Language evolution by iterated learning with bayesian agents

Languages are transmitted from person to person and generation to generation via a process of iterated learning: people learn a language from other people who once learned that language themselves. We analyze the consequences of iterated learning for learning algorithms based on the principles of Bayesian inference, assuming that learners compute a posterior distribution over languages by combining a prior (representing their inductive biases) with the evidence provided by linguistic data. We show that when learners sample languages from this posterior distribution, iterated learning converges to a distribution over languages that is determined entirely by the prior. Under these conditions, iterated learning is a form of Gibbs sampling, a widely-used Markov chain Monte Carlo algorithm. The consequences of iterated learning are more complicated when learners choose the language with maximum posterior probability, being affected by both the prior of the learners and the amount of information transmitted between generations. We show that in this case, iterated learning corresponds to another statistical inference algorithm, a variant of the expectation-maximization (EM) algorithm. These results clarify the role of iterated learning in explanations of linguistic universals and provide a formal connection between constraints on language acquisition and the languages that come to be spoken, suggesting that information transmitted via iterated learning will ultimately come to mirror the minds of the learners.     

Using category structures to test iterated learning as a method for identifying inductive biases

Many of the problems studied in cognitive science are inductive problems, requiring people to evaluate hypotheses in the light of data. The key to solving these problems successfully is having the right inductive biases—assumptions about the world that make it possible to choose between hypotheses that are equally consistent with the observed data. This article explores a novel experimental method for identifying the biases that guide human inductive inferences. The idea behind this method is simple: This article uses the responses produced by a participant on one trial to generate the stimuli that either they or another participant will see on the next. A formal analysis of this "iterated learning" procedure, based on the assumption that the learners are Bayesian agents, predicts that it should reveal the inductive biases of these learners, as expressed in a prior probability distribution over hypotheses. This article presents a series of experiments using stimuli based on a well-studied set of category structures, demonstrating that iterated learning can be used to reveal the inductive biases of human learners.     

Language Evolution Can Be Shaped by the Structure of the World

Human languages vary in many ways but also show striking cross‐linguistic universals. Why do these universals exist? Recent theoretical results demonstrate that Bayesian learners transmitting language to each other through iterated learning will converge on a distribution of languages that depends only on their prior biases about language and the quantity of data transmitted at each point; the structure of the world being communicated about plays no role (Griffiths & Kalish, 2005 , 2007 ). We revisit these findings and show that when certain assumptions about the relationship between language and the world are abandoned, learners will converge to languages that depend on the structure of the world as well as their prior biases. These theoretical results are supported with a series of experiments showing that when human learners acquire language through iterated learning, the ultimate structure of those languages is shaped by the structure of the meanings to be communicated.     

When Extremists Win: Cultural Transmission Via Iterated Learning When Populations Are Heterogeneous

How does the process of information transmission affect the cultural or linguistic products that emerge? This question is often studied experimentally and computationally via iterated learning, a procedure in which participants learn from previous participants in a chain. Iterated learning is a powerful tool because, when all participants share the same priors, the stationary distributions of the iterated learning chains reveal those priors. In many situations, however, it is unreasonable to assume that all participants share the same prior beliefs. We present four simulation studies and one experiment demonstrating that when the population of learners is heterogeneous, the behavior of an iterated learning chain can be unpredictable and is often systematically distorted by the learners with the most extreme biases. This results in group‐level outcomes that reflect neither the behavior of any individuals within the population nor the overall population average. We discuss implications for the use of iterated learning as a methodological tool as well as for the processes that might have shaped cultural and linguistic evolution in the real world.     

