Andrew D. M. Smith

Cross-Situational Learning: An Experimental Study of Word-Learning Mechanisms

Cross-situational learning is a mechanism for learning the meaning of words across multiple exposures, despite exposure-by-exposure uncertainty as to the word’s true meaning. We present experimental evidence showing that humans learn words effectively using cross-situational learning, even at high levels of referential uncertainty. Both overall success rates and the time taken to learn words are affected by the degree of referential uncertainty, with greater referential uncertainty leading to less reliable, slower learning. Words are also learned less successfully and more slowly if they are presented interleaved with occurrences of other words, although this effect is relatively weak. We present additional analyses of participants’ trial-by-trial behavior showing that participants make use of various cross-situational learning strategies, depending on the difficulty of the word-learning task. When referential uncertainty is low, participants generally apply a rigorous eliminative approach to cross-situational learning. When referential uncertainty is high, or exposures to different words are interleaved, participants apply a frequentist approximation to this eliminative approach. We further suggest that these two ways of exploiting cross-situational information reside on a continuum of learning strategies, underpinned by a single simple associative learning mechanism.

Intelligent meaning creation in a clumpy world helps communication

This article investigates the problem of how language learners decipher what words mean. In many recent models of language evolution, agents are provided with innate meanings a priori and explicitly transfer them to each other as part of the communication process. By contrast, I investigate how successful communication systems can emerge without innate or transferable meanings, and show that this is dependent on the agents developing highly synchronized conceptual systems. I present experiments with various cognitive, communicative, and environmental factors which affect the likelihood of agents achieving meaning synchronization and demonstrate that an intelligent meaning creation strategy in a clumpy world leads to the highest level of meaning similarity between agents.

Establishing Communication Systems without Explicit Meaning Transmission

This paper investigates the development of experience-based meaning creation and explores the problem of establishing successful communication systems in a population of agents. The aim of the work is to investigate how such systems can develop, without reliance on phenomena not found in actual human language learning, such as the explicit transmission of meaning or the provision of reliable error feedback to guide learning. Agents develop individual, distinct meaning structures, and although they can communicate despite this, communicative success is closely related to the proportion of shared lexicalised meaning, and the communicative systems have a large degree of redundant synonymy.

Cross-situational learning: a mathematical approach

We present a mathematical model of cross-situational learning, in which we quantify the learnability of words and vocabularies. We find that high levels of uncertainty are not an impediment to learning single words or whole vocabulary systems, as long as the level of uncertainty is somewhat lower than the total number of meanings in the system. We further note that even large vocabularies are learnable through cross-situational learning.

Learning Times for Large Lexicons Through Cross-Situational Learning

Cross-situational learning is a mechanism for learning the meaning of words across multiple exposures, despite exposure-by-exposure uncertainty as to a words true meaning. Doubts have been expressed regarding the plausibility of cross-situational learning as a mechanism for learning human-scale lexicons in reasonable timescales under the levels of referential uncertainty likely to confront real word learners. We demonstrate mathematically that cross-situational learning facilitates the acquisition of large vocabularies despite significant levels of referential uncertainty at each exposure, and we provide estimates of lexicon learning times for several cross-situational learning strategies. This model suggests that cross-situational word learning cannot be ruled out on the basis that it predicts unreasonably long lexicon learning times. More generally, these results indicate that there is no necessary link between the ability to learn individual words rapidly and the capacity to acquire a large lexicon.

The Inferential Transmission of Language

Language is a symbolic, culturally transmitted system of communication, which is learnt through the inference of meaning. In this paper, I describe the importance of meaning inference, not only in language acquisition, but also in developing a unified explanation for language change and evolution. Using an agent-based computational model of meaning creation and communication, I show how the meanings of words can be inferred through disambiguation across multiple contexts, using cross-situational statistical learning. I demonstrate that the uncertainty inherent in the process of meaning inference, moreover, leads to stable variation in both conceptual and lexical structure, providing evidence which helps to explain how language changes rapidly without losing communicability. Finally, I describe how an inferential model of communication may provide important theoretical insights into plausible explanations of the bootstrapping of, and the subsequent progressive complexification of, cultural communication systems.

