Language Evolution and Computation Bibliography

We investigate the emergence of iconicity, specifically a bouba-kiki effect in miniature artificial languages under different functional constraints: when the languages are reproduced and when they are used communicatively. We ran transmission chains of (a) participant dyads who played an interactive communicative game and (b) individual participants who played a matched learning game. An analysis of the languages over six generations in an iterated learning experiment revealed that in the Communication condition, but not in the Reproduction condition, words for spiky shapes tend to be rated by naive judges as more spiky than the words for round shapes. This suggests that iconicity may not only be the outcome of innovations introduced by individuals, but, crucially, the result of interlocutor negotiation of new communicative conventions. We interpret our results as an illustration of cultural evolution by random mutation and selection (as opposed to by guided variation).

Most theories of learning would predict a gradual acquisition and refinement of skills as learning progresses, and while some highlight exponential growth, this fails to explain why natural cognitive development typically progresses in stages. Models that do span multiple developmental stages typically have parameters to "switch" between stages. We argue that by taking an embodied view, the interaction between learning mechanisms, the resulting behavior of the agent, and the opportunities for learning that the environment provides can account for the stage-wise development of cognitive abilities. We summarize work relevant to this hypothesis and suggest two simple mechanisms that account for some developmental transitions: neural readiness focuses on changes in the neural substrate resulting from ongoing learning, and perceptual readiness focuses on the perceptual requirements for learning new tasks. Previous work has demonstrated these mechanisms in replications of a wide variety of infant language experiments, spanning multiple developmental stages. Here we piece this work together as a single model of ongoing learning with no parameter changes at all. The model, an instance of the Epigenetic Robotics Architecture (Morse et al 2010) embodied on the iCub humanoid robot, exhibits ongoing multi-stage development while learning pre-linguistic and then basic language skills.

The emergence of signaling systems has been observed in numerous experimental and real-world contexts, but there is no consensus on which (if any) shared mechanisms underlie such phenomena. A number of explanatory mechanisms have been proposed within several disciplines, all of which have been instantiated as credible working models. However, they are usually framed as being mutually incompatible. Using an exemplar-based framework, we replicate these models in a minimal configuration which allows us to directly compare them. This reveals that the development of optimal signaling is driven by similar mechanisms in each model, which leads us to propose three requirements for the emergence of conventional signaling. These are the creation and transmission of referential information, a systemic bias against ambiguity, and finally some form of information loss. Considering this, we then discuss some implications for theoretical and experimental approaches to the emergence of learned communication.

In language, recombination of a discrete set of meaningless building blocks forms an unlimited set of possible utterances. How such combinatorial structure emerged in the evolution of human language is increasingly being studied. It has been shown that it can emerge when languages culturally evolve and adapt to human cognitive biases. How the emergence of combinatorial structure interacts with the existence of holistic iconic form?meaning mappings in a language is still unknown. The experiment presented in this paper studies the role of iconicity and human cognitive learning biases in the emergence of combinatorial structure in artificial whistled languages. Participants learned and reproduced whistled words for novel objects with the use of a slide whistle. Their reproductions were used as input for the next participant, to create transmission chains and simulate cultural transmission. Two conditions were studied: one in which the persistence of iconic form?meaning mappings was possible and one in which this was experimentally made impossible. In both conditions, cultural transmission caused the whistled languages to become more learnable and more structured, but this process was slightly delayed in the first condition. Our findings help to gain insight into when and how words may lose their iconic origins when they become part of an organized linguistic system.

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.

Cross-situational learning has recently gained attention as a plausible candidate for the mechanism that underlies the learning of word-meaning mappings. In a recent study, Blythe and colleagues have studied how many trials are theoretically required to learn a human-sized lexicon using cross-situational learning. They show that the level of referential uncertainty exposed to learners could be relatively large. However, one of the assumptions they made in designing their mathematical model is questionable. Although they rightfully assumed that words are distributed according to Zipf's law, they applied a uniform distribution of meanings. In this article, Zipf's law is also applied to the distribution of meanings, and it is shown that under this condition, cross-situational learning can only be plausible when referential uncertainty is sufficiently small. It is concluded that cross-situational learning is a plausible learning mechanism but needs to be guided by heuristics that aid word learners with reducing referential uncertainty.

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.

This paper reconsiders the diphone-based word segmentation model of Cairns, Shillcock, Chater, and Levy (1997) and Hockema (2006), previously thought to be unlearnable. A statistically principled learning model is developed using Bayes theorem and reasonable assumptions about infants implicit knowledge. The ability to recover phrase-medial word boundaries is tested using phonetic corpora derived from spontaneous interactions with children and adults. The (unsupervised and semi-supervised) learning models are shown to exhibit several crucial properties. First, only a small amount of language exposure is required to achieve the model's ceiling performance, equivalent to between 1day and 1month of caregiver input. Second, the models are robust to variation, both in the free parameter and the input representation. Finally, both the learning and baseline models exhibit undersegmentation, argued to have significant ramifications for speech processing as a whole.