Individual Differences in Learning Abilities Impact Structure Addition: Better Learners Create More Structured Languages

Over the last decade, iterated learning studies have provided compelling evidence for the claim that linguistic structure can emerge from non-structured input, through the process of transmission. However, it is unclear whether individuals differ in their tendency to add structure, an issue with implications for understanding who are the agents of change. Here, we identify and test two contrasting predictions: The first sees learning as a pre-requisite for structure addition, and predicts a positive correlation between learning accuracy and structure addition, whereas the second maintains that it is those learners who struggle with learning and reproducing their input who add structure to it. This prediction is hard to test in standard iterated learning paradigms since each learner is exposed to a different input, and since structure and accuracy are computed using the same test items. Here, we test these contrasting predictions in two experiments using a one-generation artificial language learning paradigm designed to provide independent measures of learning accuracy and structure addition. Adults (N = 48 in each study) were exposed to a semi-regular language (with probabilistic structure) and had to learn it: Learning was assessed using seen items, whereas structure addition was calculated over unseen items. In both studies, we find a strong positive correlation between individuals' ability to learn the language and their tendency to add structure to it: Better learners also produced more structured languages. These findings suggest a strong link between learning and generalization. We discuss the implications of these findings for iterated language models and theories of language change more generally.     

Iterated learning: Intergenerational knowledge transmission reveals inductive biases

Cultural transmission of information plays a central role in shaping human knowledge. Some of the most complex knowledge that people acquire, such as languages or cultural norms, can only be learned from other people, who themselves learned from previous generations. The prevalence of this process of “iterated learning” as a mode of cultural transmission raises the question of how it affects the information being transmitted. Analyses of iterated learning utilizing the assumption that the learners are Bayesian agents predict that this process should converge to an equilibrium that reflects the inductive biases of the learners. An experiment in iterated function learning with human participants confirmed this prediction, providing insight into the consequences of intergenerational knowledge transmission and a method for discovering the inductive biases that guide human inferences.     

Revealing human inductive biases for category learning by simulating cultural transmission

We explored people’s inductive biases in category learning—that is, the factors that make learning category structures easy or hard—using iterated learning. This method uses the responses of one participant to train the next, simulating cultural transmission and converging on category structures that people find easy to learn. We applied this method to four different stimulus sets, varying in the identifiability of their underlying dimensions. The results of iterated learning provide an unusually clear picture of people’s inductive biases. The category structures that emerge often correspond to a linear boundary on a single dimension, when such a dimension can be identified. However, other kinds of category structures also appear, depending on the nature of the stimuli. The results from this single experiment are consistent with previous empirical findings that were gleaned from decades of research into human category learning.     

A Bayesian Model of Biases in Artificial Language Learning: The Case of a Word‐Order Universal

In this article, we develop a hierarchical Bayesian model of learning in a general type of artificial language‐learning experiment in which learners are exposed to a mixture of grammars representing the variation present in real learners’ input, particularly at times of language change. The modeling goal is to formalize and quantify hypothesized learning biases. The test case is an experiment ( Culbertson, Smolensky, & Legendre, 2012 ) targeting the learning of word‐order patterns in the nominal domain. The model identifies internal biases of the experimental participants, providing evidence that learners impose (possibly arbitrary) properties on the grammars they learn, potentially resulting in the cross‐linguistic regularities known as typological universals. Learners exposed to mixtures of artificial grammars tended to shift those mixtures in certain ways rather than others; the model reveals how learners’ inferences are systematically affected by specific prior biases. These biases are in line with a typological generalization—Greenberg's Universal 18—which bans a particular word‐order pattern relating nouns, adjectives, and numerals.     

Effects of visual information on adults' and infants' auditory statistical learning

Infant and adult learners are able to identify word boundaries in fluent speech using statistical information. Similarly, learners are able to use statistical information to identify word-object associations. Successful language learning requires both feats. In this series of experiments, we presented adults and infants with audio-visual input from which it was possible to identify both word boundaries and word-object relations. Adult learners were able to identify both kinds of statistical relations from the same input. Moreover, their learning was actually facilitated by the presence of two simultaneously present relations. Eight-month-old infants, however, do not appear to benefit from the presence of regular relations between words and objects. Adults, like 8-month-olds, did not benefit from regular audio-visual correspondences when they were tested with tones, rather than linguistic input. These differences in learning outcomes across age and input suggest that both developmental and stimulus-based constraints affect statistical learning.     

