How did language evolve? This question is among the great unsolved mysteries and a fundamental scientific challenge. We investigate ways in which artificial agents can self-organize languages with natural-language like properties and how meaning can co-evolve with language. Our research is based on the hypothesis that language is a complex adaptive system that emerges through adaptive interactions between agents and continues to evolve in order to remain adapted to the needs and capabilities of the agents. We explore this hypothesis by implementing the full cycle of speaker and hearer as they play situated language games and observing the characteristics of the languages that emerge.
How did language evolve? This question is among the great unsolved mysteries. Other big questions about our nature as human beings and our place in the universe have at least tentative approaches to answers. Questions like where the universe came from, how life evolved, how human beings evolved as a species… most people can frame at least a rough response. They have a grip on these questions. The question of the origin of language – that capacity perhaps most characteristic of our species – remains mysterious. This is a fundamental scientific challenge.
At the Sony Computer Science Laboratories a project has emerged over the last decade which tackles this question directly and offers the possibility of an answer. We explore the evolution of human language through the interactions of communities of autonomous robots over many cycles of interaction. The robots have capacities which we put in – both hardware and software. They have perceptual capacities, for example they may be able to hear and distinguish sounds, detect gestures. They can move, interact. And we can put in different software determining for example how they react to items they don’t recognize, how they respond to deviation from a pattern, and much more. We choose sets of parameters and set communities running over cycles of many thousands of interactions. From these simple interactions, and the individual responses of agents, complex systems arise which actually have the features of communication in natural languages.
This project involves achieving vast goals, including the core goal of formalizing natural language, to be able to model it. The challenge here is the sheer complexity of the whole of what makes up language, from the acts of hearing, producing sound, articulation and interpretation, parsing, incompleteness, error, the speed at which the whole process takes place. A normal adult speaker probably knows at least 100 000 units of language construction, which are manipulated in varying conditions of audibility, success and failure, with language speakers from a similar dialect group or very different, very rapidly. Yet we manage to communicate. Moreover, we innovate in communication. The complexity and multiple parameters influencing transmission of meaning are not the only barrier to creating a successful model. Language changes, shifts, modifies, is dynamic. New meanings arise, are shared, catch on, spread, die out – sometimes very rapidly, or sometimes, as in profound changes in a language structure, over thousands of years. The dynamism, the fluidity of language, must be captured in a successful model. How is it possible that agents can innovate, introduce new meanings, and these meanings can come to be shared? How is innovation possible?
After more than a decade of research CSL teams working together with the VUB AI laboratory have built a formalism (called FCG) that works, that encompasses these features, and moreover that is available to other researchers. Features described below suggest the richness of the data that can now be mined using this methodology and the perspectives for research that open up.
Why are robots the right solution for embodying the experiment? Why is modeling the problem with robots interesting?
The agents are able to act, interact and perceive – the basic situation of human beings prior to communication through language. Moreover, all these aspects of their capacities can be modified by the experimenter, for example how they process sound and “hear”. This gives the opportunity for testing different hypotheses about what might be necessary – what we might have to have in place – for language to be possible. Obviously, those capacities have to be embodied in the robots, which is a whole separate programming challenge – to have agents which are capable of perceiving the world, or responding to the world. A second key achievement of Sony CSL has been development of a framework called IRL link which relates the agents to the world through their sensory-motor apparatus.
Why robots not people?
Apart from the fact that they can be set up with chosen parameters, and start with a “clean slate”, there are two game-changing advantages: number, and time. The communities of agents can be set up with hundreds of agents, to interact over thousands of cycles of interaction, as the experimenters call it, “cooking” the experiment over a long period of time. This gives us access to peering into the dynamics of change which represent in human terms hundreds of thousands of years. We can truly turn the clock back. To give an idea of the slowness of change of some of the strata of language: the shift from the Old German case system to Modern German is an event that started around the year 1100 and is still continuing (link). With this methodology we can run simulations to try and understand the forces at work that are bringing about this change. And crucially, these experiments are scientifically rigorous. Any other researcher can repeat them. In short, using this methodology the experimenter has powerful control over the capacities of the agents, plus the ability to make large numbers of them interact over very large time cycles, producing results that can be replicated by other experimenters.
The excitement surrounding this project is the possibility of using this new methodology to actually make concrete progress in understanding the evolution of language.
Several insights and resources are key to the success of this methodology. The major breakthrough is the vision of the potential combined power of these tools when applied to the question of the origin of language. They include:
Thoughts about simplicity and complexity.
A key insight here is that advances in understanding complex systems from other disciplines – in particular evolutionary biology – can shed light on the origins of language. Consider our recent understanding of how communities of ants, for example, can create nests, or chains to transport food. There is no central organization, no prior planning. Each individual ant makes simple iterative steps involving touching, pheromones, movement. The goal is achieved through self-organization from the bottom up, not imposition from above. Local alignment of individuals results in global coherence, in a giant act of collective bootstrapping.
A key aspect of this insight is also that from simple interactions agents can come to have shared understandings. They can learn, can acquire new concepts from other agents. This is fundamental to what language actually is: a system of shared meanings which make it possible to communicate.
Moreover, modeling the evolution of language in this way gives us the tools to model the dynamism of natural languages. Variation, innovation, emergence of new structures, their adoption, spread and possible elimination can be understood within this model (link to Experiments in Cultural Language Evolution). Which holds the promise that invention of language itself can be modeled.
We have the possibility of putting this insight to work. We have the key elements of hardware, software and community. We have robots with wide-ranging perception capacities, and capacities for generating sound, gesture, reacting.
Over a decade of work we have produced softwares for modeling language, and for enabling agents to perceive.
With a multiplicity of agents we can simulate a community interacting over time.
We have a community of autonomous, self-organizing agents whose interactions over time produce features and structures of natural languages.
We now have the tools to model. Hence the relevance and power of this methodology. What we are talking about is so immensely complex that only now do we have the computational power to make models and run simulations that can mimic natural language.
Moreover, since we can model, our results are rigorous. They can be tested by other experimenters, are scientific.
Richness of data
This approach has an immense richness of material to draw on, ranging from historical linguistics to anthropology to behavioral science. There are enormous amounts of rich data to build on, which have been extensively studied, and which can be fed into these simulations. The collapse of case structures from Old German to Modern German, the appearance of articles in European languages 2 000 years ago, the differential spectra of color recognition between different languages. We are just beginning to probe the potential for this methodology.
And the results are also very rich – they feed into one another. Each part adds more than just that piece of knowledge. We can come to understand how articles come into a language or how other structures collapsed – but the understanding we gain is much wider than that. Each experiment adds to the whole body of knowledge, and can be re-used in other experiments The research is cumulative – tools developed can be used in other experiments.
The combining of insights and technology at this time holds the promise of understanding the forces at play underlying language. The very mechanisms of our humanity.
Obviously: the research carried out has far-reaching practical/commercial applications, ranging from building interfaces that are meaning-driven, new techniques for accessing information or finding meanings in data, or human learning and tutoring, to name but a few.
But there is an excitement itself in asking the question – as physics was able to articulate a theory that took us back to the origins of the Universe – Can we turn back the clock on language? Can we return, in these experiments, to the roots of language, the time before language existed, and find out what had to be in place for language to evolve.
The future research of Sony CSL in the area of language is building on the results achieved and exploring this deep question for our understanding of ourselves.
The reconstruction of phylogenies of cultural artefacts represents an open problem that mixes theoretical and computational challenges. Existing bench- marks rely on simulated phylogenies, where hypotheses on the underlying evolutionary mechanisms are unavoidable, or in real data phylogenies, for which no true evolutionary history is known. Here we introduce a web-based game, Copystree, where users create phylogenies of manuscripts, through successive copying actions, in a fully monitored setup. While players enjoy the experience, Copystree allows to build artificial phylogenies whose evolutionary processes do not obey to any pre-defined theoretical mechanisms, being generated instead with the unpredictability of human creativity. We present the analysis of the data gathered during the first set of experiments and use the artificial phylogenies gathered for a first test of existing phylogenetic algorithms.
Human language users are capable of proficiently learning new constructions and using a language for everyday communication even if they have only acquired a basic linguistic inventory. This paper argues that such robustness can best be achieved through a constructional processing model in which grammatical structures may emerge spontaneously as a side effect of how constructions are combined with each other. This claim is substantiated by a fully operational precision model for Basic English in Fluid Construction Grammar, which is available for online testing. The precision model is the first ever to incorporate key properties from construction grammar in a large-scale setting, such as argument structure constructions and the surface generalization hypothesis, and is therefore a milestone achievement in the field of construction grammar.
