Thursday, November 8, 2012

Steels - Experiments in Cultural Language Evolution

Luc Steels co-founded the Computer Science Department at Vrije Universiteit Brussel, and is part of their Artificial Intelligence Lab. In 1996 he founded the Sony Computer Science Laboratory in Paris, and is now  ICREA research professor at the Institute for Evolutionary Biology. Since about 1995, he has been heavily involved in finding practical ways to demonstrate how language evolves. His main approach, revealed in a wide range of publications, is to simulate the emergence of language in computational and, more recently, robotic agents. A recent book, Experiments in Cultural Language Evolution, published by John Benjamins, details these experiments by his team.  Here is an interview from 2006, courtesy of "Talking Robots." His work has resulted in a theory of language called "Fluid Construction Grammar" which reflects many issues brought up in usage-based approaches to language acquisition.

Review of  Experiments in Cultural Language Evolution

Reviewer:  Nick Moore
Book Title: Experiments in Cultural Language Evolution
Book Author: Luc Steels
Publisher: John Benjamins
Linguistic Field(s): Computational Linguistics; Historical Linguistics; Linguistic Theories; Text/Corpus Linguistics


The ten papers collected in “Experiments in Cultural Language Evolution” represent the state-of-the-art of research into simulated multi-agent interaction. Centered around Luc Steels’ work at the Sony Computer Science Laboratory, Paris, this volume represents the culmination of more than a decade of work dedicated to uncovering the practicalities of language evolution in a social setting. The book is divided into three sections. An introductory section comprises a Foreword and an Introduction, both by Steels, that set out the direction and the theoretical framework for the remaining papers. Part 1 describes experiments in vocabulary evolution and Part 2 details how grammatical features evolve in experiments in the same framework. Each experiment enhances results gained in previous experiments.

The Foreword places the volume in its historical context by stressing that the question of language evolution is almost as plagued by speculation today as in 1995, when Steels launched this research project. Because there is no fossil record and because we cannot allow any modern language to represent languages as they first emerged, we can only be guided by general principles of evolution when theorising the evolution of languages. Steels and his team have since synthesised an approach to language evolution that attempts to simulate the evolution of language in a cultural context by using computational agents, typically embodied as robots. The Foreword also summarises each chapter.

Chapter 1, “Self-organization and selection in cultural language evolution” by Steels, outlines the theoretical framework for the empirical descriptions in the remaining chapters. Steels demands that any theoretical description of language evolution be biologically feasible, demonstrate advantage to social reproduction, and adapt to cultural change. Language in this model is assumed to be open-ended, distributed, and transmitted non-telepathically. The key aspects of an evolutionary theory that are applied to language are fundamentally functional, i.e., Does language succeed in communicating? Agents apply general strategies that adapt language for optimum expressive adequacy, cognitive effort, learnability and social conformity. The repeated application of these strategies to instances of communicative events produces a language system based on the probability of communicative success. The language system is the combination of the general cognitive capabilities of routine processing and meta-analysis. Ready-made responses may be available to a speaker, but analysis is required to evaluate those responses. Where self-evaluation indicates a lack of success, a repair is introduced. Repair actions may require a reframing of the chosen sentence, the selection of an alternative lexical item, or the creation of a new item or structure. Self-evaluation is possible because of a routine termed ‘re-entry’ (i.e. a process that matches the mirror-neuron hypothesis; see Rizzolatti and Craighero, 2004), which allows the speaker to practice the communicative effect of the chosen sentence before it is articulated by acting as the hearer in an internal process.

While the language system adapts, constrained by language strategies, language items emerge through a self-enforcing cumulative process of invention, trial, and alignment between agents. As with repair, alignment is central to the self-organising character of language. Alignment is the social enaction of frequency, such that the communicatively successful use of a language item increases the likelihood of it being adopted by other agents. This process is demonstrated throughout the volume in various experiments. To further strengthen the centrality of self-organisation in language evolution, Steels also employs the principle of 'structural coupling' (Maturana, 2002), which facilitates alignment through linguistic transmission due to the structure of an organism and its interaction with the living, non-living and linguistic environment, without the need for intention or a central authority.

The key issue for Steels is to provide empirical evidence for the theoretical framework sketched here. Contemporary evolutionary linguistic processes, such as creolisation, can shed light on how language evolves, as can placing linguistically-competent subjects into a context where new language must be invented to complete a communicative task. However, Steels and his collaborators choose to model evolutionary processes computationally and robotically, using embodied agents to enact language games. Throughout the volume, robots engage in: acquisition experiments, where one linguistically competent robot passes on a linguistic system to another robot with a pidgin version of the language, through tutoring, although neither robot knows which has the full version; emergent experiments, where both robot agents, using the strategies described above, collaborate to converge on a non-predetermined stable linguistic system; and reconstruction experiments, where strategies are varied by agents to simulate known linguistic evolution. The remaining chapters describe these experiments for selected vocabulary (Part 1) and grammatical (Part 2) features.