Learning colour words is slow: A cross-situational learning account

Research into child language reveals that it takes a long time for children to learn the correct mapping of colour words. Steels and Belpaeme’s guessing game, however, models fast learning of words. We discuss computational studies based on cross-situational learning, which yield results that are more consistent with the empirical child language data than those obtained by Steels and Belpaeme. Page 1 of 4 Learning colour words is slow: a cross-situational learning account 27-10-06 file://C:\Documents%20and%20Settings\beerens\Local%20Settings\Temporary%20Inte... Steels and Belpaeme (hereinafter S&B) have successfully shown how computational modelling can contribute greatly to the study of the evolution of language and cognition. S&B have – in our opinion correctly – decided to write the paper from an engineer’s point of view. We feel, however, that their model of linguistic communication would have been more realistic, and therefore the results they obtained more robust, if they had used a model of acquiring colour categories through multiple contexts. S&B model the communication between agents using the guessing game model, which is, in itself, not unreasonable. Their claim, however, that this game is “equivalent” to colour chip naming experiments carried out by anthropologists (Section 2.4.2), is not justified, in our opinion. The guessing game is primarily a model of learning through corrective feedback, whereas colour chip naming experiments consist of an anthropologist (A) asking an informant (B) to point out, on a chipset, the focal colour of a colour term from B’s language. There are three important differences between the anthropological experiments and the guessing game. Firstly, B is not doing any learning in fact, A is learning about B’s representation of colour and about B’s language. Secondly, A does not correct B’s responses or provide any feedback about them. Finally, there is no negotiation between A and B about what the words should refer to. This positive feedback loop between the choice of which words to use and their success in communication is the main learning mechanism in the guessing game. Indeed, S&B claim that the feedback loop is a necessary requirement for cultural language development (Section 5, condition 1), although in fact it is widely accepted that children receive little, if any, corrective feedback while learning words (Bloom 2000, but see Chouinard & Clark (2003) for an alternative account). In computational simulations of lexicon creation and learning, similar to those presented by S&B, we have shown that agents using a cross-situational statistical learner (a variant of Siskind’s (1996) cross-situational learner) can successfully develop a shared vocabulary of grounded word meanings without corrective feedback (Smith 2003; Vogt 2004). In our model, as in guessing games, hearers have to infer what speakers are referring to, but, unlike in guessing games, the agents do not have any way of verifying the effectiveness of their attempts at communication. Instead, the agents use co-variances to learn a mapping between words and categories based on the co-occurrence of words and potential referents across multiple situations. Although young children do learn to relate colour terms to colours, it takes them a considerable length of time to find the appropriate mappings (e.g., Andrick & Tager-Flusberg 1986; Sandhofer & Smith 2001). For instance, it has been estimated that, on average, children required over 1,000 trials to learn the three basic colour terms “red”, “green” and “yellow” (Rice 1980, cited in Sandhofer & Smith 2001). Sandhofer & Smith suggest that children go through different stages in learning colour words: first they appear to learn that colour terms relate to the domain of colour, and only then can they actually learn the correct mapping. This has also been observed by Andrick & Tager-Flusberg (1986), who additionally suggest that children find it difficult to learn the boundaries of colour categories, thus slowing down the learning of colour words. Research into child lexical acquisition is, of course, dominated by the problem of referential indeterminacy, and many constraints have been suggested to explain how children reduce indeterminacy (see, e.g., Bloom 2000). Very few of these accounts, however, allow for the fact that children hear words in multiple different contexts, and can use this to determine the intended reference. Recent empirical research, indeed, shows that a cross-situational model of learning provides a robust account of lexical acquisition in general, and of the acquisition of adjectives, including colour categories, in particular. Houston-Price et al. (2003) suggest that the children in their study used cross-situational learning to disambiguate word reference, even though their experiments were designed with attentional cues. In addition, Mather & Schafer (2004) show that children can learn the reference of nouns by exploiting co-variations across multiple contexts. Akhtar & Montague (1999) demonstrate that children use cross-situational learning to discover the meanings of novel adjectives. Klibanoff & Waxman (2000), furthermore, provide empirical support for their proposal that adjectival categories are learnt cross-situationally, within the context of basic level categories. Page 2 of 4 Learning colour words is slow: a cross-situational learning account 27-10-06 file://C:\Documents%20and%20Settings\beerens\Local%20Settings\Temporary%20Inte... A comparison of the guessing game and a cross-situational statistical learner, using computational simulations, has shown that, in the guessing game, coherence in production between agents is considerably higher and that learning is much faster (Vogt & Coumans 2003). This means that agents using cross-situational statistical learning have considerable difficulties in arriving at a shared lexicon, though in the end they manage to overcome them. Note, however, that cross-situational statistical learning improves when: agents’ semantic categories are similar (Smith 2003); learners assume mutual exclusivity (Smith in press); and the context size is relatively small (Smith & Vogt 2004). This slower rate of acquisition is thus consistent with the empirical evidence that children learn colour words relatively slowly. Importantly, as yet unpublished studies have shown that the category variance among agents in the cross-situational learner tends to be much higher than that seen from the guessing games. This suggests that negotiating category boundaries in the crosssituational learner is more difficult, which could confirm Andrick & Tager-Flusberg (1986)’s finding. S&B have presented a model of learning colour words which is fast and based on corrective feedback. Research on child lexicon acquisition suggests, however, that colour categories are actually acquired slowly and through cross-situational learning. If cross-situational learning is, indeed, a more plausible model than the guessing game, then the results achieved by S&B may no longer hold for their account of cultural learning. References Akhtar, N. and Montague, L (1999) Early lexical acquisition: the role of cross-situational learning. First Language 19: 347-358 Andrick, G. R. and Tager-Flusberg, H. (1986) The acquisition of colour terms. Journal of Child Language 13: 119-134 Bloom, P. (2000) How children learn the meaning of words. Cambridge, MA: MIT Press. Chouinard, M. M. and Clark, E. V. (2003) Adult reformulation of child errors as negative evidence. Journal of Child Language 30: 637-669 Houston-Price, C., Plunkett, K., Harris, P. and Duffy, H. (2003). Developmental change in infants’ use of cues to word meaning. Paper presented to XIth European Conference on Developmental Psychology, Catholic University of Milan, Italy. Klibanoff, R. S. and Waxman, S. R. (2000) Basic level object categories support the acquisition of novel adjectives: Evidence from preschool-aged children. Child Development 7(3): 649-659 Mather, E. and Schafer, G. (2004) Object-label covariation: A cue for the acquisition of nouns? Poster presented at the meeting of the International Society of Infant Studies. Chicago. Rice, N. (1980) Cognition to language. Baltimore, MD: University Park Press. Sandhofer, C. M. and Smith, L. B. (2001) Why children learn color and size words so differently: Evidence from adults’ learning of artificial terms. Journal of Experimental Psychology: General 130 (4): 600-620. Siskind, J. M. (1996) A computational study of cross-situational techniques for learning word-tomeaning mappings. Cognition 61: 39-91. Smith, A. D. M. (2003) Intelligent Meaning Creation in a Clumpy World Helps Communication. Page 3 of 4 Learning colour words is slow: a cross-situational learning account 27-10-06 file://C:\Documents%20and%20Settings\beerens\Local%20Settings\Temporary%20Inte... Artificial Life 9(2): 559-574. http://www.ling.ed.ac.uk/~andrew/publications/publications.html Smith, A. D. M. (in press) Mutual Exclusivity: Communicative success despite conceptual divergence. In M. Tallerman (Ed.) Language origins: perspectives on evolution. Oxford: Oxford University Press. http://www.ling.ed.ac.uk/~andrew/publications/publications.html Smith, A. D. M. & Vogt, P. (2004) Lexicon acquisition in an uncertain world. Paper presented at the 5th Evolution of language conference. Leipzig. http://www.ling.ed.ac.uk/~paulv/publications.html Vogt, P. (2004) Minimum cost and the emergence of the Zipf-Mandelbrot law. In J. Pollack, M. Bedau, P. Husbands, T. Ikegami and R. A. Watson (Eds.) Artificial Life IX Proceedings of the Ninth International Conference on the Simulation and Synthesis of Living Systems. The MIT Press. http://www.ling.ed.ac.uk/~paulv/publications.html Vogt, P. and Coumans, H. (2003) Investigating social interaction strategies for bootstrapping lexicon development Journal of Artificial Societies and Social Simulation 6(1). http://jasss.soc.surrey.ac.uk/6/1/4.html Pa

