In preparation: Journal for Evolutionary Linguistics 1(1). 2010

Can Evolutionary Linguistics Become a Science?

Luc Steels∗1,2

1AI Laboratory, Vrije Universiteit Brussels, Pleinlaan 2 1050 Brussels, Belgium 2Sony Computer Science Laboratory, 6 Rue Amyot, 75005 Paris, France

Email: Luc Steels∗- [email protected];

∗Corresponding author

Abstract The paper introduces a methodology for developing evolutionary explanations for features of human natural languages. The methodology is inspired by Evolutionary Biology but maps the Darwinian selectionist framework to the cognitive and linguistic level. Point of departure is a particular language system found in human languages, such as tense and aspect marking or the expression of spatial relations. To remain empirically grounded, the language system is studied by means of a concrete realisation in a particular human language, for example case marking in Icelandic or subordinate clauses in German. The methodology suggests to first survey what kind of conceptual, lexical and grammatical structures we find in the chosen human language and to reconstruct the production and comprehension processes that can handle them. It proposes next to investigate the role of the language system in linguistic communication by designing a language game where it is beneficial. Then the learning, alignment, and expansion strategies that individuals need to acquire and self-organize a language system are reconstructed, and the processes by which these strategies could originate by the recruitment of cognitive mechanisms are identified. To explain why a language strategy has become adopted by a population, the final stage in the methodology studies the selectionist advantage of the language system by showing how it helps to increase communicative success, provide enough expressive power, and minimise cognitive effort. The paper gives a concrete example of this methodology, and discusses the general selectionist framework underlying it.

1 Introduction [7, 12, 13, 18, 21, 45, 49, 52]. There are now regular ‘Evolution of Language’ conferences that bring these How did human natural languages originate and why diverse fields together in an interdisciplinary dialog, and how do they continue to evolve? This fas- starting from 1996 [46], and there is a growing liter- cinating question has received a growing amount ature with monographs, collections of articles, and of interest over the past decade from many dis- publications in a wide range of journals. Out of all ciplines. We see contributions from anthropology, these activities, a new field of research is beginning archeology, historical linguistics, neurobiology, cog- to crystalise that does not approach the study of lan- nitive psychology, evolutionary biology, genetics, ar- guage evolution from the perspective of an existing tificial intelligence, and complex systems research

1 discipline but sees this research topic as its central like genetic drift, punctuated equilibria, exaptation, core. I believe this field is best called ‘Evolutionary epigenesis, genetic hitch-hiking, the Baldwin effect, Linguistics’. The goal of this paper is to reflect on transposons, and so on. However, recent research on the methodology that could most profitably be used speciation has seemed to confirm rather than further to develop this field and I will suggest that Evolu- question the role of natural selection in the origins tionary Biology can act as a source of inspiration. of species [51], and so I will mainly restrict the dis- Evolutionary Biology is one of the best developed cussion to that. areas of contemporary science and it acts as a glue We need a comparable general framework for that pulls together all other biological disciplines: Evolutionary Linguistics. One approach might be ethology, ecology, molecular biology, population dy- to apply the Darwinian proposal literally, in other namics, development, and genetics. As Dobzhansky words, to consider language to be a biological adap- famously put it: “Nothing in biology makes sense tation which has evolved through genetic evolution except in the light of evolution.” [28] Evolutionary and natural selection. This can only be the case Linguistics could possibly play a similar assembling if the following conditions are met: (i) There are role for the study of language and meaning, which dedicated phenotypic traits for language. It is usu- is currently fragmented into different subdisciplines ally argued that this is a highly specialized modular that hardly interact with each other. ‘language organ’ in the brain. (ii) There is a crit- Evolutionary Linguistics draws necessarily on ical role of ‘language genes’ for building these spe- many existing linguistic subdisciplines: Historical cialized brain structures. And (iii) greater linguistic linguistics produces data and insight in the actual aptitude leads to higher biological fitness, so that evolution of human languages. Cognitive linguis- the language genes spread and remain in the gene tics and construction grammar contribute studies pool. Some researchers are pursuing this biological on the rich semantics and pragmatics underlying hypothesis [6, 55, 57, 63, 90], but it is not the only human languages and how they are grammatically possibility. expressed. Computational linguistics provides con- It could also be that language evolution is a spe- crete models of syntactic parsing, production and cial case of cultural evolution based on the recruit- language learning. Developmental linguistics docu- ment of cognitive mechanisms that are also useful ments the pathways by which children or second lan- in other domains [20, 36, 53, 60, 79, 81, 93]. In that guage learners acquire language. And neurolinguis- case the possible analogy between biological evolu- tics makes concrete proposals about the neural basis tion and language evolution is at a more abstract of language, including the way the many cognitive level. The explanatory targets are not biological functions needed for language might have evolved. phenotypic traits but language features. We still ex- But Evolutionary Linguistics wants to go be- pect to see heredity, variation, and selection to am- yond these efforts by developing explanations of how plify those variants that are more adapted to a par- and why certain linguistic phenomena could have ticular set of challenges, but heredity will be based evolved, not only in terms of empirical findings of on memory and learning rather than genetic trans- actual human language evolution but also in terms mission. Variation will come from the unavoidable of formal models and systematic experiments. Thus divergence in linguistic behavior due to the different it helps to move linguistics from a descriptive en- development histories of the members of a popula- deavour to an explanatory science, from describing tion and performance deviations in actual discourse. features of human languages or their evolution to And selection will not be natural selection (i.e. based explaining why they are there. on biological fitness in an ecosystem) but linguistic The first reason why Evolutionary Biology has selection that amplifies the use of conceptualisations been so successful is because Darwin established a and linguistic constructions that lead to higher com- clear general framework to explain adaptation and municative success, greater expressive power, and speciation and to develop concrete evolutionary ex- less cognitive effort. In the rest of this paper, I will planations for phenotypic features of living beings, focus on this socio-cultural perspective on language such the color patterns of the wings of butterflies evolution, as I believe it has so far yielded the most or organs like the liver. There have been many re- explanatory power. finements and enhancements of this framework, sug- The epistemological status of Darwinian evolu- gesting supplementary causes or factors in evolution, tionary theory is quite different from other natural