Steels and Martin Loetzsch start Part 1 with the simplest language game: the naming game. In the “non-grounded” version of this emergent experiment, two agents share the same viewpoint of a set of objects. The speaker offers a name for an object, to which the hearer points. If the hearer matches the speaker’s object, a new round is played. However, a number of repairs may occur. The speaker may identify an object with no known name, in which case it has to invent one. The hearer may not know the word, so it guesses the object. If the guess is correct, the new word is remembered, but if it is incorrect, the speaker points out the object, and the new word is remembered. If the hearer knows a different word for the same object, scores are given to the different words so that, through usage, agents converge on agreed words. Thus, in one experiment, twelve words for five objects after 50 games become five to six words, on average, after 200 games. In the “grounded” version of the game, the agents are mobile and may see the same objects from different angles. Identifying objects through luminance, yellow/blue and red/green scores, x and y coordinates, and height and width measurements, agents store prototypes of objects which they then collaborate to name with other agents, using similar strategies and repairs as in the previous game. Aggregate results produce close to 100% communicative success after 1,000 games producing 20 terms after 18 views of 10 objects. Adding the ability to both track moved objects and update prototype models results in about 90% success with 11 terms from 1,500 games after 16 views of 10 objects. That is, these two learning heuristics produce far less ambiguity and synonymy.

In “Language Strategies for Color”, Joris Bleys engages robot agents in naming games for colour, thereby accounting for how categories emerge from a natural spectrum. Agents carry out the same language games as in the previous experiment, but here the objects are distinguishable only by colour. Robot agents use a learning strategy that adjusts, rather than replaces, the current prototypical colour towards the speaker’s use of the colour word whenever communication is unsuccessful. Using English words based on scores for brightness, red/green and yellow/blue scores, robot agents score about 83% communicative success, matching baseline or target scores set by human agents. In an emergent experiment using only hue (or brightness), robot agents achieve about 72% success. To make the experiments more closely match natural language, Bleys also investigates graded membership of colour categories (e.g. “only slightly”, “somewhat” or “very” red). In a reconstruction experiment, robot agents produce words that were “qualitatively similar” (p.74) to their human counterparts in baseline data. Similarly, in acquisition experiments, robot agents demonstrate communicative success at rates marginally below humans. An emergent experiment for colour produces almost 95% communicative success with little variance for 5 words after about 15,00 games. The impressive results for graded membership demonstrate another important aspect of these evolutionary experiments: language strategies adapt to give selective advantage. In this case, graded membership of colours allows a higher rate of success than brightness-only or hue-and-brightness systems.

The experiments in the next two chapters, “Emergent mirror systems for body language” by Steels and Michael Spranger and “The co-evolution of basic spatial terms and categories” by Spranger, add complexity to the linguistic models developed in the previous two chapters by adding verbal and adverbial options (Steel and Spranger) and prepositional meanings (Spranger). Spranger’s experimental embodied-robotic subjects achieve 98% communicative success when reconstructing German spatial terms. Steel and Spranger claim that “It is only by the full integration of all aspects of language with sophisticated sensory-motor intelligence that agents were able to arrive at a shared communicative system that is adequate for the game” (107) of correctly ordering a fellow robot agent to strike a particular pose. That is, communicative success is achieved by: grounding the agents in a sensory experience relative to their own body and its parts; employing a prototypical, rather than categorical, approach to language; simulating mirror neurons (by enabling robots to simulate and monitor, without enacting, a motor programme); and providing feedback loops for the motor system.

Part 1 culminates in the chapter “Multi-dimensional meanings in lexical formation”, by Pieter Wellens and Loetzsch, which attempts to simulate a more natural environment for lexical emergence and demonstrate the adaptive benefits of the strategies adopted in the studies in this volume. The language games played by robot agents in the preceding chapters all focus on one aspect of language, but this does not reflect natural language use, when speakers must select the most suitable linguistic features to distinguish objects. The most favourable results are obtained when agents use a probability-based ‘Adaptive Strategy’ for word learning, whereby a fuzzy-logic algorithm for ‘best fit’ is used in naming objects as speaker or hearer. In experiments where 25 agents able to distinguish 16 features per object play 4,000 games each, totalling 50,000 games over 10 repetitions, the agents achieve 90% communicative success after 10,000 games, and approach a 98% success rate after 30,000 games. Another measure, lexicon coherence, which quantifies the alignment between agents’ lexicons at any time, reaches 0.4 after 10,000 games and averages only as high as 0.45 on a scale of -1 to +1 after 50,000 games. This reflects natural language, where high levels of communicative success are achieved even when agents do not totally agree on word meanings.

Part Two of the book, ‘Emergence of Grammatical Systems’, opens with Remi van Trijp’s ‘The evolution of case systems for marking event structure’, which posits three bold hypotheses: 1. “Case evolves because it has selective advantage for communication” (170); 2. case emerges when a population shares a ‘case strategy’; and 3. “Case markers can be repurposed for a different language system if the original selective advantages of a case system have been ‘usurped’ by more dominant, competing systems in the language” (170). In experiments where one robot agent describes a scene that the two agents have just watched together, robot agents acquire the case system for German, although van Trijp rejects the need for 'a priori' grammatical categories. After 5,000 games, coherence scores are above 0.95, the language system is highly systematic, and cognitive effort is at a minimum, thus providing support for the first hypothesis. Moreover, the evolution of the Spanish personal pronoun system is reconstructed in experiments that provide evidence for hypotheses 2 and 3 above. As with native speakers, grammatical variation is accommodated by robot agents who produce language with preferences for certain structures. Similarly, subsequent experiments demonstrate a paradigm shift in the population, with preferences moving from one system to another. In the conclusion, van Trijp is careful to emphasise that these experiments demonstrate a high level of communicative success using general shared cognitive strategies – typically, “analogical reasoning or similarity-based categorization” (202).

