constrained, and minimally interactive nature of the task, and (b) that RSA accurately captures this behavior. In this work, we reevaluate the contribution of the pragmatic reasoning formalized by RSA in explaining listener behavior by comparing RSA to a baseline literal listener model that is only driven by literal word meaning and the prior probability of referring to an object. Across three experiments we observe only modest evidence of pragmatic behavior in one-shot web-based language games, and only under very limited circumstances. We find that although RSA provides a strong fit to listener responses, it does not perform better than the baseline literal listener model. Our results suggest that while participants playing the role of the Speaker are informative in these one-shot web-based reference games, participants playing the role of the Listener only rarely take this Speaker behavior into account to reason about the intended referent. In addition, we show that RSA’s fit is primarily due to a combination of non-pragmatic factors, perhaps the most surprising of which is that in the majority of conditions that are amenable to pragmatic reasoning, RSA (accurately) predicts that listeners will behave non-pragmatically. This leads us to conclude that RSA’s strong overall correlation with human behavior in one-shot web-based language games does not reflect listener’s pragmatic reasoning about informative speakers.

A probabilistic pragmatics for English singular <em>some</em>

When English some combines with a singular NP, the resulting phrases reliably convey affective meanings not shared by variants with a(n) or plural NPs. Prior research has traced these effects to semantic properties of some that entail that the speaker cannot identify a unique referent for the phrase. In this paper, we present attested examples that conflict with this generalization. In addition, we argue that semantic accounts miss an important generalization: some is reliably affective only if a is available as an alternative. These facts suggest a pragmatic source for the relevant meanings. To capture them, we argue that a given context can make different modes of identification for entities relevant and that singular some signals a lack of engagement with these modes. We analyze the pragmatics of this signaling using the "lexical uncertainty" version of the Rational Speech Acts model and show how it can be used to characterize the observed affective meanings.

Free choice disjunction as a rational speech act

The so-called free choice inference (from You may take an apple or a pear to You may take an apple) is mysterious because it does not follow from ordinary modal logic. We show that this inference arises in the Rational Speech Act framework (Frank & Goodman 2012). Our basic idea is inspired by exhaustification-based models of free choice (Fox 2007) and by game-theoretic accounts based on iterated best response (Franke 2011). We assume that when the speaker utters You may take an apple or a pear, the hearer reasons about why the speaker did not choose alternative utterances such as You may take an apple. A crucial ingredient in our explanation is the idea of semantic uncertainty (Bergen, Levy & Goodman 2016). Specifically, we assume that the speaker is uncertain whether or not the hearer will interpret You may take an apple as forbidding them from taking a pear. This uncertainty can be thought of as resulting from Fox’s (2007) optional covert exhaustification. Uttering the disjunction is a way for the speaker to prevent the hearer from concluding that any fruit is forbidden to take. Knowing this, the hearer concludes that they may choose either fruit.

A Minimal Turing Test

We introduce the Minimal Turing Test, an experimental paradigm for studying perceptions and meta-perceptions of different social groups or kinds of agents, in which participants must use a single word to convince a judge of their identity. We illustrate the paradigm by having participants act as contestants or judges in a Minimal Turing Test in which contestants must convince a judge they are a human, rather than an artificial intelligence. We embed the production data from such a large-scale Minimal Turing Test in a semantic vector space, and construct an ordering over pairwise evaluations from judges. This allows us to identify the semantic structure in the words that people give, and to obtain quantitative measures of the importance that people place on different attributes. Ratings from independent coders of the production data provide additional evidence for the agency and experience dimensions discovered in previous work on mind perception. We use the theory of Rational Speech Acts as a framework for interpreting the behavior of contestants and judges in the Minimal Turing Test.

A Minimal Parsimonious Model of Pragmatic Comprehension

Computational accounts of how we infer what a speaker means by what she says have recently spurred significant progress in our understanding of linguistic pragmatics. These models, motivated by general assumptions about human cognition, have been shown to account for a wide range of phenomena in pragmatic comprehension. Despite their successes, these models incorporate a number of structures and assumptions that diminish their parsimony as well as their consistency with the complex cognitive features of human communication. In this paper, I demonstrate these weaknesses through a number of empirical test cases and theoretical arguments. I then propose an alternative computational modeling approach, derived from the popular Rational Speech Act (RSA) theory, that makes fewer and simpler assumptions about pragmatic comprehension and accounts for a wide range of pragmatic phenomena without requiring ad hoc modifications or theoretically unmotivated assumptions. This Speaker Centric Inference Model of Pragmatic Interpretation (SCIMPI) not only provides the basis for a simple and parsimonious account of human pragmatic inference, but suggests that formal linguistic pragmatics should be practiced by incrementally constructing theory through the introduction of minimal and motivated components.

A Psycholinguistic Model for the Marking of Discourse Relations

Discourse relations can either be explicitly marked by discourse connectives (DCs), such as therefore and but, or implicitly conveyed in natural language utterances. How speakers choose between the two options is a question that is not well understood. In this study, we propose a psycholinguistic model that predicts whether or not speakers will produce an explicit marker given the discourse relation they wish to express. Our model is based on two information-theoretic frameworks: (1) the Rational Speech Acts model, which models the pragmatic interaction between language production and interpretation by Bayesian inference, and (2) the Uniform Information Density theory, which advocates that speakers adjust linguistic redundancy to maintain a uniform rate of information transmission. Specifically, our model quantifies the utility of using or omitting a DC based on the expected surprisal of comprehension, cost of production, and availability of other signals in the rest of the utterance. Experiments based on the Penn Discourse Treebank show that our approach outperforms the state-of-the-art performance at predicting the presence of DCs (Patterson and Kehler, 2013), in addition to giving an explanatory account of the speaker’s choice.

Pragmatic reasoning through semantic inference

A number of recent proposals have used techniques from game theory and Bayesian cognitive science to formalize Gricean pragmatic reasoning (Frank & Goodman, 2012; Franke, 2009; Goodman & Stuhlmuller, 2013; Jager, 2012). We discuss two phenomena which pose a challenge to these accounts of pragmatics: M-implicatures (Horn, 1984) and embedded implicatures which violate Hurford’s constraint (Chierchia, Fox, & Spector, 2012; Hurford, 1974). While techniques have been developed for deriving M-implicatures, Hurford-violating em- bedded implicatures pose a more fundamental challenge, because of basic limitations in the models’ architecture. In order to explain these phenomena, we propose a realignment of the division between semantic content and pragmatic content. Under this proposal, the semantic content of an utterance is not fixed independent of pragmatic inference; rather, pragmatic inference partially determines an utterance’s semantic content. We show how semantic inference can be realized as an extension to the Rational Speech Acts framework (Goodman & Stuhlmuller, 2013). The addition of lexical uncertainty derives both M-implicatures and the relevant embedded implicatures, and preserves the derivations of more standard implicatures. We use this principle to explain a novel class of implicature, non-convex disjunctive implicatures, which have several theoretically interesting properties. In particular, these implicatures can be preserved in downward-entailing contexts in the absence of accenting, a property which is predicted by lexical uncertainty, but which violates prior generalizations in the literature (Fox & Spector, in press; Horn, 1989). EARLY ACCESS VERSION