Rules are an efficient feature of natural languages
which allow speakers to use a finite set of
instructions to generate a virtually infinite set of
utterances. Yet, for many regular rules, there are
irregular exceptions. There has been lively debate
in cognitive science about how individual learners
acquire rules and exceptions; for example, how they
learn the past tense of preach is preached, but for
teach it is taught. However, for most population or
language-level models of language structure,
particularly from the perspective of language
evolution, the goal has generally been to examine
how languages evolve stable structure, and neglects
the fact that in many cases, languages exhibit
exceptions to structural rules. We examine the
dynamics of regularity and irregularity across a
population of interacting agents to investigate how,
for example, the irregular teach coexists beside the
regular preach in a dynamic language system. Models
show that in the absence of individual biases
towards either regularity or irregularity, the
outcome of a system is determined entirely by the
initial condition. On the other hand, in the
presence of individual biases, rule systems exhibit
frequency dependent patterns in regularity
reminiscent of patterns found in natural
language. We implement individual biases towards
regularity in two ways: through ?child? agents who
have a preference to generalise using the regular
form, and through a memory constraint wherein an
agent can only remember an irregular form for a
finite time period. We provide theoretical arguments
for the prediction of a critical frequency below
which irregularity cannot persist in terms of the
duration of the finite time period which constrains
agent memory. Further, within our framework we also
find stable irregularity, arguably a feature of most
natural languages not accounted for in many other
cultural models of language structure.
After several decades in scientific purgatory, language evolution has reclaimed its place as one of the most important branches in linguistics. This renewed interest is accompanied by powerful new methods for making empirical observations. At the same time, construction grammar is increasingly embraced in all areas of linguistics as a fruitful way of making sense of all these new data, and it has enthused formal and computational linguists, who have developed sophisticated tools for exploring issues in language processing and learning. Separately, linguists and computational linguists are able to explain which changes take place in language and how these changes are possible. When working together, however, they can also address the question of why language evolves over time and how it emerged in the first place. This special issue therefore brings together key contributions from both fields to put evidence and methods from both perspectives on the table.
Word order, argument structure and unbounded dependencies are among the most important topics in linguistics because they touch upon the core of the syntax-semantics interface. One question is whether ?marked? word order patterns, such as The man I talked to vs. I talked to the man, require special treatment by the grammar or not. Mainstream linguistics answers this question affirmatively: in the marked order, some mechanism is necessary for ?extracting? the man from its original argument position, and a special placement rule (e.g. topicalization) is needed for putting the constituent in clause-preceding position. This paper takes an opposing view and argues that such formal complexity is only required for analyses that are based on syntactic trees. A tree is a rigid data structure that only allows information to be shared between local nodes, hence it is inadequate for non-local dependencies and can only allow restricted word order variations. A construction, on the other hand, offers a more powerful representation device that allows word order variations ? even unbounded dependencies ? to be analyzed as the side-effect of how language users combine the same rules in different ways in order to satisfy their communicative needs. This claim is substantiated through a computational implementation of English argument structure constructions in Fluid Construction Grammar that can handle both comprehension and formulation.
Human languages have multiple strategies that allow us to discriminate objects in a vast variety of contexts. Colours have been extensively studied from this point of view. In particular, previous research in artificial language evolution has shown how artificial languages may emerge based on specific strategies to distinguish colours. Still, it has not been shown how several strategies of diverse complexity can be autonomously managed by artificial agents . We propose an intrinsic motivation system that allows agents in a population to create a shared artificial language and progressively increase its expressive power. Our results show that with such a system agents successfully regulate their language development, which indicates a relation between population size and consistency in the emergent communicative systems.
Long-distance dependencies belong to the most controversial challenges in linguistics. These patterns seem to contain constituents that have left their original position in a sentence and that have landed in a different place. A typical example is the relative clause the person I have talked to yesterday, in which the direct object (the person) is not situated in an argument position following the verb, but instead is located at the beginning of the utterance. Upon closer inspection, however, all problems related to long-distance dependencies can be reduced to the limits of phrase structural analyses. A phrase structure tree is a rigid data structure in which information is shared between local nodes. These analyses therefore need to resort to more complex formal machinery in order to overcome this locality constraint, such as using transformations or positing filler-gap constructions. However, there exists a more intuitive alternative within the tradition of cognitive-functional linguistics in which long-distance dependencies do not require special treatment. Instead, these patterns are simply the side effect of how grammatical constructions combine with each other in order to satisfy the communicative needs of language users. Through a computational implementation in Fluid Construction Grammar, this article demonstrates that it is perfectly feasible to formalize this alternative in a model that is capable of both formulating and comprehending utterances.
Rules are an efficient feature of natural languages
which allow speakers to use a finite set of
instructions to generate a virtually infinite set of
utterances. Yet, for many regular rules, there are
irregular exceptions. There has been lively debate
in cognitive science about how individual learners
acquire rules and exceptions; for example, how they
learn the past tense of preach is preached, but for
teach it is taught. In this paper, we take a
different perspective, examining the dynamics of
regularity and irregularity across a population of
interacting agents to investigate how inflectional
rules are applied to verbs. We show that in the
absence of biases towards either regularity or
irregularity, the outcome is determined by the
initial condition, irrespective of the frequency of
usage of the given lemma. On the other hand, in
presence of biases, rule systems exhibit frequency
dependent patterns in regularity reminiscent of
patterns in natural language corpora. We examine the
case where individuals are biased towards linguistic
regularity in two ways: either as child learners, or
through a memory constraint wherein irregular forms
can only be remembered by an individual agent for a
finite time period. We provide theoretical arguments
for the prediction of a critical frequency below
which irregularity cannot persist in terms of the
duration of the finite time period which constrains
Creole languages offer an invaluable opportunity to
study the processes leading to the emergence and
evolution of Language, thanks to the short –
typically a few generations – and reasonably well
defined time-scales involved in their
emergence. Another well-known case of a very fast
emergence of a Language, though referring to a much
smaller population size and different ecological
conditions, is that of the Nicaraguan Sign
Language. What these two phenomena have in common is
that in both cases what is emerging is a contact
language, i.e., a language born out of the
non-trivial interaction of two (or more) parent
languages. This is a typical case of what is known
in biology as horizontal transmission. In many
well-documented cases, creoles emerged in large
segregated sugarcane or rice plantations on which
the slave labourers were the overwhelming
majority. Lacking a common substrate language,
slaves were naturally brought to shift to the
economically and politically dominant European
language (often referred to as the lexifier) to
bootstrap an effective communication system among
themselves. Here, we focus on the emergence of
creole languages originated in the contacts of
European colonists and slaves during the 17th and
18th centuries in exogenous plantation colonies of
especially the Atlantic and Indian Ocean, where
detailed census data are available. Those for
several States of USA can be found at
http://www.census.gov/history, while for Central
America and the Caribbean can be found at
entering in the details of the creole formation at a
fine-grained linguistic level, we aim at uncovering
some of the general mechanisms that determine the
emergence of contact languages, and that
successfully apply to the case of creole formation.
The complex organization of syntax in hierarchical
structures is one of the core design features of
human language. Duality of patterning refers for
instance to the organization of the meaningful
elements in a language at two distinct levels: a
combinatorial level where meaningless forms are
combined into meaningful forms and a compositional
level where meaningful forms are composed into
larger lexical units. The question remains wide open
regarding how such a structure could have
emerged. Furthermore a clear mathematical framework
to quantify this phenomenon is still lacking. The
aim of this paper is that of addressing these two
aspects in a self-consistent way. First, we
introduce suitable measures to quantify the level of
combinatoriality and compositionality in a language,
and present a framework to estimate these
observables in human natural languages. Second, we
show that the theoretical predictions of a
multi-agents modeling scheme, namely the Blending
Game, are in surprisingly good agreement with
empirical data. In the Blending Game a population of
individuals plays language games aiming at success
in communication. It is remarkable that the two
sides of duality of patterning emerge simultaneously
as a consequence of a pure cultural dynamics in a
simulated environment that contains meaningful
relations, provided a simple constraint on message
transmission fidelity is also considered.
Natural languages enable humans to engage in highly complex social and conversational interactions with each other. Alife approaches to the origins and emergence of language typically manage this complexity by carefully staging the learning paths that embodied artificial agents need to follow in order to bootstrap their own communication system from scratch. This paper investigates how these scaffolds introduced by the experimenter can be removed by allowing agents to autonomously set their own challenges when they are driven by intrinsic motivation and have the capacity to self-assess their own skills at achieving their communicative goals. The results suggest that intrinsic motivation not only allows agents to spontaneously develop their own learning paths, but also that they are able to make faster transitions from one learning phase to the next.
Sign languages (SL) require a fundamental rethinking of many basic assumptions about human language processing because instead of using linear speech, sign languages coarticulate facial expressions, shoulder and hand movements, eye gaze and usage of a three-dimensional space. SL researchers have therefore advocated SL-specific approaches that do not start from the biases of models that were originally developed for vocal languages. Unfortunately, there are currently no processing models that adequately achieve both language comprehension and formulation, and the SL-specific developments run the risk of becoming alienated from other linguistic research. This paper explores the hypothesis that a construction grammar architecture offers a solution to these problems because constructions are able to simultaneously access and manipulate information coming from many different sources. This claim is illustrated by a proof-of-concept implementation of a basic grammar for French Sign Language in Fluid Construction Grammar.