Models of language evolution and change

UNLABELLED In the absence of direct evidence of the emergence of language, the explicitness of formal models which allow the exploration of interactions between multiple complex adaptive systems has proven to be an important tool. Computational simulations have been at the heart of the field of evolutionary linguistics for the past two decades, particularly through the language game and iterated learning paradigms, but these are now being extended and complemented in a number of directions, through formal mathematical models, language-ready robotic agents, and experimental simulations in the laboratory. For further resources related to this article, please visit the WIREs website. CONFLICT OF INTEREST The author has declared no conflicts of interest for this article.

Regularity in Mappings Between Signals and Meanings

We combine information theory and cross-situational learning to develop a novel metric for quantifying the degree of regularity in the mappings between signals and meanings that can be inferred from exposure to language in context. We illustrate this metric using the results of two artificial language learning experiments, which show that learners are sensitive, with a high level of individual variation, to systematic regularities in the input. Analysing language using this measure of regularity allows us to explore in detail how language learning and language use can both generate linguistic variation, leading to language change, and potentially complexify language structure, leading to qualitative language evolution.

The joy of sacs

This paper investigates an idea that was put forward (and hinted at in Fitch, 2000) by Tecumseh Fitch at the Cradle of Language conference in Stellenbosch, South Africa. The idea is that air sacs may have played an important role in early hominid vocalizations. Many other primates have air sacs, notably chimpanzees, gorillas and orangutans. It is therefore likely that our latest common ancestor also had air sacs, and the shape of a recently discovered Australopithecine hyoid bone (Alemseged et al., 2006) also points in this direction. Many functions have been proposed for air sacs, among them resonance chambers, sound radiators, CO2 buffers to prevent hyperventilation and means to help exaggerate size. Whatever their function in other primates, the fact that humans are the only apes that do not have air sacs might be related to the fact that we have speech. Here I investigate the influence of the presence of an air sac on the set of (vowel) signals that can be produced.