Atypical birdsong and artificial languages provide insights into how communication systems are shaped by learning,use, and transmission

In this article, I argue that a comparative approach focusing on the cognitive capacities and behavioral mechanisms that underlie vocal learning in songbirds and humans can provide valuable insights into the evolutionary origins of language. The experimental approaches I discuss use abnormal song and atypical linguistic input to study the processes of individual learning, social interaction, and cultural transmission. Atypical input places increased learning and communicative pressure on learners, so exploring how they respond to this type of input provides a particularly clear picture of the biases and constraints at work during learning and use. Furthermore, simulating the cultural transmission of these unnatural communication systems in the laboratory informs us about how learning and social biases influence the structure of communication systems in the long run. Findings based on these methods suggest fundamental similarities in the basic social–cognitive mechanisms underlying vocal learning in birds and humans, and continuing research promises insights into the uniquely human mechanisms and into how human cognition and social behavior interact, and ultimately impact on the evolution of language.     

Amount of Learning and Signal Stability Modulate Emergence of Structure and Iconicity in Novel Signaling Systems

Iterated language learning experiments that explore the emergence of linguistic structure in the laboratory vary considerably in methodological implementation, limiting the generalizability of findings. Most studies also restrict themselves to exploring the emergence of combinatorial and compositional structure in isolation. Here, we use a novel signal space comprising binary auditory and visual sequences and manipulate the amount of learning and temporal stability of these signals. Participants had to learn signals for meanings differing in size, shape, and brightness; their productions in the test phase were transmitted to the next participant. Across transmission chains of 10 generations each, Experiment 1 varied how much learning of auditory signals took place, and Experiment 2 varied temporal stability of visual signals. We found that combinatorial structure emerged only for auditory signals, and iconicity emerged when the amount of learning was reduced, as an opportunity for rote-memorization hampers the exploration of the iconic affordances of the signal space. In addition, compositionality followed an inverted u-shaped trajectory raising across several generations before declining again toward the end of the transmission chains. This suggests that detection of systematic form-meaning linkages requires stable combinatorial units that can guide learners toward the structural properties of signals, but these combinatorial units had not yet emerged in these unfamiliar systems. Our findings underscore the importance of systematically manipulating training conditions and signal characteristics in iterated language learning experiments to study the interactions between the emergence of iconicity, combinatorial and compositional structure in novel signaling systems.     

Getting it right by getting it wrong: When learners change languages

When natural language input contains grammatical forms that are used probabilistically and inconsistently, learners will sometimes reproduce the inconsistencies; but sometimes they will instead regularize the use of these forms, introducing consistency in the language that was not present in the input. In this paper we ask what produces such regularization. We conducted three artificial language experiments, varying the use of determiners in the types of inconsistency with which they are used, and also comparing adult and child learners. In Experiment 1 we presented adult learners with scattered inconsistency - the use of multiple determiners varying in frequency in the same context - and found that adults will reproduce these inconsistencies at low levels of scatter, but at very high levels of scatter will regularize the determiner system, producing the most frequent determiner form almost all the time. In Experiment 2 we showed that this is not merely the result of frequency: when determiners are used with low frequencies but in consistent contexts, adults will learn all of the determiners veridically. In Experiment 3 we compared adult and child learners, finding that children will almost always regularize inconsistent forms, whereas adult learners will only regularize the most complex inconsistencies. Taken together, these results suggest that regularization processes in natural language learning, such as those seen in the acquisition of language from non-native speakers or in the formation of young languages, may depend crucially on the nature of language learning by young children.     

Acquiring and processing verb argument structure: distributional learning in a miniature language

Adult knowledge of a language involves correctly balancing lexically-based and more language-general patterns. For example, verb argument structures may sometimes readily generalize to new verbs, yet with particular verbs may resist generalization. From the perspective of acquisition, this creates significant learnability problems, with some researchers claiming a crucial role for verb semantics in the determination of when generalization may and may not occur. Similarly, there has been debate regarding how verb-specific and more generalized constraints interact in sentence processing and on the role of semantics in this process. The current work explores these issues using artificial language learning. In three experiments using languages without semantic cues to verb distribution, we demonstrate that learners can acquire both verb-specific and verb-general patterns, based on distributional information in the linguistic input regarding each of the verbs as well as across the language as a whole. As with natural languages, these factors are shown to affect production, judgments and real-time processing. We demonstrate that learners apply a rational procedure in determining their usage of these different input statistics and conclude by suggesting that a Bayesian perspective on statistical learning may be an appropriate framework for capturing our findings.     