One of the most salient hallmarks of construction grammar is its approach to argument structure and coercion: rather than positing many different verb senses in the lexicon, the same lexical construction may freely interact with multiple argument structure constructions. This view has however been criticized from within the construction grammar movement for leading to overgeneration. This paper argues that this criticism falls flat for two reasons: (1) lexicalism, which is the alternative solution proposed by the critics, has already been proven to overgenerate itself, and (2) the argument of overgeneration becomes void if grammar is implemented as a problem-solving model rather than as a generative competence model; a claim that the paper substantiates through a computational operationalization of argument structure and coercion in Fluid Construction Grammar. The paper thus shows that the current debate on argument structure is hiding a much more fundamental rift between practitioners of construction grammar that touches upon the role of grammar itself.
Empirical evidence shows that the rate of irregular
usage of English verbs exhibits discontinuity as a
function of their frequency: the most frequent verbs
tend to be totally irregular. We aim to
qualitatively understand the origin of this feature
by studying simple agent-based models of language
dynamics, where each agent adopts an inflectional
state for a verb and may change it upon interaction
with other agents. At the same time, agents are
replaced at some rate by new agents adopting the
regular form. In models with only two inflectional
states (regular and irregular), we observe that
either all verbs regularise irrespective of their
frequency, or a continuous transition occurs between
a low-frequency state, where the lemma becomes fully
regular, and a high-frequency one, where both forms
coexist. Introducing a third (mixed) state, wherein
agents may use either form, we find that a third,
qualitatively different behaviour may emerge,
namely, a discontinuous transition in frequency. We
introduce and solve analytically a very general
class of three-state models that allows us to fully
understand these behaviours in a unified
framework. Realistic sets of interaction rules,
including the well-known naming game (NG) model,
result in a discontinuous transition, in agreement
with recent empirical findings. We also point out
that the distinction between speaker and hearer in
the interaction has no effect on the collective
behaviour. The results for the general three-state
model, although discussed in terms of language
dynamics, are widely applicable.
Several recent theories have suggested that an
increase in the number of non-native speakers in a
language can lead to changes in morphological
rules. We examine this experimentally by contrasting
the performance of native and non-native English
speakers in a simple Wug-task, showing that
non-native speakers are significantly more likely to
provide non -ed (i.e., irregular) past-tense forms
for novel verbs than native speakers. Both groups
are sensitive to sound similarities between new
words and existing words (i.e., are more likely to
provide irregular forms for novel words which sound
similar to existing irregulars). Among both natives
and non-natives, irregularizations are non-random;
that is, rather than presenting as truly irregular
inflectional strategies, they follow identifiable
sub-rules present in the highly frequent set of
irregular English verbs. Our results shed new light
on how native and non-native learners can affect
Language universals have long been attributed to an
innate Universal Grammar. An alternative explanation
states that linguistic universals emerged
independently in every language in response to
shared cognitive or perceptual biases. A
computational model has recently shown how this
could be the case, focusing on the paradigmatic
example of the universal properties of colour naming
patterns, and producing results in quantitative
agreement with the experimental data. Here we
investigate the role of an individual perceptual
bias in the framework of the model. We study how,
and to what extent, the structure of the bias
influences the corresponding linguistic universal
patterns. We show that the cultural history of a
group of speakers introduces population-specific
constraints that act against the pressure for
uniformity arising from the individual bias, and we
clarify the interplay between these two forces
Contact languages are born out of the non-trivial
interaction of two (or more) parent
languages. Nowadays, the enhanced possibility of
mobility and communication allows for a strong
mixing of languages and cultures, thus raising the
issue of whether there are any pure languages or
cultures that are unaffected by contact with
others. As with bacteria or viruses in biological
evolution, the evolution of languages is marked by
horizontal transmission; but to date no reliable
quantitative tools to investigate these phenomena
have been available. An interesting and well
documented example of contact language is the
emergence of creole languages, which originated in
the contacts of European colonists and slaves during
the 17th and 18th centuries in exogenous plantation
colonies of especially the Atlantic and Indian
Ocean. Here, we focus on the emergence of creole
languages to demonstrate a dynamical process that
mimics the process of creole formation in American
and Caribbean plantation ecologies. Inspired by the
Naming Game (NG), our modeling scheme incorporates
demographic information about the colonial
population in the framework of a non-trivial
interaction network including three populations:
Europeans, Mulattos/Creoles, and Bozal slaves. We
show how this sole information makes it possible to
discriminate territories that produced modern
creoles from those that did not, with a surprising
accuracy. The generality of our approach provides
valuable insights for further studies on the
emergence of languages in contact ecologies as well
as to test specific hypotheses about the peopling
and the population structures of the relevant
territories. We submit that these tools could be
relevant to addressing problems related to contact
phenomena in many cultural domains: e.g., emergence
of dialects, language competition and hybridization,
Long-distance dependencies are notoriously diffi cult to analyze in a formally explicit way because they involve constituents that seem to have been extracted from their canonical position in an utterance. The most widespread solution is to identify a GAP at an EXTRACTION SITE and to communicate information about that gap to its FILLER, as in What_FILLER did you see_GAP? This paper rejects the filler?gap solution and proposes a cognitive-functional alternative in which long-distance dependencies spontaneously emerge as a side eff ect of how grammatical constructions interact with each other for expressing diff erent conceptualizations. The proposal is supported by a computational implementation in Fluid Construction Grammar that works for both parsing and production.
Computational experiments in cultural language evolution are important because they help to reveal the cognitive mechanisms and cultural processes that continuously shape and reshape the structure and knowledge of language. However, understanding the intricate relations between these mechanisms and processes can be a daunting challenge. This paper proposes to recruit the concept of fitness landscapes from evolutionary biology and computer science for visualizing the ?linguistic fitness? of particular language systems. Through a case study on the German paradigm of definite articles, the paper shows how such landscapes can shed a new and unexpected light on non-trivial cases of language evolution. More specifically, the case study falsifies the widespread assumption that the paradigm is the accidental by-product of linguistic erosion. Instead, it has evolved to optimize the cognitive and perceptual resources that language users employ for achieving successful communication.
Human languages are rule governed, but almost
invariably these rules have exceptions in the form
of irregularities. Since rules in language are
efficient and productive, the persistence of
irregularity is an anomaly. How does irregularity
linger in the face of internal (endogenous) and
external (exogenous) pressures to conform to a rule?
Here we address this problem by taking a detailed
look at simple past tense verbs in the Corpus of
Historical American English. The data show that the
language is open, with many new verbs entering. At
the same time, existing verbs might tend to
regularize or irregularize as a consequence of
internal dynamics, but overall, the amount of
irregularity sustained by the language stays roughly
constant over time. Despite continuous vocabulary
growth, and presumably, an attendant increase in
expressive power, there is no corresponding growth
in irregularity. We analyze the set of irregulars,
showing they may adhere to a set of minority rules,
allowing for increased stability of irregularity
over time. These findings contribute to the debate
on how language systems become rule governed, and
how and why they sustain exceptions to rules,
providing insight into the interplay between the
emergence and maintenance of rules and exceptions in
Fluid Construction Grammar (FCG) is an open-source computational grammar formalism that is becoming increasingly popular for studying the history and evolution of language. This demonstration shows how FCG can be used to operationalise the cultural processes and cognitive mechanisms that underly language evolution and change.
Construction Grammar has reached a stage of maturity where many researchers are looking for an explicit formal grounding of their work. Recently, there have been exciting developments to cater for this demand, most notably in Sign-Based Construction Grammar (SBCG) and Fluid Construction Grammar (FCG). Unfortunately, like playing a music instrument, the formalisms used by SBCG and FCG take time and effort to master, and linguists who are unfamiliar with them may not always appreciate the far-reaching theoretical consequences of adopting this or that approach. This paper undresses SBCG and FCG to their bare essentials, and offers a linguist-friendly comparison that looks at how both approaches define constructions, linguistic knowledge and language processing.
The German definite article paradigm, which is notorious for its case syncretism, is widely considered to be the accidental by-product of diachronic changes. This paper argues instead that the evolution of the paradigm has been motivated by the needs and constraints of language usage. This hypothesis is supported by experiments that compare the current paradigm to its Old High German ancestor (OHG; 900?1100ad) in terms of linguistic assessment criteria such as cue reliability, processing efficiency and ease of articulation. Such a comparison has been made possible by ?bringing back alive? the OHG system through a computational reconstruction
in the form of a processing model.The experiments demonstrate that syncretism has made the New High German system more efficient for processing, pronunciation and perception than its historical predecessor, without harming the language?s strength at disambiguating utterances.