2 science theories such as Newtonian mechanics, in the 2 Language systems and Language sense that it provides a framework for explaining Strategies phenomena but the details still need to be worked The first useful concept of evolutionary theory, as out for each specific case. For example, it does not currently conceptualized, is the distinction between directly explain why a particular species of butter- genotype and phenotype. A genotype consists of a flies has certain color patterns on the wings. It only network of genes that builds a concrete organism, suggests how we can go about finding such an expla- a phenotype, through a very complex developmental nation. In contrast, the equations of mechanics can process that typically integrates epigenetic factors be applied to any falling body by simply plugging in and environmental influences. The genotype also in- the relevant quantities and parameters. fluences the further operation, adaptation, and re- The explanatory framework that we can hope for pair of the phenotype as its life unfolds. in Evolutionary Linguistics might be more similar to An organism finds itself in a particular ecosys- that of Evolutionary Biology than that of physics, tem, a portion of the world that is populated by in the sense that there might not be a single cog- other species and has certain physical characteristics nitive mechanism, like recursion, or a single most like temperature or humidity. The organism gets a important function of early language, like hunting chance to play the game of life: It moves around in or courtship, or a single set of equations that could the ecosystem, fends off predators, finds food, co- explain the cultural dynamics of language. But operates with conspecifics, finds mates, gives birth, we may be able to establish a general explanatory cares for offspring, and so on. The organism’s behav- framework, comparable to the Darwinian selectionist ior yields success or failure for these various tasks. framework, which we can apply to understand how The cumulative success of an organism’s behaviors specific features of a language or a set of languages confers a certain selective fitness, a measure how well might have evolved. Every new case that can be the organism has been able to survive and reproduce treated satisfactorily gives us further evidence that within this specific ecosystem. Fitness is of course the overarching framework is right, just like every not only a function of genes. You might be unlucky successful evolutionary explanation of a biological and fall out of the nest or get run over by a truck trait gives more weight to the Darwinian approach. before you can show your worth. However, for ex- The fact that an evolutionary explanation for plaining long term adaptation and speciation, only a concrete phenotypic feature needs a lot of work traits that are under genetic influence play a role. could be seen as a weakness, but evolutionary biolo- The first question in transposing the Dar- gists have now developed a clear set of methodolog- winian selectionist framework to language is whether ical stages, an impressive battery of highly efficient we find something which is similar to the geno- and sophisticated tools to deal with each stage, as type/phenotype distinction and to the notion of fit- well as stringent evaluation criteria to attest whether ness. I argue in this section that this is the dis- a stage has been adequately dealt with. The sec- tinction between a language strategy and a language ond reason for the success of Evolutionary Biology system [72]. is that everybody in the field accepts and shares this methodology and hence that arguments can be set- tled by scientific means instead of appeal to author- 2.1 Language Subsystems ity or story telling. We need a comparable method- Every utterance combines a set of lexical and gram- ology for Evolutionary Linguistics and I will suggest matical features that deal with different aspects of that this can be similar to the one used by evolu- meaning and function. For example, the sentence tionary biologists. “Sophie walked home” introduces a number of en- The first part of the paper (section 2 and 3) de- tities (Sophie, home), an event (walk), its partici- velops the general theoretical framework for a theory pant roles (agent, target), and the time of the event of language evolution by linguistic selection. Section (past). There is a long tradition in linguistics to de- 4 then proposes a methodology for developing evolu- compose language into different subsystems. Each tionary explanations assuming this framework. And subsystem concerns a feature of utterances that cen- section 5 and 6 provide more detailed explanations ters around the same meaning or function, for exam- of this methodology as well as examples. ple, the tense system of English, the case grammar

3 of Latin, the reflexive and reciprocal pronoun sys- • A grammatical system for marking tem of Spanish, the classifier system of Swahili. It topic/comment structure (what is fore- is true that these subsystems or paradigms often in- ground/what is background), for example teract, in the sense that the same linguistic element through different sentence patterns (as the can have multiple functions, just like the same phe- field topology in German) or special particles notypic trait of an organism can be important to (as “wa” in Japanese). achieve different organismic functions. Nevertheless the decomposition is extremely useful, both in anal- Since the rise of structuralism in the early 20th ysis and in understanding how the subsystem could century, linguists focus on such paradigmatic sub- have evolved. systems because the linguistic elements involved Here are some examples of language subsystems: (for example French “Pass’e simple”) can only be understood in contrast to others (French “Pass’e • A lexical system of basic color terms (“red”, compos’e) and language evolution typically includes “green”, “blue”, etc.) expressing perceptually shifts and interacts between the elements. For exam- grounded color categories. ple, if a new color term appears in a language (such as “orange” which became a color word in English • A grammatical system expressing the abstract in the 16th century), then the perceptually grounded role of participants in an event (agent, patient, category it expresses will be pushing away other cat- instrument, beneficiary, etc.) through word egories (like yellow and red) that covered the same order or prepositions (as in “Joachim gives a terrain earlier. So the semantics of the basic color book to Vanessa” versus “Vanessa gives a book terms must be viewed as a system. Or if certain to Joachim”), or with cases (as in German “die cases in the case grammar of a language, for exam- Frau gibt ihrem Mann diesen Buch”), or other ple the dative in German, are beginning to expand means. their ‘territory’ to express a broader field of seman- tic roles, then this will be in competition with other • A lexico-grammatical system for expressing cases, for example the genitive [68], and so we can- spatial relations, such as using prepositions as not just consider a single case in isolation but always “next to” in “the block next to me”. need to consider cases as forming part of a paradig- matic case system. • A lexical system of posture distinctions and bodily action words, useful for commanding In order to understand the evolution of language, or describing body postures, such as “stand”, it is not enough to take only a descriptive stance. “sit” and “lie”. We need to understand also how language users are able to produce and comprehend utterances and how • A system of tense and aspect that expresses these faculties may have evolved. Hence I will from the temporal structure of events (for example now on use the term language subsystem to refer to in terms of present/past/future) or the internal the mental machinery that is needed for handling structure of events (for example distinguishing a particular subsystem of a language. So when I perfective/imperfective). Such systems often talk later about the Russian aspect system I will use morphological markers attached to verbs mean the set of data structures needed to represent or constructions with auxiliaries. the concepts expressed by aspect, the functions by which these representations are grounded in sensory • A system of determiners for being more precise experiences, the inferences generated by aspectual about the possible referents of a noun phrase, notions, the set of morphological aspect markers and for example, through articles, demonstratives, the syntactic constraints on their usage, as well as or quantifiers, as in “the book” (definite) ver- the procedures that can make use of all this informa- sus “a book” (indefinite) tion to produce and parse those features of Russian utterances that have to do with aspect. The no- • A system for structuring subordinate clauses, tion of a language subsystem therefore cuts across in which sentences appear as recursive parts of all layers that have been traditionally distinguished other sentences, such as in relative clauses like separately in linguistics: meaning, lexicon, morphol- “the first girl who saw a bird got a present”. ogy, syntax, grammar, and pragmatics, and it goes