In “Emergent functional grammar for space”, Spranger and Steels demonstrate the selective advantage of grammaticalising spatial relationships over the solely lexical variant in experiments that reconstruct German and that self-organise into an emergent system. Crucially, they show how a semantically-oriented strategy towards grammaticalised spatial relationships requires less cognitive effort for greater communicative success. Similarly, Katrien Beuls, Steels and Sebastian Höfer’s experiments into “The emergence of internal agreement systems” produce results that reduce cognitive effort and ambiguity by grouping related words into groups or phrases. Kateryna Gerasymova, Spranger and Beuls investigate the Russian system of Aktionsarten in “A language strategy for aspect”. Although the Russian system of aspect is considered complex and elaborate, robot agents are able to reconstruct and acquire the system, partially aided by the ability to accept holophrases (a learned combination of words) for later analysis. Robot agents then demonstrate how an entirely new aspect system can emerge. As in the experiments by Wellens and Loetzsch, the final chapter ''The emergence of quantifiers'', by Simon Pauw and Joseph Hilferty, demonstrates the selective advantage of fuzzy categories by focusing on quantifying expressions. Experiments in acquisition and formation compare the alternative strategies of absolute quantification and scalable quantification, resulting in the conclusion that the more unpredictable the environment, the more likely a scalable strategy will prevail.


Although each paper has different authors, the volume exhibits both a remarkable sense of consistency and a clear sense of progression from one chapter to the next. The research reveals a sense of direction shared by Steels and the other contributors that is laid out in Chapter One. In fact, it is advisable to read Chapter One again after examining the results of later experiments, in order to fully appreciate the significance of the bold approach taken by this team of researchers.

The greatest danger of depending on functional explanations to support a hypothesis is that evidence can only be interpreted as supporting an inert status quo. Fortunately, Steels and colleagues avoid this theoretical blind alley by incorporating the dynamics of alignment and the explanations and mechanisms for linguistic change. For instance, in van Trijp’s chapter, experimental evidence provides support for the hypothesis that the advantages provided by grammatical case in Spanish have been replaced by other grammatical features, freeing case markers to function in new ways. Perhaps my only concern with some of the papers in the volume is that there is an over-reliance on formal, rather than functional models of language. While some functional models may be difficult to model computationally, there are solutions, such as Halliday and Matthiessen (1999), which may provide the research team with grammatical models more aligned with the non-representational approach to language that is central to the research reported here.

This book and other experiments by the same team provide empirical evidence for the emergence of language based on evolutionary principles, on what we currently understand about brain structure and organisation (e.g. Edelman 1999; 2004) and, significantly, without the need for language-specific acquisition strategies; in all of the experiments here, the learning strategies employed are general cognitive strategies rather than language-specific. The experiments repeatedly demonstrate that: language can emerge without a priori conditions; current language systems can be aligned within a community through structural coupling; known developments in language can be modelled in embodied robotic agents with simulated mirror neurons; and language functions probabilistically, not categorically. I am unaware of any other series of falsifiable experiments that provide verifiable evidence to counter these conclusions, despite many theoretical claims to the contrary. Consequently, this volume should be of value to anyone interested in language evolution, in the application of natural languages to robotic agents, and in general linguistic theory.


Edelman, G.M. 1999. Building a picture of the brain. Annals of the New York Academy of Sciences 882 June 1999, p.68-89

Edelman, G.M. 2004. Wider Than the Sky - The Phenomenal Gift of Consciousness. New Haven: Yale University Press

Halliday, M.A.K. and Matthiessen, C.M.I.M. 1999. Construing Experience through Meaning: A Language-based Approach to Cognition. London: Continuum

Maturana Romesin, H. 2002. Autopoiesis, Structural coupling and cognition: A history of these and other notions in the biology of cognition. Cybernetics and Human Knowing 9(4), pp.5-34

Rizzolatti, G. and Craighero, L. 2004. The Mirror-Neuron system. Annual Review of Neuroscience 27, pp.169-92

Nick Moore has worked in Brazil, Oman, Turkey, the UAE and the UK with students and teachers of English as a foreign language, English for specific and academic purposes, and linguistics. His PhD in applied linguistics from the University of Liverpool addressed information structure in written English. His other research interests include systemic functional linguistics, corpus linguistics, theories of embodiment, lexis and skills in language teaching, and reading programs. He is the co-editor of 'READ', maintains a blog on language, linguistics and learning at and has recently joined the TESOL unit at Sheffield Hallam University.

The review for this book is posted here on 

No comments:

Post a Comment