The naming game (NG) describes the agreement
dynamics of a population of N agents interacting
locally in pairs leading to the emergence of a
shared vocabulary. This model has its relevance in
the novel fields of semiotic dynamics and
specifically to opinion formation and language
evolution. The application of this model ranges from
wireless sensor networks as spreading algorithms,
leader election algorithms to user-based social
tagging systems. In this paper, we introduce the
concept of overhearing (i.e., at every time step of
the game, a random set of N-delta individuals are
chosen from the population who overhear the
transmitted word from the speaker and accordingly
reshape their inventories). When delta = 0 one
recovers the behavior of the original NG. As one
increases delta, the population of agents reaches a
faster agreement with a significantly low-memory
requirement. The convergence time to reach global
consensus scales as log N as delta approaches
1. Copyright (C) EPLA, 2013
Despite centuries of research, the origins of grammatical case are more mysterious than ever. This paper addresses some unanswered questions through language game experiments in which a multi-agent population self-organizes a morphosyntactic case system. The experiments show how the formal part of grammatical constructions may pressure such emergent systems to become more economical.
Case has fascinated linguists for centuries without however revealing its most important secrets. This paper offers operational explanations for case through language game experiments in which autonomous agents describe real-world events to each other. The experiments demonstrate (a) why a language may develop a case system, (b) how a population can self-organize a case system, and (c) why and how an existing case system may take on new functions in a language.
German case syncretism is often assumed to be the accidental by-product of historical development. This paper contradicts this claim and argues that the evolution of German case is driven by the need to optimize the cognitive effort and memory required for processing and interpretation. This hypothesis is supported by a novel kind of computational experiments that reconstruct and compare attested variations of the German definite article paradigm. The experiments show how the intricate interaction between those variations and the rest of the German ?linguistic landscape? may direct language change.
Linguistic utterances are full of errors and novel expressions, yet linguistic communication is remarkably robust. This paper presents a double-layered architecture for open-ended language processing, in which ?diagnostics? and ?repairs? operate on a meta-level for detecting and solving problems that may occur during habitual processing on a routine layer. Through concrete operational examples, this paper demonstrates how such an architecture can directly monitor and steer linguistic processing, and how language can be embedded in a larger cognitive system.
Almost all languages in the world have a way to formulate commands. Commands specify actions that the body should undertake (such as “stand up”), possibly involving other objects in the scene (such as “pick up the red block”). Action language involves various competences, in particular (i) the ability to perform an action and recognize which action has been performed by others (the so-called mirror problem), and (ii) the ability to identify which objects are to participate in the action (e.g. “the red block” in “pick up the red block”) and understand what role objects play, for example whether it is the agent or undergoer of the action, or the patient or target (as in “put the red block on top of the green one”). This chapter describes evolutionary language game experiments exploring how these competences originate, can be carried out and acquired, by real robots, using evolutionary language games and a whole systems approach.
Cognitive linguistics has reached a stage of maturity where many researchers are looking for an explicit formal grounding of their work. Unfortunately, most current models of deep language processing incorporate assumptions from generative grammar that are at odds with the cognitive movement in linguistics. This demonstration shows how Fluid Construction Grammar (FCG), a fully operational and bidirectional unification-based grammar formalism, caters for this increasing demand. FCG features many of the tools that were pioneered in computational linguistics in the 70s-90s, but combines them in an innovative way. This demonstration highlights the main differences between FCG and related formalisms.
This chapter introduces a new experimental paradigm for studying issues in the grounding of language and robots, and the integration of all aspects of intelligence into a single system. The paradigm is based on designing and implementing artificial agents so that they are able to play language games about situations they perceive and act upon in the real world. The agents are not pre-programmed with an existing language but with the necessary cognitive functions to self-organize communication systems from scratch, to learn them from human language users if there are sufficiently frequent interactions, and to participate in the on-going cultural evolution of language.
This chapter introduces very briefly the framework and tools for lexical and grammatical processing that have been used in the evolutionary language game experiments reported in this book. This framework is called Fluid Construction Grammar (FCG) because it rests on a constructional approach to language and emphasizes flexible grammar application. Construction grammar organizes the knowledge needed for parsing or producing utterances in terms of bi-directional mappings between meaning and form. In line with other contemporary linguistic formalisms, FCG uses feature structures and unification and includes several innovations which make the formalism more adapted to implement flexible and robust language processing systems on real robots. This chapter is an introduction to the formalism and how it is used in processing.
This chapter introduces the computational infrastructure that is used to bridge the gap between results from sensorimotor processing and language. It consists of a system called Incremental Recruitment Language (IRL) that is able to configure a network of cognitive operations to achieve a particular communicative goal. IRL contains mechanisms for finding such networks, chunking subnetworks for more efficient later reuse, and completing partial networks (as possibly derived from incomplete or only partially understood sentences).
This chapter describes key aspects of a visual perception system as a key component for language game experiments on physical robots. The vision system is responsible for segmenting the continuous flow of incoming visual stimuli into segments and computing a variety of features for each segment. This happens by a combination of bottom-up way processing that work on the incoming signal and top-down processing based on expectations about what was seen before or objects stored in memory. This chapter consists of two parts. The first one is concerned with extracting and maintaining world models about spatial scenes, without any prior knowledge of the possible objects involved. The second part deals with the recognition of gestures and actions which establish the joint attention and pragmatic feedback that is an important aspect of language games.
This chapter explores a semantics-oriented approach to the origins of syntactic structure. It reports on preliminary experiments whereby speakers introduce hierarchical constructions and grammatical markers to express which conceptualization strategy hearers are supposed to invoke. This grammatical information helps hearers to avoid semantic ambiguity or errors in interpretation. A simulation study is performed for spatial grammar using robotic agents that play language games about objects in their shared world. The chapter uses a reconstruction of a fragment of German spatial language to identify the niche of spatial grammar, and then reports on acquisition and formation experiments in which agents seeded with a `pidgin German’ without grammar are made to interact until rudiments of hierarchical structure and grammatical marking emerge.
Grounding language in sensorimotor spaces is an important and difficult task. In order, for robots to be able to interpret and produce utterances about the real world, they have to link symbolic information to continuous perceptual spaces. This requires dealing with inherent vagueness, noise and differences in perspective in the perception of the real world. This paper presents two case studies for spatial language and quantification that show how cognitive operations – the building blocks of grounded procedural semantics – can be efficiently grounded in sensorimotor spaces.
Russian requires speakers of the language to conceptualize events using temporal language devices such as Aktionsarten and aspect, which relate to particular profiles and characteristics of events such as whether the event just started, whether it is ongoing or it is a repeated event. This chapter explores how such temporal features of events can be processed and learned by robots through grounded situated interactions. We use a whole systems approach, tightly integrating perception, conceptualization grammatical processing and learning and demonstrate how a system of Aktionsarten can be acquired.
Basic postures such as sit, stand and lie are ubiquitous in human interaction. In order to build robots that aid and support humans in their daily life, we need to understand how posture categories can be learned and recognized. This paper presents an unsupervised learning approach to posture recognition for a biped humanoid robot. The approach is based on Slow Feature Analysis (SFA), a biologically inspired algorithm for extracting slowly changing signals from signals varying on a fast time scale. Two experiments are carried out: First, we consider the problem of recognizing static postures in a multimodal sensory stream which consists of visual and proprioceptive stimuli. Secondly, we show how to extract a low-dimensional representation of the sensory state space which is suitable for posture recognition in a more complex setting. We point out that the beneficial performance of SFA in this task can be related to the fact that SFA computes manifolds which are used in robotics to model invariants in motion and behavior. Based on this insight, we also propose a method for using SFA components for guided exploration of the state space.
This chapter introduces the modular humanoid robot Myon, covering its mechatronical design, embedded low-level software, distributed processing architecture, and the complementary experimental environment. The Myon humanoid is the descendant of various robotic hardware platforms which have been built over the years and therefore combines the latest research results on the one hand, and the expertise of how a robot has to be built for experiments on embodiment and language evolution on the other hand. In contrast to many other platforms, the Myon humanoid can be used as a whole or in parts. Both the underlying architecture and the supportive application software allow for ad hoc changes in the experimental setup.
This chapter studies how basic spatial categories such as left-right, front-back, far-near or north-south can emerge in a population of robotic agents in co-evolution with terms that express these categories. It introduces various language strategies and tests them first in reconstructions of German spatial terms, then in acquisition experiments to demonstrate the adequacy of the strategy for learning these terms, and finally in language formation experiments showing how a spatial vocabulary and the concepts expressed by it can emerge in a population of embodied agents from scratch.
This chapter investigates how a vocabulary for talking about body actions can emerge in a population of grounded autonomous agents instantiated as humanoid robots. The agents play a Posture Game in which the speaker asks the hearer to take on a certain posture. The speaker either signals success if the hearer indeed performs an action to achieve the posture or he shows the posture himself so that the hearer can acquire the name. The challenge of emergent body language raises not only fundamental issues in how a perceptually grounded lexicon can arise in a population of autonomous agents but also more general questions of human cognition, in particular how agents can develop a body model and a mirror system so that they can recognize actions of others as being the same as their own.