4 beyond description to include issues of language pro- syntactic or lexical aspects, but also with meaning: cessing, perception, and action. How should the concepts be acquired that are needed I propose that these language subsystems are the in this language subsystem? How should the ontol- ‘phenotypic traits’ for which evolutionary linguists ogy be expanded when no adequate meanings can be seek explanations. They allow language users to pro- found to achieve the desired communicative goals? duce and comprehend utterances to achieve the com- And it incorporates procedures to deal with prag- municative goals that are helpful in their coopera- matic issues: How should meanings be interpreted tive interactions with other members of the language and acted upon, what kind of interaction could re- community and within a specific range of contexts. pair a failed interaction? How can additional ques- tions help to fix misunderstandings or generate data for learning? 2.2 Language Strategies To do all this, a language strategy encompasses Where do the many language subsystems that make three kinds of functions (see Figure 1). First it up the language faculty of an individual come from? must contain a learning function, which the hearer I suggest here the notion of a ‘language subsystem exercises to acquire the linguistic aspects of a lan- strategy’ or language strategy for short. A language guage subsystem as well as the conceptualisations strategy is the set of procedures that a language employed by it. The learning function typically in- user employs to learn, align, expand and adapt one cludes ways to extract enough information from con- subsystem of his or her language. The notion of textualized utterances to reconstruct the conceptual a language strategy is reminiscent of the Language and linguistic structures that are used by the speaker Acquisition Device (LAD) that has been the tar- but unknown to the hearer. get of investigations in generative grammar during Languages are not fixed static systems but evolve the past decades [17]. However there are two im- and change all the time as new semantic challenges portant differences. (i) I am not assuming that come up, constructions get out of fashion, or changes there is a single universal language strategy, instead in one subsystem trigger changes in another one [39]. there are many language specific strategies, spe- So a language strategy should not only cover how a cialized and optimized for dealing with a language novice language user learns the system in place at a system as it occurs in a particular language or a particular moment in time, but also how a speaker range of languages. For example, Russian speak- may flexibly adapt and expand his own language sub- ers have a language strategy for dealing with as- system in order to deal with novel cases, without pect, German speakers have a language strategy for losing the systematicity present in the language sub- handling how topic/comment structure is expressed system. These innovations could then be picked up through the ordering of phrasal constituents, Swahili by listeners or they may die out as soon as they have speakers have a language strategy for dealing with come up. classifier systems. Of course components of lan- Third, a language strategy includes an alignment guage strategies might be shared across domains and function by which speakers and hearers are able to the same strategy (for example storing and re-using coordinate their language subsystems. This is nec- holophrases) might work to some extent for initially essary to dampen the unavoidable variation that oc- dealing with several language subsystems. (ii) I am curs in language use. Two speakers of the same lan- not assuming that language strategies are innately guage never use exactly the same constructions or given. They are configured and tried out by individ- conceptualisations. One speaker may use the present ual language users in a developmental process. Even perfect in a sentence like “I have just written her a in a single population we might see significant diver- letter”, whereas another speaker may prefer to use a gence of what language strategies are employed by simple past in the same circumstances, as in “I just different individuals for acquiring the same language wrote her a letter.” One speaker may treat “agree” subsystem. as a transitive verb, combinable with a direct object, A language strategy handles not only the rou- as in “I ask you to formally agree this arrangement”, tine application of ready-made solutions but includes whereas another speaker may prefer to treat “agree” ways to handle failure, invent novel solutions, and as an intransitive verb so that the object of agree- coordinate an individual’s language subsystem with ment must be introduced with a preposition, as in those of others. It not only helps out with purely “I ask you to formally agree with this arrangement”.

5 function description learning (by the hearer) to acquire aspects of his language subsys- tem from the speaker expansion (by the speaker) to extend or adapt aspects of his language subsystem for dealing with new situations alignment (speaker and hearer) to coordinate aspects of their language sub- systems so that they become more similar

Figure 1: This table summarizes the different functions of a language strategy.

Nevertheless, language coherence is sufficiently high differ from one language to the next and they may so that even speakers who have never met each other change over time (“red” was called “read” in Old have a reasonably high chance of communicative suc- English). There is also considerable variation in the cess. This can only be because speakers and hearers basic color prototypes lexicalized in a language and have ways to align their linguistic choices as part of there is usually an evolution towards more and more situated language interactions. refined color prototypes, often by borrowing from The distinction between language strategies and another language [4]. For example, English speakers language subsystems is analogous to that between make a rather clear distinction between green and genes (or gene networks) and phenotypic traits, in blue, but in Chinese and Japanese there is a sin- the sense that language strategies participate in gle color category which covers both areas, named building and maintaining language subsystems just “ao” in Japanese or “q¯ing” in Chinese. It is used for like genotypes participate in building and maintain- the (green) traffic light (“ao shingo”) or the color ing phenotypes. The complete language faculty of of unripe bananas, but also for a blue skie (“ao- an individual consists of a large network of inter- zora”). There can even be significant differences in acting language subsystems and they are built and terms of which color prototypes are adopted. The maintained by a network of interacting language Berinmo, a Papua New Guinea indigineous culture, strategies. The complete language faculty is there- has a word “wor” which covers some of the green re- fore analogous to a complete organism that exhibits gion, a word “nol” which covers much of green, blue many phenotypic traits orchestrated by a large num- and blue/purple, a word “wap” which covers almost ber of genes and gene regulatory networks. And just all the lightest colors, and a word “kel” which covers as the organism as a whole generates a particular be- almost all dark colors. [64] havior in response to the opportunities and demands These examples could be multiplied easily for of the situations occurring in its ecosystem, the lan- any language subsystem we care to investigate [33]. guage faculty produces and comprehends complete For example, if we investigate tense and aspect sys- utterances appropriate in the situated interactions tems in the world’s languages, we will find that they that the individual has. do not all make similar distinctions with respect to Why do we need to distinguish different language temporal structure. Some language are very fine- strategies? Would it not be easier to assume that grained (as aspect in Russian or tense in French) there is a single powerful strategy that can be used whereas others use a very shallow classification. The to learn all of language, for example, a powerful sta- details of expression may vary widely from a com- tistical induction method, like Minimal Description plex system of morphological markers to a few aux- Length Learning [11] or a recurrent neural network iliaries. Or, if we investigate the expression of argu- algorithm [15]? ment structure we will find that there is huge vari- First of all, we see that the same language strat- ation both in terms of the abstract semantic roles egy can be instantiated in quite different ways. For that are expressed and in the way how they are ex- example, “red” is the name of a prototypical color in pressed [95]. All these examples show that learning English, roughly in the 625-740 nanometer range of a language subsystem is more than learning its lex- the color spectrum. The same color range is called icon and its grammar. It requires also learning the “rojo” in Spanish, “aka” in Japanese, “ˇcerven´y”in inventory of concepts expressed by the language, and Czech, or “merah” in Indonesian. So color words how they are semantically grounded.