Phonological Concept Learning

Linguistic and non‐linguistic pattern learning have been studied separately, but we argue for a comparative approach. Analogous inductive problems arise in phonological and visual pattern learning. Evidence from three experiments shows that human learners can solve them in analogous ways, and that human performance in both cases can be captured by the same models. We test GMECCS (Gradual Maximum Entropy with a Conjunctive Constraint Schema), an implementation of the Configural Cue Model (Gluck & Bower, 1988a ) in a Maximum Entropy phonotactic‐learning framework (Goldwater & Johnson, 2003 ; Hayes & Wilson, 2008 ) with a single free parameter, against the alternative hypothesis that learners seek featurally simple algebraic rules (“rule‐seeking”). We study the full typology of patterns introduced by Shepard, Hovland, and Jenkins ( 1961 ) (“SHJ”), instantiated as both phonotactic patterns and visual analogs, using unsupervised training. Unlike SHJ, Experiments 1 and 2 found that both phonotactic and visual patterns that depended on fewer features could be more difficult than those that depended on more features, as predicted by GMECCS but not by rule‐seeking. GMECCS also correctly predicted performance differences between stimulus subclasses within each pattern. A third experiment tried supervised training (which can facilitate rule‐seeking in visual learning) to elicit simple rule‐seeking phonotactic learning, but cue‐based behavior persisted. We conclude that similar cue‐based cognitive processes are available for phonological and visual concept learning, and hence that studying either kind of learning can lead to significant insights about the other.     

Cross-situational Learning From Ambiguous Egocentric Input Is a Continuous Process: Evidence Using the Human Simulation Paradigm

Recent laboratory experiments have shown that both infant and adult learners can acquire word-referent mappings using cross-situational statistics. The vast majority of the work on this topic has used unfamiliar objects presented on neutral backgrounds as the visual contexts for word learning. However, these laboratory contexts are much different than the real-world contexts in which learning occurs. Thus, the feasibility of generalizing cross-situational learning beyond the laboratory is in question. Adapting the Human Simulation Paradigm, we conducted a series of experiments examining cross-situational learning from children's egocentric videos captured during naturalistic play. Focusing on individually ambiguous naming moments that naturally occur during toy play, we asked how statistical learning unfolds in real time through accumulating cross-situational statistics in naturalistic contexts. We found that even when learning situations were individually ambiguous, learners’ performance gradually improved over time. This improvement was driven in part by learners’ use of partial knowledge acquired from previous learning situations, even when they had not yet discovered correct word-object mappings. These results suggest that word learning is a continuous process by means of real-time information integration.     

Social cues modulate the representations underlying cross-situational learning

Because children hear language in environments that contain many things to talk about, learning the meaning of even the simplest word requires making inferences under uncertainty. A cross-situational statistical learner can aggregate across naming events to form stable word-referent mappings, but this approach neglects an important source of information that can reduce referential uncertainty: social cues from speakers (e.g., eye gaze). In four large-scale experiments with adults, we tested the effects of varying referential uncertainty in cross-situational word learning using social cues. Social cues shifted learners away from tracking multiple hypotheses and towards storing only a single hypothesis (Experiments 1 and 2). In addition, learners were sensitive to graded changes in the strength of a social cue, and when it became less reliable, they were more likely to store multiple hypotheses (Experiment 3). Finally, learners stored fewer word-referent mappings in the presence of a social cue even when given the opportunity to visually inspect the objects for the same amount of time (Experiment 4). Taken together, our data suggest that the representations underlying cross-situational word learning of concrete object labels are quite flexible: In conditions of greater uncertainty, learners store a broader range of information.     