This chapter explores a possible language strategy for verbalizing aspect: the encoding of Aktionsarten by means of morphological markers. Russian tense-aspect system is used as a model. We first operationalize this system and reconstruct the learning operators needed for acquiring it. Then we perform a first language formation experiment in which a novel system of Aktionsarten emerges and gets coordinated between the agents, driven by a need for higher expressivity.
Language change is increasingly recognized as one of the most crucial sources of evidence for understanding human cognition. Unfortunately, despite sophisticated methods for documenting which changes have taken place, the question of why languages evolve over time remains open for speculation. This paper presents a
novel research method that addresses this issue by combining agent-based experiments with deep language processing, and demonstrates the approach through a case study on German definite articles. More specifically, two populations of autonomous agents are equipped with a model of Old High German (500?1100 AD) and Modern High
German definite articles respectively, and a set of self-assessment criteria for evaluating their own linguistic performances. The experiments show that inefficiencies detected in the grammar by the Old High German agents correspond to grammatical forms that have actually undergone the most important changes in the German language.
The results thus suggest that the question of language change can be reformulated as an optimization problem in which language users try to achieve their communicative goals while allocating their cognitive resources as efficiently as possible.
The question how a shared vocabulary can arise in a multi-agent population despite the fact that each agent autonomously invents and acquires words has been solved. The solution is based on alignment: Agents score all associations between words and meanings in their lexicons and update these preference scores based on communicative success. A positive feedback loop between success and use thus arises which causes the spontaneous self-organization of a shared lexicon. The same approach has been proposed for explaining how a population can arrive at a shared grammar, in which we get the same problem of variation because each agent invents and acquires their own grammatical constructions. However, a problem arises if constructions reuse parts that can also exist on their own. This happens particularly when frequent usage patterns, which are based on compositional rules, are stored as such. The problem is how to maintain systematicity. This paper identifies this problem and proposes a solution in the form of multilevel alignment. Multilevel alignment means that the updating of preference scores is not restricted to the constructions that were used in the utterance but also downward and upward in the subsumption hierarchy.
Becoming a proficient speaker of a language requires more than just learning a set of words and grammar rules, it also implies mastering the ways in which speakers of that language typically innovate: stretching the meaning of words, introducing new grammatical constructions, introducing a new category, and so on. This paper demonstrates that such meta-knowledge can be represented and applied by reusing similar representations and processing techniques as needed for routine linguistic processing, which makes it possible that language processing makes use of computational reflection.
The fascinating question of the origins and evolution of language has been drawing a lot of attention recently, not only from linguists, but also from anthropologists, evolutionary biologists, and brain scientists. This groundbreaking book explores the cultural side of language evolution. It proposes a new overarching framework based on linguistic selection and self-organization and explores it in depth through sophisticated computer simulations and robotic experiments. Each case study investigates how a particular type of language system can emerge in a population of language game playing agents and how it can continue to evolve in order to cope with changes in ecological conditions. Case studies cover on the one hand the emergence of concepts and words for proper names, color terms, names for bodily actions, spatial terms and multi-dimensional words. The second set of experiments focuses on the emergence of grammar, specifically case grammar for expressing argument structure, functional grammar for expressing different uses of spatial relations, internal agreement systems for marking constituent structure, morphological expression of aspect, and quantifiers expressed as articles. The book is ideally suited as study material for an advanced course on language evolution and it will be of interest to anyone who wonders how human languages may have originated.
Written by leading international experts, this volume presents contributions establishing the feasibility of human language-like communication with robots. The book explores the use of language games for structuring situated dialogues in which contextualized language communication and language acquisition can take place. Within the text are integrated experiments demonstrating the extensive research which targets artificial language evolution. Language Grounding in Robots uses the design layers necessary to create a fully operational communicating robot as a framework for the text, focusing on the following areas: Embodiment; Behavior; Perception and Action; Conceptualization; Language Processing; Whole Systems Experiments. This book serves as an excellent reference for researchers interested in further study of artificial language evolution.
The lexicons of human languages organize their units
at two distinct levels. At a first combinatorial
level, meaningless forms (typically referred to as
phonemes) are combined into meaningful units
(typically referred to as morphemes). Thanks to
this, many morphemes can be obtained by relatively
simple combinations of a small number of
phonemes. At a second compositional level of the
lexicon, morphemes are composed into larger lexical
units, the meaning of which is related to the
individual meanings of the composing morphemes. This
duality of patterning is not a necessity for
lexicons and the question remains wide open
regarding how a population of individuals is able to
bootstrap such a structure and the evolutionary
advantages of its emergence. Here we address this
question in the framework of a multi-agents model,
where a population of individuals plays simple
naming games in a conceptual environment modeled as
a graph. We demonstrate that errors in communication
as well as a blending repair strategy, which
crucially exploits a shared conceptual
representation of the environment, are sufficient
conditions for the emergence of duality of
patterning, that can thus be explained in a pure
cultural way. Compositional lexicons turn out to be
faster to lead to successful communication than
purely combinatorial lexicons, suggesting that
meaning played a crucial role in the evolution of
One of the fundamental problems in cognitive science
is how humans categorize the visible color
spectrum. The empirical evidence of the existence of
universal or recurrent patterns in color naming
across cultures is paralleled by the observation
that color names begin to be used by individual
cultures in a relatively fixed order. The origin of
this hierarchy is largely unexplained. Here we
resort to multiagent simulations, where a population
of individuals, subject to a simple perceptual
constraint shared by all humans, namely the human
Just Noticeable Difference, categorizes and names
colors through a purely cultural negotiation in the
form of language games. We found that the time
needed for a population to reach consensus on a
color name depends on the region of the visible
color spectrum. If color spectrum regions are ranked
according to this criterion, a hierarchy with [red,
(magenta)-red], [violet], [green/yellow], [blue],
[orange], and [cyan], appearing in this order, is
recovered, featuring an excellent quantitative
agreement with the empirical observations of the
WCS. Our results demonstrate a clear possible route
to the emergence of hierarchical color categories,
confirming that the theoretical modeling in this
area has now attained the required maturity to make
significant contributions to the ongoing debates
concerning language universals.
All languages of the world have a way to talk about space and spatial relations of objects. Cross-culturally, immense variation in how people conceptualize space for language has been attested. Different spatial conceptualization strategies such as proximal, projective and absolute have been identified to underlie peoples conception of spatial reality. This paper argues that spatial conceptualization strategies are negotiated in a cultural process of linguistic selection. Conceptualization strategies originate in the cognitive capabilities of agents. The ecological conditions and the structure of the environment influence the conceptualization strategy agents invent and which corresponding system of lexicon and ontology of spatial relations is selected for. The validity of these claims is explored using populations of humanoid robots.
This paper compares two prominent approaches in artificial language evolution: Iterated Learning and Social Coordination. More specifically, the paper contrasts experiments in both approaches on how populations of artificial agents can autonomously develop a grammatical case marking system for indicating event structure (i.e. ?who does what to whom?). The comparison demonstrates that only the Social Coordination approach leads to a shared communication system in a multi-agent population. The paper concludes with an analysis and discussion of the results, and argues that Iterated Learning in its current form cannot explain the emergence of more complex natural language-like phenomena.
The paper surveys recent research on language evolution, focusing in particular on models of cultural evolution and how they are being developed and tested using agent-based computational simulations and robotic experiments. The key challenges for evolutionary theories of language are outlined and some example results are discussed, highlighting models explaining how linguistic conventions get shared, how conceptual frameworks get coordinated through language, and how hierarchical structure could emerge. The main conclusion of the paper is that cultural evolution is a much more powerful process that usually assumed, implying that less innate structures or biases are required and consequently that human language evolution has to rely less on genetic evolution.
This paper presents a design pattern for handling argument structure and offers a concrete operationalization of this pattern in Fluid Construction Grammar. Argument structure concerns the mapping between ?participant structure? (who did what to whom) and instances of ?argument realization? (the linguistic expression of participant structures). This mapping is multilayered and indirect, which poses great challenges for grammar design. In the proposed design pattern, lexico-phrasal constructions introduce their semantic and syntactic potential of linkage. Argument structure constructions, then, select from this potential the values that they require and implement the actual linking.
This paper illustrates the use of ?feature matrices?, a technique for handling ambiguity and feature indeterminacy in feature structure grammars using unification as the single mechanism for processing. Both phenomena involve forms that can be mapped onto multiple, often conflicting values. This paper illustrates their respective challenges through German case agreement, which has become the litmus test for demonstrating how well a grammar formalism deals with multifunctionality. After reviewing two traditional solutions, the paper demonstrates how complex grammatical categories can be represented as feature matrices instead of single-valued features. Feature matrices allow a free flow of constraints on possible feature-values coming from any part of an utterance, and they postpone commitment to any particular value until sufficient constraints have been identified. All examples in this paper are operationalized in Fluid Construction Grammar, but the design principle can be extended to other unification-grammars as well.