6 Second, we see that not all language strategies der of a few thousands. Most of these strategies appear in all languages. For example, many lan- center around the lexicalisation of categories that guages do not at all have a strategy built on ba- carve up a particular conceptual space, for example, sic color prototypes, which integrates hue as well spatial relations, emotions, properties of moving ob- as brightness, but focus instead only on brightness, jects, kinship relations, smells, body parts. Others which means that they primarily employ words like use grammatical constructions to express additional “dark”, “shiny”, “dull”, or “light”. Most languages aspects of meaning, such as temporal properties of build spatial prepositions based on body-related cat- events, restrictions in the scope of a referring expres- egories (such as left/right and front/back), but oth- sion, participant roles, etc. ers use only global landmarks (such as ‘towards the sea’/‘towards the mountain’) [47]. Russian has a very elaborate aspect system but a simple tense sys- 2.3 Communicative success tem, whereas in French it is the reverse. Complex So far I suggested a mapping from the geno- classifier systems are found in Bantu languages and type/phenotype distinction to the distinction be- Asian languages but are entirely absent from Euro- tween language strategies and language subsystems. pean languages. Japanese and Chinese do not have Is it possible to talk about success and fitness in an elaborate system of determiners. the case of language? First of all it is clear that Third, we see that there is evolution in terms an utterance can have success or not within a par- of which language strategies a language commu- ticular context of usage. Success implies that the nity employs and hence which language subsystems hearer can reconstruct sufficiently the meaning con- they will maintain. Here are some examples: (i) A veyed by the speaker in order to act upon it, and, kind of switch has often happened or is happening more importantly, that the consequences of his un- in brightness-based color languages towards a full- derstanding satisfy the intentions of the speaker [19]. color system [50]. Today’s hue terms, like “yellow”, For example, if the speaker asks “Could you give me “brown”, or “blue”, were all expressing brightness- some water” and gets the water, the utterance was based distinctions in Old English before they became successful. If he gets nothing at all or something else used as part of a full color strategy in the late Middle instead, it is not. Of course, speakers and hearers do English period (1350-1500) [16]. (ii) At the moment not all the time get feedback on whether the com- there is a struggle going on in Spanish between the munication was successful, but, particularly in the traditional ‘etymological’ strategy that uses noun earliest phases of bootstrapping or learning a lan- markers for argument structure, with a distinction guage, pragmatic feedback appears crucial. for example between accusative and dative ( as in There are several factors that contribute to com- “lo” versus “le”), and a ‘referential’ strategy that municative success: First of all, there must be uses noun makers for helping in anaphora resolution enough expressive power to achieve the objectives [97]. (iii) English lost its case system and then devel- of the utterance. For example, if a speaker wants to oped an alternative scheme for expressing the roles draw the attention of the listener to an object in the of participants in events based on word order and world by using a hue distinction (as between “red” prepositions [94]. Often we see that a particular lan- and “green”), then the objects to be distinguished guage strategy is productive for a while but then dies must indeed differ with respect to hue. Moreover out to be replaced by another one. For example, En- the speaker and the hearer must master the rele- glish still has a set of irregular verbs which form the vant perceptually grounded hue categories, and they past tense by changing the inner vowel of the verb must have stored words to express these categories stem, as in “think/thought”, “break/broke”, etc. in- in their lexicon. Second, the language subsystems stead of taking the verb stem and adding “-ed”, as of speaker and hearer must be sufficiently shared to in “pull/pulled” or “walk/walked”. The irregular allow understanding. The color categories must be verbs are remnants of a strategy that was dominant similar and the color words must have similar mean- in older Indo-European languages and died out to ings. Third, both speaker and hearer should try to be replaced by the “-ed” strategy. minimise the cognitive effort that needs to be ap- It is hard to estimate at this point how many plied. If an utterance requires too much processing, language strategies are needed to cover a complete either because it generates a large search space or natural language, but I guess it must be on the or- requires a lot of memory in parsing or producing,

7 or because it requires a lot of additional inference nation of its genome and this leads to the inheritance to figure out or apply the meaning of the utterance, of phenotypic traits. But there are several sources then the risk of misunderstanding increases. The of innovation and variation: Both parents make con- context plays of course a very important role in how tributions and the novel combination may have new precise and elaborate an utterance has to be. If the side effects, glitches may occur in the copying process situation is highly restricted and speaker and hearer which could give rise to novel genes or gene networks, already focus strongly on the same tasks and aspects one organism may get more opportunities to learn, of the context, the utterance can be minimal. and the environment plays a role in which genes are Expressive power, convention sharing, and cogni- expressed and how [10]. tive effort co-determine the success of an utterance. Each variant reacts differently to the pressures If an utterance is successful, this means that the and opportunities of the ecosystem. This influences language subsystems that participated in producing success in survival and reproduction, and hence fit- or comprehending the utterance deserve credit, and ness. Genes that lead to increased fitness, which consequently that the strategies that were used to means genes that participated directly or indirectly build and maintain these language subsystems were in building phenotypes that lead to more successful appropriate for the range of contexts and the com- behavior, proliferate, and hence the traits of better munity of language users in which the individual op- adapted variants spread in the population. Learn- erates. It is therefore possible to define the succes ing aids to achieve further adaptation and can po- rate of an individual’s language subsystem as being tentially trigger a Baldwinian type of evolution in the success rate of the utterances which this lan- which skills that have to be learned at some stage guage subsystem helped to produce or comprehend become genetically coded [24]. over a particular window of time. And it becomes The change in gene frequency driven by differen- then possible to define the fitness of a language strat- tial fitness in an ecosystem not only explains why a egy for a particular language user as equivalent to species is well adapted to its ecosystem but also why the success rate of the language subsystem that this there can be speciation. Speciation means that two strategy helped to build and maintain. Note that subpopulations diverge up to a point where mem- the communicative fitness of a strategy is not an bers of each group can no longer produce offspring absolute notion because it depends on the kinds of together. This kind of divergence happens as soon communicative contexts that are encountered and as two subpopulations start to adapt to significantly on what language strategies and subsystem choices different ecological niches. They progressively grow have been adopted by the rest of the population. apart and this may eventually create the conditions that members of each subpopulation are no longer responsive to each other’s mating signals or that their offspring has become sterile [56]. 3 Language evolution by linguistic selec- Can we map this selectionist logic to the domain tion of language in order to explain the origins and evo- So far I introduced a mapping from the biological lution of language subsystems adapted to the needs notions of genotype, phenotype, success, and fitness of their users? Can we use it to explain how lan- to the linguistic notions of language strategy, lan- guage subsystem choices or language strategies be- guage subsystem, communicative success, and strat- come shared in a population? Is it helpful to under- egy fitness. I now turn to the second key compo- stand how languages may evolve different dialects nent of Darwinian evolutionary theory, namely the which could grow sufficiently apart to induce mutual causal interaction between heridity, variation, and incomprehension and hence two different languages? selection. Biologists argue that this causal interac- And can it help us to understand the origins of early tion explains how adaptations may arise in a species language, which would mean a state where there are and how speciation can occur, so it is crucial for the no language subsystems yet, and the first strategies present discussion to understand its nature and to arise to deal with some of the communicative needs see how it can be mapped onto the linguistic do- of the population? main. Before exploring the analogy further, it is impor- Genotypic information is preserved from an or- tant to be aware of a very big difference between biol- ganism to its offspring by the copying and recombi- ogy and language. There is clearly heredity involved