Recent research suggests that language evolution is a process of cultural change, in which linguistic structures are shaped through repeated cycles of learning and use by domain-general mechanisms. This paper draws out the implications of this viewpoint for understanding the problem of language acquisition, which is cast in a new, and much more tractable, form. In essence, the child faces a problem of induction, where the objective is to coordinate with others (C-induction), rather than to model the structure of the natural world (N-induction). We argue that, of the two, C-induction is dramatically easier. More broadly, we argue that understanding the acquisition of any cultural form, whether linguistic or otherwise, during development, requires considering the corresponding question of how that cultural form arose through processes of cultural evolution. This perspective helps resolve the 'logical' problem of language acquisition and has far-reaching implications for evolutionary psychology.

The emergence of shared symbol systems is considered to be a pivotal moment in human evolution and human development. These changes are normally explained by reference to changes in people's internal cognitive processes. We present 2 experiments which provide evidence that changes in the external, collaborative processes that people use to communicate can also affect the structure and organization of symbol systems independently of cognitive change. We propose that mutual-modifiability--opportunities for people to edit or manipulate each other's contributions--is a key constraint on the emergence of complex symbol systems. We discuss the implications for models of language development and the origins of compositionality.

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.

It has been suggested that iconic graphical signs evolve into symbolic graphical signs through repeated usage. This article reports a series of interactive graphical communication experiments using a ``pictionary'' task to establish the conditions under which the evolution might occur. Experiment 1 rules out a simple repetition based account in favor of an account that requires feedback and interaction between communicators. Experiment 2 shows how the degree of interaction affects the evolution of signs according to a process of grounding. Experiment 3 confirms the prediction that those not involved directly in the interaction have trouble interpreting the graphical signs produced in Experiment 1. On the basis of these results, this article argues that icons evolve into symbols as a consequence of the systematic shift in the locus of information from the sign to the users' memory of the sign's usage supported by an interactive grounding process.

The grounding of symbols in computational models of linguistic abilities is one of the fundamental properties of psychologically-plausible cognitive models. This paper presents an embodied model for the grounding of language in action based on epigenetic robots. Epigenetic robotics is one of the new cognitive modeling approaches to modeling autonomous mental development. The robot model is based on an integrative vision of language, in which linguistic abilities are strictly dependent on, and grounded in, other behaviors and skills. It uses simulated robots that learn through imitation the names of basic actions. Robots also learn higher-order action concepts through the process of grounding transfer. The simulation demonstrates how new, higher-order behavioral abilities can be autonomously built upon previously-grounded basic action categories, following linguistic interaction with human users.

The emergence of human communication systems is typically investigated via 2 approaches with complementary strengths and weaknesses: naturalistic studies and computer simulations. This study was conducted with a method that combines these approaches. Pairs of participants played video games requiring communication. Members of a pair were physically separated but exchanged graphic signals through a medium that prevented the use of standard symbols (e.g., letters). Communication systems emerged and developed rapidly during the games, integrating the use of explicit signs with information implicitly available to players and silent behavior-coordinating procedures. The systems that emerged suggest 3 conclusions: (a) signs originate from different mappings; (b) sign systems develop parsimoniously; (c) sign forms are perceptually distinct, easy to produce, and tolerant to variations.

How do communities establish shared communication systems? The Common Knowledge view assumes that symbolic conventions develop through the accumulation of common knowledge regarding communication practices among the members of a community. In contrast with this view, it is proposed that coordinated communication emerges a by-product of local interactions among dyads. A set of multi-agent computer simulations show that a population of 'egocentric' agents can establish and maintain symbolic conventions without common knowledge. In the simulations, convergence to a single conventional system was most likely and most efficient when agents updated their behavior on the basis of local rather than global, system-level information. The massive feedback and parallelism present in the simulations gave rise to phenomena that are often assumed to result from complex strategic processing on the part of individual agents. The implications of these findings for the development of theories of language use are discussed.

Naturally occurring speech contains only a limited amount of complex recursive structure, and this is reflected in the empirically documented difficulties that people experience when processing such structures. We present a connectionist model of human performance in processing recursive language structures. The model is trained on simple artificial languages. We find that the qualitative performance profile of the model matches human behavior, both on the relative difficulty of center-embedding and cross-dependency, and between the processing of these complex recursive structures and right-branching recursive constructions. We analyze how these differences in performance are reflected in the internal representations of the model by performing discriminant analyses on these representations both before and after training. Furthermore, we show how a network trained to process recursive structures can also generate such structures in a probabilistic fashion. This work suggests a novel explanation of people's limited recursive performance, without assuming the existence of a mentally represented competence grammar allowing unbounded recursion.

Time underlies many interesting human behaviors. Thus, the question of how to represent time in connectionist models is very important. One approach is to represent time implicitly by its effects on processing rather than explicitly (as in a spatial representation). The current report develops a proposal along these lines first described by Jordan (1986) which involves the use of recurrent links in order to provide networks with a dynamic memory. In this approach, hidden unit patterns are fed back to themselves; the internal representations which develop thus reflect task demands in the context of prior internal states. A set of simulations is reported which range from relatively simple problems (temporal version of XOR) to discovering syntactic/semantic features for words. The networks are able to learn interesting internal representations which incorporate task demands with memory demands; indeed, in this approach the notion of memory is inextricably bound up with task processing. These representations reveal a rich structure, which allows them to be highly context-dependent while also expressing generalizations across classes of items. These representations suggest a method for representing lexical categories and the type/token distinction.