Natural languages are fluid. New conventions may arise and there is never absolute consensus in a population. How can human language users nevertheless have such a high rate of communicative success? And how do they deal with the incomplete sentences, false starts, errors and noise that is common in normal discourse? Fluidity, ungrammaticality and error are key problems for formal descriptions of language and for computational implementations of language processing because these seem to be necessarily rigid and mechanical. This chapter discusses how these issues are approached within the framework of Fluid Construction Grammar. Fluidity is not achieved by a single mechanism but through a combination of intelligent grammar design and flexible processing principles.
One of the key components for achieving flexible, robust, adaptive and open-ended language-based communication between humans and robots – or between robots and robots – is rich deep semantics. AI has a long tradition of work in the representation of knowledge, most of it within the logical tradition. This tradition assumes that an autonomous agent is able to derive formal descriptions of the world which can then be the basis of logical inference and natural language understanding or production. This paper outlines some difficulties with this logical stance and reports alternative research on the development of an ?embodied cognitive semantics? that is grounded in the world through a robot?s sensori-motor system and is evolutionary in the sense that the conceptual frameworks underlying language are assumed to be adapted by agents in the course of dialogs and thus undergo constant change.
This chapter presents an operational grammar for German spatial language, in particular German locative phrases, as a case study for processing distributed information. It investigates the complex interplay of syntactic phenomena and spatial semantics, with a specific emphasis on efficient processing of syntactic indeterminacy and semantic ambiguity. Since FCG applies constructions in a sequence one after the other, the main challenge lies in mutual dependencies between constructions, that is, some constructions require pieces of information in order to make decisions that are only later on provided by other constructions. We present solutions and design patterns for dealing with these processing issues, which all have in common the strategy of postponing decisions as long as possible in processing until all the necessary information for making the decision is available.
This thesis contributes to our understanding of the origins of spatial language by carrying out language game experiments with artificial agents instantiated as humanoid robots. It tests the theory of language evolution by linguistic selection, which states that language emerges through a cultural process based on the recruitment of various cognitive capacities in the service of language. Agents generate possible paradigmatic choices in their language systems and explore different language strategies. Which ones survive and dominate depends on linguistic selection criteria, such as expressive adequacy with respect to the ecological challenges and conditions in the environment, minimization of cognitive effort, and communicative success. To anchor this case study in empirical phenomena, the thesis reconstructs the syntax and semantics of German spatial language, in particular German locative phrases. Syntactic processing is organized using Fluid Construction Grammar (FCG), a computational formalism for representing linguistic knowledge. For the semantics the thesis focusses in particular on proximal, projective and absolute spatial categories as well as perspective, perspective reversal and frame of reference. The semantic investigations use the perspective of Embodied Cognitive Semantics. The spatial semantics is grounded in the sensorimotor experiences of the robot and made compositional by using the Incremental Recruitment Language (IRL) developed for this purpose. The complete reconstructed system allows humanoid robots to communicate successfully and efficiently using the German locative system and provides a performance base line. The reconstruction shows that the computational formalisms, i.e. FCG and IRL, are sufficient for tackling complex natural language phenomena. Moreover, the reconstruction efforts reveal the tight interaction of syntax and semantics in German locative phrases. The second part of the thesis concentrates on the evolution of spatial language. First the focus is on the formation and acquisition of spatial language by proposing strategies in the form of invention, adoption, and alignment operators. The thesis shows the adequacy of these strategies in acquisition experiments in which some agents act as learners and others as tutors. It shows next in language formation experiments that these strategies are sufficient to allow a population to self-organize a spatial language system from scratch. The thesis continues by studying the origins and competition of language strategies. Different conceptual strategies are considered and studied systematically, particularly in relation to the properties of the environment, for example, whether a global landmark is available. Different linguistic strategies are studied as well, for instance, the problem of choosing a particular reference object on the scene can be solved by the invention of markers, which allows many different reference objects, or by converging to a standard single reference object, such as a global landmark. The thesis demonstrates that the theory of language evolution by linguistic selection leads to operational experiments in which artificial agents self-organize semantically rich and syntactically complex language. Moreover, many issues in cognitive science, ranging from perception and conceptualization to language processing, had to be dealt with to instantiate this theory, so that this thesis contributes not only to the study of language evolution but to the investigation of the cognitive bases of spatial language as well.
Construction Grammar is enthusiastically embraced by a growing group of linguists who find it a natural way to formulate their analyses. But so far there is no widespread formalization of construction grammar with a solid computational implementation. Fluid Construction Grammar attempts to fill this gap. It is a fully operational computational framework capturing many key concepts in construction grammar. The present book is the first extensive publication describing this framework. In addition to general introductions, it gives a number of concrete examples through a series of linguistically challenging case studies, including phrase structure, case grammar, and modality. The book is suited both for linguists who want to know what Fluid Construction Grammar looks like and for computational linguists who may want to use this computational framework for their own experiments or applications.
Pronouns form a particularly interesting part-of-speech for evolutionary linguistics because
their development is often lagging behind with respect to other changes in their language. Many
hypotheses on pronoun evolution exist ? both for explaining their initial resilience to change as
well as for why they eventually cave in to evolutionary pressures ? but so far, no one has proposed
a formal model yet that operationalizes these explanations in a unified theory. This paper
therefore presents a computational model of pronoun evolution in a multi-agent population;
and argues that pronoun evolution can best be understood as an interplay between the level
of language strategies, which are the procedures for learning, expanding and aligning particular
features of language, and the level of the specific language systems that instantiate these
strategies in terms of concrete words, morphemes and grammatical structures. This claim is
supported by a case study on Spanish pronouns, which are currently undergoing an evolution
from a case- to a referential-based system, the latter of which there exist multiple variations
(which are called le�smo, la�smo and lo�smo depending on the type of change).
Artificial agents trying to achieve communicative goals in situated interactions in the real-world need powerful computational systems for conceptualizing their environment. In order to provide embodied artificial systems with rich semantics reminiscent of human language complexity, agents need ways of both conceptualizing complex compositional semantic structure and actively reconstructing semantic structure, due to uncertainty and ambiguity in transmission. Furthermore, the systems must be open-ended and adaptive and allow agents to adjust their semantic inventories in order to reach their goals. This paper presents recent progress in modeling open-ended, grounded semantics through a unified software system that addresses these problems.
Over the past several decades, psycholinguists have gained countless insights into the process of child language acquisition. Can these findings be used for the development of language competence in autonomous artificial systems? This paper reports on our attempt to apply insights from developmental psychology in order to enable artificial systems to acquire language. We consider a comprehensive chain of computational processes, starting from conceptualization and extending through language generation and interpretation, and show how they can be intertwined to allow for acquisition of complex aspects of grammar.
Semantic maps have offered linguists an appealing and empirically rooted methodology for describing recurrent structural patterns in language development and the multifunctionality of grammatical categories. Although some researchers argue that semantic maps are universal and given, others provide evidence that there are no fixed or universal maps. This paper takes the position that semantic maps are a useful way to visualize the grammatical evolution of a language (particularly the evolution of semantic structuring) but that this grammatical evolution is a consequence of distributed processes whereby language users shape and reshape their language. So it is a challenge to find out what these processes are and whether they indeed generate the kind of semantic maps observed for human languages. This work takes a design stance towards the question of the emergence of linguistic structure and investigates how grammar can be formed in populations of autonomous artificial ?agents? that play ?language games? with each other about situations they perceive through a sensori-motor embodiment. The experiments reported here investigate whether semantic maps for case markers could emerge through grammaticalization processes without the need for a universal conceptual space.
Grammatical agreement is one of the most puzzling aspects found in natural language. Its acquisition requires intensive linguistic exposure and capacities to deal with outliers that break regular patterns. Other than relying on statistical methods to deal with agreement in a computational application, this paper demonstrates how agreement can be learned by artificial agents in a simulated environment in such a way that the openendedness of natural language can be captured by their language processing mechanisms.
In this paper we test the dominant paradigm for modeling the semantics of determined noun phrases called Generalized Quantifier Theory in embodied interactions with robots. We contrast the traditional approach with a new approach, called Clustering Determination, which is heavily inspired by research on grounding of sensorimotor categories, and we show that our approach performs better in noisy, real world, referential communication.
This paper presents a software system that integrates different computational paradigms to solve cognitive tasks of different levels. The system has been employed to empower research on very different platforms ranging from simple two-wheeled structures with only a few cheap sensors, to complex two-legged humanoid robots, with many actuators, degrees of freedom and sensors. It is flexible and adjustable enough to be used in part or as a whole, to target different research domains
PROJECTS and questions, including Evolutionary Robotics, RoboCup and Artificial Language Evolution on Autonomous Robots (ALEAR, an EU funded cognitive systems project). In contrast to many other frameworks, the system is such that researchers can quickly adjust the system to different problems and platforms, while allowing maximum reuse of components and abstractions, separation of concerns and extensibility.