8 in language in the sense that language subsystems hearer and sufficient to construct shared meaning, and language strategies get preserved through the but also on whether both partners sufficiently share memories of the individuals using them. But there the conceptual and linguistic structures being em- is no direct physical copying going on of language ployed. Consequently, we get a self-enforcing effect: strategies between the brains of individuals. There The more success a particular language subsystem is no telepathy. Every individual has to reconstruct choice has, the more it will be preferred by those the strategies necessary to deal with the language who have used it in their communication (because of his community and then enact them to acquire it was successful), and hence the more it will spread the specific conceptualisations and conventions that further. There will be a general tendency towards are in common use [81]. Hence, the innovation and a winner-take-all situation, for example, synonyms variation in language strategies cannot arise from get damped so that there tends to be a single dom- errors in physical copying or from the recombina- inating word to express a particular meaning. But tion of strategies from the two parents. It is due there is always going to be residual language varia- to the fact that the constructions or reconstructions tion remaining, due to the large number of individu- of strategies carried out by different individuals is als speaking a language and their different histories always a matter of guessing and trying. Neverthe- of language use. less, the analogy remains still useful because we are The escalating positive feedback between success dealing in both cases with a selectionist system. I and use coordinates the language subsystem choices will now argue that the analogy applies at two lev- within the population and explains how a shared sys- els: that of specific language choices (for example tem can arise even if there is no telepathy or cen- which temporal categories will be adopted in a lan- tral control. This process of reaching global coher- guage and how are they going to be expressed in a ence despite only local interactions through a self- grammar of tense) and that of language strategies. enforcing feedback loop is actually quite common in physico-chemical and biological systems and usually referred to as self-organisation. We find it for ex- 3.1 Linguistic selection of language choices ample in the path formation in ant societies or the Language strategies build, adapt and maintain lan- emergence of bird flocks [14]. Self-organization has guage subsystems but there are usually still a huge been studied intensely the past decades in the field number of possible choices given the same language of complex systems science and the techniques devel- strategy (think for example about the many different oped there can be applied to the study of the emer- ways in which the color space can be divided up into gence of conventions in symbolic systems. different color categories or the many ways in which The process of linguistic selection of language time can be conceptualised and grammatically ex- choices can also explain how sublanguages or di- pressed). So a first problem we face is to explain how alects may emerge or disappear again. The more a language community can settle on a shared lan- interactions the members of a population have, the guage subsystem assuming that they already share more their individual languages will become coordi- the same language strategy and are confronted with nated. On the other hand, if certain barriers are or similar contexts and communicative goals. This is become present that split a population, either due the first level at which linguistic selection applies. to geographic or to social strata, differences in the Language users can monitor the success that they language are maintained or appear [83]. have with a particular choice in their language sub- system and then prefer those conceptualisations or constructions that have had the highest communica- 3.2 Linguistic selection of language strategies tive success. This ensures that their language sub- A language strategy contains all the functions that system is adapted to the communicative challenges are necessary for learning, expanding, adapting and and contexts effectively encountered. Moreover com- aligning a language subsystem, not only its lexicon municative success not only depends on whether the and its grammar but also the ontologies expressed by distinctions being expressed are adequate for reach- the words and grammatical constructions that make ing the communicative goals in the shared environ- up the system. But where do language strategies ment of speaker and hearer or whether grammar come from? A biological/genetic approach to lan- and lexicon were correctly applied by speaker and guage origins would argue that there is only a single

9 language strategy, which is part of the human in- very high stress due to rapid population change, nate endowment. In the socio-cultural approach ad- population collapse, or the intense mixing of pop- vocated here, language strategies are specialized for ulations. This can not only lead to a loss or re- the different subsystems of a language and hypothe- placement of specific linguistic forms. The language sized to be the outcome of processes that recruit and strategy underlying a particular subsystem can be configure existing cognitive functions, also useful in lost and then true innovation has to take place, in other domains, for the purposes of linguistic com- the sense that an alternative language strategy needs munication [81]. A recruitment process can create a to emerge and become conventionalised, or the lan- variety of possible language strategies, but then we guage strategies of one language (known by one part still need to explain how certain variants can become of the population) become applied to the lexical ma- dominant in the population. This is the second level terial of another language. Examples abound in the at which linguistic selection applies. formation of creole languages [53], but can also be Cognitive functions that might be recruited for found in on-going language change, like the shift in the semantic aspects of language include particular English from morphological cases to the syntactic ways for conceptualising reality that are also use- expression of participant roles [94]. ful in decision making, planning, spatial navigation, tool design, tool use, motor control, food selection, problem solving, etc. Indeed recent neurological ev- The overall framework, summarized in Figure 2, idence has shown that the same brain areas become is not completely blind. Language users are stimu- active both in a non-linguistic task (e.g. controlling lated by the needs in communication and the con- movement of the toe) and in language processing texts they encounter, and they always try to glean for words or sentences that talk about this task or evidence from the language use of others to steer about objects involved in it (e.g. hearing the world the recruitment or use of language strategies. So “toe”). [34]. Cognitive functions that can be re- we are dealing with a two-level selectionist system: cruited for dealing with the lexical and grammatical Cognition comes up with possible language strate- aspects of language include bi-directional associative gies without knowing in advance which ones will be memories, mechanisms for recognising or producing effective or effectively used in the community, and hierarchically structured events, priming, heuristic each strategy comes up with possible language sub- search, etc. Each of these has a broad range of other system choices without knowing in advance which applications as well, which may explain why the ones would be most useful and accepted by the com- brain areas traditionally associated with language munity. The two levels interact with each other be- are involved in so many other tasks [54]. cause communicative fitness of language strategies There are many possible language strategies even is determined by the communicative success of the for the same language functions. Language users language subsystems built with them. could however track the ‘fitness’ or rate of commu- nicative success they have with the language subsys- tem built with a particular strategy and then prefer The analogy I suggest here between biological strategies with a higher fitness for continuing to ac- evolution and linguistic evolution is summarized in quire or build their language, just like they track the following table. The communicative fitness of the success of individual language choices and prefer the language of an individual is equal to the cumu- those that had the highest communicative success in lative fitness of all the language strategies used by the past. Language strategies that lead to higher the individual which depends in turn on the commu- communicative fitness will be used more and this nicative success of the language system choices and will stimulate others to adopt the same or a simi- on the context of application. lar strategy. Hence linguistic selection of language biological linguistic strategies can explain how they become shared in the population. Of course we should always expect genes language strategies there to be residual variation as well as remnants of phenotypic traits language subsystems older strategies that are still known but not prefered features of behavior features of utterances by the majority of speakers. behavioral success communicative success It is well known that a language may undergo biological fitness communicative fitness