Acquiring variation in an artificial language: Children and adults are sensitive to socially conditioned linguistic variation

Languages exhibit sociolinguistic variation, such that adult native speakers condition the usage of linguistic variants on social context, gender, and ethnicity, among other cues. While the existence of this kind of socially conditioned variation is well-established, less is known about how it is acquired. Studies of naturalistic language use by children provide various examples where children’s production of sociolinguistic variants appears to be conditioned on similar factors to adults’ production, but it is difficult to determine whether this reflects knowledge of sociolinguistic conditioning or systematic differences in the input to children from different social groups. Furthermore, artificial language learning experiments have shown that children have a tendency to eliminate variation, a process which could potentially work against their acquisition of sociolinguistic variation. The current study used a semi-artificial language learning paradigm to investigate learning of the sociolinguistic cue of speaker identity in 6-year-olds and adults. Participants were trained and tested on an artificial language where nouns were obligatorily followed by one of two meaningless particles and were produced by one of two speakers (one male, one female). Particle usage was conditioned deterministically on speaker identity (Experiment 1), probabilistically (Experiment 2), or not at all (Experiment 3). Participants were given tests of production and comprehension. In Experiments 1 and 2, both children and adults successfully acquired the speaker identity cue, although the effect was stronger for adults and in Experiment 1. In addition, in all three experiments, there was evidence of regularization in participants’ productions, although the type of regularization differed with age: children showed regularization by boosting the frequency of one particle at the expense of the other, while adults regularized by conditioning particle usage on lexical items. Overall, results demonstrate that children and adults are sensitive to speaker identity cues, an ability which is fundamental to tracking sociolinguistic variation, and that children’s well-established tendency to regularize does not prevent them from learning sociolinguistically conditioned variation.     

Induction, overhypothesis, and the origin of abstract knowledge. Evidence from 9-month-old infants

Human cognition relies on the ability to extract generalizable knowledge from limited evidence. One type of inductive learning, overhypothesis formation, allows learners to make inferences that take them beyond the limits of direct experience, leading to the creation of abstract knowledge. The developmental roots of this ability have yet to be investigated. We report three experiments examining whether 9-month-old infants are capable of forming overhypotheses. Our results show that when given evidence about a few objects in some category, infants formed a second-order generalization about categories in general. These findings provide evidence that infants possess a powerful mechanism for inductive learning-a mechanism that may be applied to many domains and that can account for the development of many inductive biases later on.     

Learning words’ sounds before learning how words sound: 9-Month-olds use distinct objects as cues to categorize speech information

One of the central themes in the study of language acquisition is the gap between the linguistic knowledge that learners demonstrate, and the apparent inadequacy of linguistic input to support induction of this knowledge. One of the first linguistic abilities in the course of development to exemplify this problem is in speech perception: specifically, learning the sound system of one’s native language. Native-language sound systems are defined by meaningful contrasts among words in a language, yet infants learn these sound patterns before any significant numbers of words are acquired. Previous approaches to this learning problem have suggested that infants can learn phonetic categories from statistical analysis of auditory input, without regard to word referents. Experimental evidence presented here suggests instead that young infants can use visual cues present in word-labeling situations to categorize phonetic information. In Experiment 1, 9-month-old English-learning infants failed to discriminate two non-native phonetic categories, establishing baseline performance in a perceptual discrimination task. In Experiment 2, these infants succeeded at discrimination after watching contrasting visual cues (i.e., videos of two novel objects) paired consistently with the two non-native phonetic categories. In Experiment 3, these infants failed at discrimination after watching the same visual cues, but paired inconsistently with the two phonetic categories. At an age before which memory of word labels is demonstrated in the laboratory, 9-month-old infants use contrastive pairings between objects and sounds to influence their phonetic sensitivity. Phonetic learning may have a more functional basis than previous statistical learning mechanisms assume: infants may use cross-modal associations inherent in social contexts to learn native-language phonetic categories.