The empirical evidence that human color
categorization exhibits some universal patterns
beyond superficial discrepancies across different
cultures is a major breakthrough in cognitive
science. As observed in the World Color Survey
(WCS), indeed, any two groups of individuals develop
quite different categorization patterns, but some
universal properties can be identified by a
statistical analysis over a large number of
populations. Here, we reproduce the WCS in a
numerical model in which different populations
develop independently their own categorization
systems by playing elementary language games. We
find that a simple perceptual constraint shared by
all humans, namely the human Just Noticeable
Difference (JND), is sufficient to trigger the
emergence of universal patterns that unconstrained
cultural interaction fails to produce. We test the
results of our experiment against real data by
performing the same statistical analysis proposed to
quantify the universal tendencies shown in the WCS
[Kay P & Regier T. (2003) Proc. Natl. Acad. Sci. USA
100: 9085-9089], and obtain an excellent
quantitative agreement. This work confirms that
synthetic modeling has nowadays reached the maturity
to contribute significantly to the ongoing debate in
Aspect is undoubtedly the most capricious grammatical category of the Russian language. It has often been asserted as a mystery accessible only to native speakers, leaving all the others lost in its apparently infinite clutter. Recent work in cognitive linguistics has tried to bring order to the seeming chaos of the Russian aspectual system. But these approaches have not been operationalized so far. This paper demonstrates how the aspectual derivation of Russian verbs can be handled successfully with Fluid Constructional Grammar, a computational formalism recently developed for the representation and processing of constructions.
In this paper, we propose a concrete operationalization which incorporates data from the FrameNet database into Fluid Construction Grammar, currently the only computational implementation of construction grammar that can achieve both production and parsing using the same set of constructions. As a proof of concept, we selected an annotated sentence from the FrameNet database and transcribed its frame annotation analysis into an FCG grammar. The paper illustrates the proposed constructions and discusses the value and results of these formalization efforts.
In this paper we demonstrate (1) how a group of embodied artificial agents can learn to construct abstract conceptual representations of body postures from their continuous sensorimotor interaction with the environment, (2) how they can metaphorically extend these bodily concepts to visual experiences of external objects and (3) how they can use their acquired embodied meanings for self-organizing a communication system about postures and objects. For this, we endow the agents with cognitive mechanisms and structures that are instantiations of specific ideas in cognitive linguistics (namely image schema theory) about how humans relate motor and visual space. We show that the agents are indeed able to perform well in the task and thus the experiment offers a concrete operationalization of these theories and increases their explanatory power.
Open-ended language communication remains an enormous challenge for autonomous robots. This paper argues that the notion of a language strategy is the appropriate vehicle for addressing this challenge. A language strategy packages all the procedures that are necessary for playing a language game. We present a specific example of a language strategy for playing an Action Game in which one robot asks another robot to take on a body posture (such as stand or sit), and show how it effectively allows a population of agents to self-organise a perceptually grounded ontology and a lexicon from scratch, without any human intervention. Next, we show how a new language strategy can arise by exaptation from an existing one, concretely, how the body posture strategy can be exapted to a strategy for playing language games about the spatial position of objects (as in ?the bottle stands on the table?).
Colour naming games are idealised communicative interactions within a population of artificial agents in which a speaker uses a single colour term to draw the attention of a hearer to a particular object in a shared context. Through a series of such games, a colour lexicon can be developed that is sufficiently shared to allow for successful communication, even when the agents start out without any predefined categories. In previous models of colour naming games, the shared context was typically artificially generated from a set of colour stimuli and both agents in the interaction perceive this environment in an identical way. In this paper, we investigate the dynamics of the colour naming game in a robotic setup in which humanoid robots perceive a set of colourful objects from their own perspective. We compare the resulting colour ontologies to those found in human languages and show how these ontologies reflect the environment in which they were developed.
This paper is part of an ongoing research program to understand the cognitive and functional bases for the origins and evolution of spatial language. Following a cognitive-functional approach, we first investigate the cross-linguistic variety in spatial language, with special attention for spatial perspective. Based on this language-typological data, we hypothesize which cognitive mechanisms are needed to explain this variety and argue for an interdisciplinary approach to test these hypotheses. We then explain how experiments in artificial language evolution can contribute to that and give a concrete example.
This paper shows how experiments on artificial language evolution can provide highly relevant results for important debates in linguistic theories. It reports on a series of experiments that investigate how semantic roles can emerge in a population of artificial embodied agents and how these agents can build a network of constructions. The experiment also includes a fully operational implementation of how event-specific participant-roles can be fused with the semantic roles of argument-structure constructions and thus contributes to the linguistic debate on how the syntax-semantics interface is organized.
Learning the meanings of words requires coping with referential uncertainty ? a learner hearing a novel word cannot be sure which aspects or properties of the referred object or event comprise the meaning of the word. Data from developmental psychology suggest that human learners grasp the important aspects of many novel words after just a few exposures, a phenomenon known as fast mapping. Traditionally, word learning is viewed as a mapping task, in which the learner has to map a set of forms onto a set of pre-existing concepts. We criticise this approach and argue instead for a flexible nature of the coupling between form and meanings as a solution to the problem of referential uncertainty. We implemented and tested the model in populations of humanoid robots that play situated language games about objects in their shared environment. Results show that the model can handle an exponential increase in uncertainty and allows scaling towards very large meaning spaces, while retaining the ability to grasp an operational meaning almost instantly for a great number of words. In addition, the model captures some aspects of the flexibility of form-meaning associations found in human languages. Meanings of words can shift between being very specific (names) and general (e.g. ?small?). We show that this specificity is biased not by the model itself but by the distribution of object properties in the world.
Humans maintain a body image of themselves, which plays a central role in controlling bodily movement, planning action, recognising and naming actions performed by others, and requesting or executing commands. This paper explores through experiments with autonomous humanoid robots how such a body image could form. Robots play a situated embodied language game called the Action Game in which they ask each other to perform bodily actions. They start without any prior inventory of names, without categories for visually recognising body movements of others, and without knowing the relation between visual images of motor behaviors carried out by others and their own motor behaviors. Through diagnostic and repair strategies carried out within the context of action games, they progressively self-organise an effective lexicon as well as bi-directional mappings between the visual and the motor domain. The agents thus establish and continuously adapt networks linking perception, body representation, action, and language.
This chapter briefly discusses the issues of symbols, meanings, and embodiment. It explains the solution to the symbol grounding problem. It illustrates the ingredients that are employed in the experiments about language emergence using a specific example of a color guessing game. It argues that these experiments show that there is an effective solution to the symbol grounding problem. The objective test for this claim is in the increased success of agents in the language games.
Exploratory activities seem to be intrinsically rewarding for children and crucial for their cognitive development. Can a machine be endowed with such an intrinsic motivation system? This is the question we study in this paper, presenting a number of computational systems that try to capture this drive towards novel or curious situations. After discussing related research coming from developmental psychology, neuroscience, developmental robotics, and active learning, this paper presents the mechanism of Intelligent Adaptive Curiosity, an intrinsic motivation system which pushes a robot towards situations in which it maximizes its learning progress. This drive makes the robot focus on situations which are neither too predictable nor too unpredictable, thus permitting autonomous mental development. The complexity of the robot’s activities autonomously increases and complex developmental sequences self-organize without being constructed in a supervised manner. Two experiments are presented illustrating the stage-like organization emerging with this mechanism. In one of them, a physical robot is placed on a baby play mat with objects that it can learn to manipulate. Experimental results show that the robot first spends time in situations which are easy to learn, then shifts its attention progressively to situations of increasing difficulty, avoiding situations in which nothing can be learned. Finally, these various results are discussed in relation to more complex forms of behavioral organization and data coming from developmental psychology Keywords: active learning, autonomy, behavior, complexity, curiosity, sensorimotor development, cognitive development, developmental trajectory, epigenetic robotics, intrinsic motivation, learning, reinforcement learning, values.
Language can be viewed as a complex adaptive system which is continuously shaped and reshaped by the actions of its users as they try to solve communicative problems. To maintain coherence in the overall system, different language elements (sounds, words, grammatical constructions) compete with each other for global acceptance. This paper examines what happens when a language system uses systematic structure, in the sense that certain meaning-form conventions are themselves parts of larger units. We argue that in this case multi-level selection occurs: at the level of elements (e.g. tense affixes) and at the level of larger units in which these elements are used (e.g. phrases). Achieving and maintaining linguistic coherence in the population under these conditions is non-trivial. This paper shows that it is nevertheless possible when agents take multiple levels into account both for processing meaning-form associations and for consolidating the language inventory after each interaction.
This paper presents computational experiments that illustrate how one can precisely conceptualize language evolution as a Darwinian process. We show that there is potentially a wide diversity of replicating units and replication mechanisms involved in language evolution. Computational experiments allow us to study systemic properties coming out of populations of linguistic replicators: linguistic replicators can adapt to specific external environments; they evolve under the pressure of the cognitive constraints of their hosts, as well as under the functional pressure of communication for which they are used; one can observe neutral drift; coalitions of replicators may appear, forming higher level groups which can themselves become subject to competition and selection.