10 Figure 2: A language strategy gives rise to a large set of possible language subsystem choices. Which choice is adopted by a language user should depend on the outcome of his communicative interactions. Hence we get a selectionist effect from communicative success to language subsystem choices. Many different language strategies are possible, even for the same tasks and domains. Which strategy is adopted by a language user should depend on its communicative fitness, defined as the rate of success of the utterances in which the language system acquired and maintained by the strategy played a role. Hence we get a second selectionist effect from communicative fitness to language strategy choice.

4 Methodological Stages today. Why is it interesting to apply a selectionist frame- work to language? What does it buy us? First of all, the origins and evolution of language now be- 4.1 Methodological stages of Evolutionary Biol- comes an instance of a general process found else- ogy where in Nature. Selectionist processes have been I pointed out earlier that the Darwinian selectionist argued to underly the economy, sociology, culture logic is a framework for developing evolutionary ex- and even cosmology. And so if we hypothesise them planations, but the details of each explanation still as the basis for explaining the origins of language we need to be worked out for each case. Evolutionary do not need to evoke entirely new principles. biologists have developed a methodology to do this Second, it becomes possible to reuse the mathe- in a disciplined way. This methodology contains the matical machinery that has been developed in Evo- following stages: lutionary Biology and Complex Systems Science for 1. Select and describe the phenotypic trait of studying the origins and growth of complexity of interest, for example, the color of butterfly symbolic conventions. I have called this type of in- wings, patterns on the fin of a fish, lungs in vestigation Semiotic Dynamics and it is now a thriv- vertebrates, songs of finches. The trait may ing research field [45]. For example, the theory of be chosen because it is particularly relevant replicator dynamics [58] provides a framework for for understanding the evolution of a species or studying how language strategies or language sub- because it may shed light on evolutionary pro- system choices (given a strategy) may become shared cesses in general. in a population, assuming that payoff corresponds 2. Understand the ecological or functional signif- to communicative success. Or the techniques for icance of the trait. This is done by looking at investigating scaling laws by systematically varying the role of the trait within the functioning of system size (population, environmental complexity, the organism (for example oxygen supply) or population flux, etc.) can be applied to study the within the behaviors and interactions the or- self-organization of naming conventions [1]. ganism has within the ecosystem in which it But maybe the most important benefit from the attempts to survive and reproduce (for exam- viewpoint of Evolutionary Linguistics, is that the ple color may play a role in attracting a mate). analogy suggests a particular methodology for de- veloping evolutionary explanations of features of hu- 3. Understand how the trait gets established. For man languages which is quite similar to the method- physiological traits, this is through a combina- ology so brilliantly practized in Evolutionary Biology tion of genetic and developmental processes.

11 For behavioral traits, like strategies for catch- eggs already in her mouth. So this explains the eco- ing prey, this is through neuronal growth pro- logical function of the egg spots. cesses and learning. Next the complete genetics and developmental process could be established that show how these 4. Understand how the trait may have appeared in eggspots are laid down in a process of pattern forma- evolution. This amounts to figuring out when tion under genetic control [67]. And once the genetic and where in evolution the genetic basis for the basis was known, it could be reconstructed that these trait has appeared and what kind of genomic eggspots appeared once in the ancestral line of this changes might have taken place. cichlid branch, although some further species lost it again, for example because they invaded deep-water 5. Show that the trait has a selective advantage. habitats that were too dark for the eggspots to play This is achieved by comparing the effect of hav- a role, and that the genetic change involved a ‘co- ing or not having the trait on the fitness of option’ or exaptation of an already existing genetic individuals in the ecosystem, for example by mechanism inducing pearly spots on anal or other investigating how different variants in a popu- fins. lation are selectively able to survive in different The reproduction strategy based on mouth- circumstances. breeding has a clear selective advantage for the male because he can be much more sure that he is the one Once all these points have been clarified, the Dar- fertilising the eggs. There is also a selective advan- winian selectionist logic provides the explanatory tage for the female because she is more sure to be glue: When the trait has a selective advantage, the fertilized by the male she had chosen after courtship. relevant genes will proliferate in the gene pool and Consequently the Darwinian selectionist loop can be be preserved in subsequent generations. Thanks to closed. The selective advantage will cause the rele- heredity, the trait can also be further refined and vant eggspots genes to proliferate and we thus get an used as a building block for more complex traits. evolutionary explanation for this phenotypic trait. Examples how this methodology is being applied can be found abundantly in biological journals. Here is one example: the explanation for the ‘eggspots’ on 4.2 Stages in the methodology of Evolutionary the haplochromines, certain types of cichlid fish [66] Linguistics (Figure 3). These branches of cichlid fish have drawn I argued earlier that the target of explanations in the attention of evolutionary biologists because they Evolutionary Lingusitics are language subsystems: have very rapidly diversified within the lakes of East Why do we see aspect in the grammar of Russian Africa into several hundreds of species and therefore speakers and how could such an aspect system have seem to defy the normal pace of evolution. One dis- originated and evolved? Why do we see spatial tinctive phenotypic trait of haplochromines are cir- prepositions based on bodily coordinates in English cular eggspots on the anal fin in males. They are (as in “front/back”, “left/right”) and the use of per- called eggspots because they look like the eggs that spective reversal (as in “to your left”), and how could female cichlids produce. Why are these spots there such a system have originated? Why and how did and how did they contribute to successful specia- a determiner system evolve in Italian, even though tion? there was none in Latin? The first step is to understand the ecological or The selectionist framework applied to language, functional significance of this trait. We first need to as summarized in Figure 2, suggests the following know that these cichlid fish have developed mater- set of stages to develop such explanations: nal mouth-brooding of eggs. Females lay eggs and then suck them up again and further brood the eggs 1. Identify a language subsystem in their mouth. This gives a selective advantage be- 2. Identify its function in symbolic communica- cause it helps to protect the eggs against predators. tion The male needs to fertilize these eggs. By presenting his anal fin which contains the eggspots, the female 3. Understand how the language subsystem gets ‘believes’ there are more eggs to be sucked up, and learned, expanded and aligned, given a lan- thus receives the male’s sperm which fertilizes the guage strategy.

12 Figure 3: Males of one species of cichlid fish (figure A) have spots on their anal fin that look like eggs (figures D and E). Females suck up eggs and brood them in their mouth (figure B). Because the eggspots act as dummy eggs, females get enticed to suck them up but get the male’s sperm instead (figure C).

4. Reconstruct how this strategy may have origi- fashion, zooming in on a particular language subsys- nated. tem (for example classifier systems) as instantiated in a concrete natural language (for example Swahili), 5. Show that language subsystems built with a and deconstructing the language strategy that gives strategy have a selective advantage. language users the ability to handle the chosen lan- guage subsystem. A concrete language subsystem is Just as evolutionary biologists have tools and taken as point of departure in order to avoid that evaluation criteria for each stage, we need to de- theorising remains too abstract and vague to be rel- velop formal tools here as well, in order to be precise evant to understand ‘real’ natural languages. A spe- about what exactly is being proposed, and we need cific language is picked either because it shows the to develop clear evaluation criteria to test objectively phenomenon of interest in a proto-form so that we whether a stage has been reached. can study a simple instantiation of it, or, just the op- I will now discuss each stage in more detail. The posite, because the phenomenon is particularly well next section focuses on the first three stages, which developed in that particular language. Ideally, we are intended to give us an understanding how a should have phylogenetic data about the historical population of individuals is able to self-organize a evolution of the phenomenon or look take a case of language subsystem given a language strategy for ongoing language change so that we can gather more it. A subsequent section then focuses on how lan- data to validate evolutionary models. guage strategies may originate and become part of the toolkit of the population. 5.1 Stages I Once a language subsystem has been chosen, the first 5 The Origins of Language Subsystems three stages of the methodology can be tackled: (1) Evolutionary biologists start their investigations by Identify the language subsystem itself in a system- zooming in on a phenotypic trait that is intriguing or atic and exhaustive way, including the information promising from the viewpoint of understanding evo- processing needed to produce and comprehend ut- lution. Evolutionary linguists could start in a similar terances that require the language subsystem, (2)

13 identify the role of the language subsystem in com- resentation of linguistic knowledge needs to be such munication, and (3) reconstruct the language strat- that it can be used both for parsing and produc- egy that can build and maintain such a language ing language. Many formalisms have a strong bias subsystem. towards one or the other and generate unnecessar- ily huge search spaces when used in a bi-directional fashion. The third requirement is flexibility. The 5.1.1 Identify a language subsystem lexical and grammatical constructions are seen as We can fortunately rely on the impressive work of constraints on possible structures but these con- many descriptive linguists for gathering data on the straints should be relaxable to a certain degree be- language subsystems occuring in a language. A huge cause a hearer may be confronted with a speaker literature exists in the form of dictionaries and gram- that does not know or use slight variants of the con- mars for specific languages as well as world-wide sur- structions he used himself. And the speaker may veys [33]. The descriptive viewpoints advocated in lack a way to handle a certain linguistic challenge Cognitive Linguistics [43] and Construction Gram- but nevertheless needs to apply as many lexical and mar [31] are particularly relevant for the enterprise grammatical constraints as possible. The fourth re- of Evolutionary Linguistics because they emphasise quirement is fluidity. If language is seen as a selec- the communicative functions of language and pay at- tionist that is forever on the move, with conventions tention not only to the purely syntactic aspects but coming up and disappearing, then the formalisms we to conceptualisation and how conceptualisations are use to operationalise lexicons and grammars need to expressed [92]. Research in sociolinguistics provides be able to represent competing alternatives. They additional data on the variation that occurs within must support the scoring of elements of the system a community of speakers [42] and typological stud- so that agents can track the degree with which con- ies can provide data on the variation across different structions are conventionalized and mechanisms to languages [33]. adjust language subsystems based on the commu- Once a description of the regularities found in a nicative outcome of the utterances in which they are language are available, we can zoom in on the first used so as to implement the selectionist logic shown task, which is to identify the information structures in Figure 2. The development of formalisms and and the information processing that is necessary for language processing systems satisfying the require- handling the language subsystem under investiga- ments of Evolutionary Linguistics is currently an ac- tion. For example, if we are interested in Russian tive research area, and there are already quite good aspect, we must come up with a representation of the solutions available [5, 82]. information that Russian speakers need about which How can we evaluate whether a language subsys- aspectual markers occur, what shades of meaning tem has been properly captured? With respect to they represent, and how the markers are syntacti- empirical adequacy, we need to rely on the method- cally combined with verb forms, and we must figure ology and discipline of descriptive linguists. Re- out the information processing steps that make use cently, large linguistic resources have come on line of these representations in order to produce and com- (see http://ldc.upenn.edu/) and computational im- prehend utterances in which aspect marking appears plementations of statistical methods make it possi- [30]. The main linguistic subdiscipline on which we ble to handle huge corpus data. These techniques can draw here is computational linguistics, as this help to give descriptive linguistics a firmer empiri- field has developed the sophisticated techniques and cal basis. With respect to processing, the obvious implementation technologies needed to make precise way to test an information processing model is to models of language processing. see whether an implementation of it adequately han- The formalisms we need in Evolutionary Linguis- dles the conceptualisation and production as well as tics research have a number of requirements. The interpretation and parsing of a representative set of first one is obviously that they need to have adequate sentences that exhibits the phenomenon we are inter- representational and computational power to handle ested in. For example, if we are interested in French phenomena observed in human languages, such as hi- tense, we should implement a system for produc- erarchical structure, recursion, agreement, etc. This ing and comprehending sentences where tense mark- requirement holds for all linguistic formalisms. The ing occurs and test its adequacy against a corpus of second requirement is the mirror property: The rep- data. These implementations should not only deal