Our social behaviour has evolved primarily through contact with a limited number of other individuals. Yet as a species we exhibit uniformities on a global scale. This kind of emergent behaviour is familiar territory for statistical physicists.
What processes can explain how very large populations are able to converge on the use of a particular word or grammatical construction without global coordination? Answering this question helps to understand why new language constructs usually propagate along an S-shaped curve with a rather sudden transition towards global agreement. It also helps to analyse and design new technologies that support or orchestrate self-organizing communication systems, such as recent social tagging systems for the web. The article introduces and studies a microscopic model of communicating autonomous agents performing language games without any central control. We show that the system undergoes a disorder/order transition, going through a sharp symmetry breaking process to reach a shared set of conventions. Before the transition, the system builds up non-trivial scale-invariant correlations, for instance in the distribution of competing synonyms, which display a Zipf-like law. These correlations make the system ready for the transition towards shared conventions, which, observed on the timescale of collective behaviours, becomes sharper and sharper with system size. This surprising result not only explains why human language can scale up to very large populations but also suggests ways to optimize artificial semiotic dynamics.
What processes can explain how very large
populations are able to converge on the use of a
particular word or grammatical construction without
global coordination? Answering this question helps
to understand why new language constructs usually
propagate along an S-shaped curve with a rather
sudden transition towards global agreement. It also
helps to analyse and design new technologies that
support or orchestrate self-organizing communication
systems, such as recent social tagging systems for
the web. The article introduces and studies a
microscopic model of communicating autonomous agents
performing language games without any central
control. We show that the system undergoes a
disorder/order transition, going through a sharp
symmetry breaking process to reach a shared set of
conventions. Before the transition, the system
builds up non-trivial scale-invariant correlations,
for instance in the distribution of competing
synonyms, which display a Zipf-like law. These
correlations make the system ready for the
transition towards shared conventions, which,
observed on the timescale of collective behaviours,
becomes sharper and sharper with system size. This
surprising result not only explains why human
language can scale up to very large populations but
also suggests ways to optimize artificial semiotic
Sound is a medium used by humans to carry information. The existence of this kind of medium is a pre-requisite for language. It is organized into a code, called speech, which provides a repertoire of forms that is shared in each language community. This code is necessary to support the linguistic interactions that allow humans to communicate. How then may a speech code be formed prior to the existence of linguistic interactions? Moreover, the human speech code is characterized by several properties: speech is digital and compositional (vocalizations are made of units re-used systematically in other syllables); phoneme inventories have precise regularities as well as great diversity in human languages; all the speakers of a language community categorize sounds in the same manner, but each language has its own system of categorization, possibly very different from every other. How can a speech code with these properties form? These are the questions we will approach in the paper. We will study them using the method of the artificial. We will build a society of artificial agents, and study what mechanisms may provide answers. This will not prove directly what mechanisms were used for humans, but rather give ideas about what kind of mechanism may have been used. This allows us to shape the search space of possible answers, in particular by showing what is sufficient and what is not necessary. The mechanism we present is based on a low-level model of sensorymotor interactions. We show that the integration of certain very simple and non language-specific neural devices allows a population of agents to build a speech code that has the properties mentioned above. The originality is that it pre-supposes neither a functional pressure for communication, nor the ability to have coordinated social interactions (they do not play language or imitation games). It relies on the self-organizing properties of a generic coupling between perception and production both within agents, and on the interactions between agents.
The speech code is a vehicle of language: it defines a set of forms used by a community to carry information. Such a code is necessary to support the linguistic interactions that allow humans to communicate. How then may a speech code be formed prior to the existence of linguistic interactions? Moreover, the human speech code is discrete and compositional, shared by all the individuals of a community but different across communities, and phoneme inventories are characterized by statistical regularities. How can a speech code with these properties form? We try to approach these questions in the paper, using the ?methodology of the artificial?. We build a society of artificial agents, and detail a mechanism that shows the formation of a discrete speech code without pre-supposing the existence of linguistic capacities or of coordinated interactions. The mechanism is based on a low-level model of sensory-motor interactions. We show that the integration of certain very simple and non language-specific neural devices leads to the formation of a speech code that has properties similar to the human speech code. This result relies on the self-organizing properties of a generic coupling between perception and production within agents, and on the interactions between agents. The artificial system helps us to develop better intuitions on how speech might have appeared, by showing how self-organization might have helped natural selection to find speech.
This paper shows how a society of agents can self-organise a shared vocalisation system which is discrete, combinatorial, and has a form of primitive phonotactics, starting from holistic inarticulate vocalisations. The originality of the system is that: 1) it does not include any explicit pressure for communication; 2) agents do not possess capabilities of coordinated interactions, in particular they do not play language games; 3) agents possess no specific linguistic capacities; 4) initially there exist no convention that agent can use. As a consequence, the system shows how a primitive speech code may bootstrap in the absence of a communication system between agents, i.e. before the appearance of language.
This article proposes a number of models to examine through which mechanisms a population of autonomous agents could arrive at a repertoire of perceptually grounded categories that is sufficiently shared to allow successful communication. the models are inspired by the main approaches to human categorisation being discussed in the literature: nativism, empiricism, and culturalism. colour is taken as a case study. although we take no stance on which position is to be accepted as final truth with respect to human categorisation and naming, we do point to theoretical constraints that make each position more or less likely and we make clear suggestions on what the best engineering solution would be. specifically, we argue that the collective choice of a shared repertoire must integrate multiple constraints, including constraints coming from communication.
The interesting and deep commentaries on our target article reflect the continued high interest in the problem of colour categorisation and naming. clearly, colour remains for many cognitive science related disciplines a fascinating microworld in which some of the most fundamental issues for cognition and culture can be studied. although our target article took the stance of practically oriented engineers who are trying to find the best solution for orchestrating the self-organisation of communication systems in artificial agents, most commentators focus on the implications for cognitive science and we will do the same in our reply.
The computational and robotic modeling of language evolution is emerging as a new exciting subfield in cognitive science. The objective is to come up with precise operational models how communities of agents, equiped with a cognitive apparatus, a sensori-motor system, and a body, can arrive at shared grounded communication systems that have similar characteristics as human languages. Apart from its technological interest for building novel applications in the domain of human-robot or robot-robot interaction, this research is potentially relevant to the many disciplines interested in the origins and evolution of language.
As soon as we stop talking aloud, we seem to experience a kind of `inner voice’, a steady stream of verbal fragments expressing ongoing thoughts. What kind of information processing structures are required to explain such a phenomenon? Why would an �nner voice’ be useful? How could it have arisen? This paper explores these questions and reports briefly some computational experiments to help elucidate them.
Behavior-based robotics has always been inspired by earlier cybernetics work such as that of Grey Walter. It emphasises that intelligence can be achieved without the kinds of representations common in symbolic AI systems. The paper argues that such representations might indeed not be needed for many aspects of sensori-motor intelligence but become a crucial issue when bootstrapping to higher levels of cognition. It proposes a scenario in the form of evolutionary language games by which embodied agents develop situated grounded representations adapted to their needs and the conventions emerging in the population.
The paper reports on experiments with a population of visually grounded robotic agents capable of bootstrapping their own ontology and shared lexicon without prior design nor other forms of human intervention. The agents do so while playing a particular language game called the guessing game. We show that synonymy and ambiguity arise as emergent properties in the lexicon, due to the situated grounded character of the agent-environment interaction, but that there are also tendencies to dampen them so as to make the language more coherent and thus more optimal from the viewpoints of communicative success, cognitive complexity, and learnability.
We have been conducting large-scale public experiments with artificial robotic agents to explore what the necessary and sufficient prerequisites are for word-meaning pairs to evolve autonomously in a population of agents through a self-organized process. We focus not so much on the question of why language has evolved but rather on how. Our hypothesis is that when agents engage in particular interactive behaviors which in turn require specific cognitive structures, they automatically arrive at a language system. We study this topic by performing experiments based on artificial systems. One such experiment, known as the Talking Heads Experiment, employs a set of visually grounded autonomous robots into which agents can install themselves to play language games with each other.
Human sound systems are invariably phone- mically coded. Furthermore, phoneme invento- ries follow very particular tendancies. To ex- plain these phenomena, there existed so far three kinds of approaches : Chomskyan”/cognitive innatism, morpho-perceptual innatism and the more recent approach of language as a com- plex cultural system which adapts under the pres- sure of e�cient communication”. The two first approaches are clearly not satisfying, while the third, even if much more convincing, makes a lot of speculative assumptions and did not really bring answers to the question of phonemic cod- ing. We propose here a new hypothesis based on a low-level model of sensory-motor interac- tions. We show that certain very simple and non language-specific neural devices allow a popula- tion of agents to build signalling systems without any functional pressure. Moreover, these systems are phonemically coded. Using a realistic vowel articulatory synthesizer, we show that the inven- tories of vowels have striking similarities with hu- man vowel systems.