14 with morphological and syntactic processing but also members of a population that have regular interac- with semantic interpretation and conceptualisation. tions with each other. Each agent can alternatively play the role of speaker and of hearer so that they can build up competence both in the production and 5.1.2 Identify the function of a language subsystem the understanding of language. The language game in communication takes place within a shared setting which is a slice of The second stage in the methodology is to under- the real world, further called a micro-ecology, which stand the role of a language subsystem in linguistic has been selected in such a way that the issues ad- communication, just like evolutionary biologists at- dressed by the language subsystem show up. tempt to understand the role of a phenotypic trait For example, if we want to see the use of an as- in survival and reproduction. In some cases this pect system as in Russian, we could set up a lan- is rather obvious. For example, languages have guage game, where the speaker and the hearer both words for expressing spatial relations because speak- get to see two situations (shown perhaps as video- ers sometimes need to draw attention to an object on clips) which are only distinguished by the differences the basis of its location in space. Or languages often in Aktionsart [89]. For example, in one situation a have ways to indicate the perspective from which child Masha is reading a book the whole time and a spatial relation is to be viewed, such as “your” in the other a child Misha starts to read a book and in “the block to your left”, because speakers and then stops immediately. The language game con- hearers are embodied individuals which unavoidably sists then in answering a question which only makes see the world from their own individual perspective. sense for one of these situations and which requires Hence without an expression of the perspective from the expression of Aktionsart, such as “Who is read- which a spatial relation is conceptualized ambigui- ing a book the whole time?” [30]. ties may appear [87]. Understanding the functional A particular language subsystem is seldom abso- role of linguistic features has been of particular inter- lutely necessary. Listeners are intelligent and so they est to functional linguists [32], [27] and pragmatics can often use the context or additional inference to researchers [98]. Symbolic communication is always fill in details that are not expressed explicitly [69], embedded within the broader context of a coopera- or they can convey the same meaning using alter- tive activity, and so ‘linguistic’ goals or speech acts native non-grammatical devices. This explains why always find their ultimate justification in cooperative we find languages without determiners, without a interaction. case system for expressing participant roles, without In other cases, it is not so obvious to understand tense or aspect, without basic color terms, etc. How- the role of a language subsystem or certain aspects ever, if a language subsystem is available, language of a language subsystem. To understand many syn- users can potentially reach a higher level of commu- tactic features of human languages it is necessary nicative success, for example because ambiguity is to take processing issues into account. Hierarchical avoided. Or they may need less effort, for example structuring, or agreement phenomena, as between because less inference is needed or the search space articles and nouns in French, may be puzzling, un- in syntactic parsing is greatly reduced. less we realise that it helps to dampen the combi- What operational criterion can we use to test natorial explosions that occur unavoidably during whether a language game is adequate to study the parsing and it increases robustness in parsing be- language subsystem we want to explain? This can be cause the same information is communicated in a done through language game experiments. Human redundant way [88]. language users can be asked to play the language How can we clearly isolate the role of a language game and if their utterances exhibit the feature of subsystem in such a way that we can then study it in interest, we know that it is relevant to them within a systematic and objective way? I suggest that this this game setting and micro-ecology. For example, if can be done by defining a language game in which the language game for investigating Russian aspect the language subsystem under investigation might is properly designed and exercized in an appropriate be beneficiary to language users [74]. A language micro-ecology, speakers of Russian should indeed use game is an interaction between two individuals (fur- the aspectual markings that we expect to see. ther called linguistic agents or simply agents) that We can also set up systematic computational ex- have a common cooperative goal. The agents are periments in which simulated agents play the lan-

15 guage game with each other within a virtual envi- work, but the general principles are clear. ronment. We can then not only show that the lan- I argued earlier that the distinction between lan- guage subsystem is relevant in the language game, guage strategies and language subsystems is analo- but validate at the same time whether the recon- gous to that between genotype and phenotype. Lan- structed language subsystem (from stage 1) is ade- guage strategies are procedures for acquiring, align- quate for reaching communicative success. We can ing, expanding and adapting a language subsystem, even go one step further and set up robotic experi- and they are therefore analogous to genes or gene ments in which the agents are embodied physically regulatory networks. language subsystems are the as autonomous robots and interact with each other stored information structures that are built up by in a physical environment [73]. This is particularly these language strategies and the procedures for ac- relevant for investigating language phenomena per- tivating them during the conceptualisation, interpre- taining to grounded semantics, i.e. semantics that tation, production and parsing of sentences. They is anchored in the sensori-motor experiences of lan- are therefore like the phenotypic traits of an organ- guage users, such as action words, body language, ism that determine what behavior the organism is color terms, spatial language, etc. able to generate in response to demands within the Computational or robotic experiments are a pow- ecosystem. erful tool because we can not only control completely How can we operationalise language strategies? the experimental settings and perform repeatable ex- There is now a considerable body of experience on periments, but because we can inspect completely how to do this. Implementations rely on compu- the internal states of the virtual or physical agents tational linguistics formalisms that are adapted to and thus collect data that may be relevant for statis- Evolutionary Linguistics research [82], as well as a tical analysis. Moreover we can systematically vary computational architecture that has the capacity for key parameters, such as population size, population computational reflection and re-entrance. structure, population flux, environmental complex- Computational reflection implies that there are ity, available cognitive resources, and so on, and ex- two levels of information processing and that the amine their effect, either on the structure of the lan- system can change itself its own programs (in this guage subsystem that gets built or on various scaling case the inventory of concepts, lexical entries, gram- properties and the resilience of the language subsys- matical constructions and even language strategies) tem against change. [70]. At the first level, ‘normal’ language process- ing takes place: Situations are conceptualized, lex- ical and grammatical rules applied to produce a 5.1.3 Understand how a language subsystem gets sentence, sentences get parsed, meanings are inter- built, given a strategy preted and acted upon. If everything goes well, an Once evolutionary biologists understand the role of utterance is efficiently produced or comprehended. a trait in the ecological or physiological survival and This process is monitored by a set of diagnostics, reproduction of the organisms in a species they turn which are procedures that test whether the different their attention to the question how that particular steps in language processing have been adequately trait gets built, in other words what kind of genetic achieved. They operate at the meta-level. For ex- or developmental process may be at work to see the ample, they test whether there was a lexical entry formation of eggspots on the backfin of cichlid fish, for every word, whether there was a way to cover or color pattern on a butterfly’s wings. This is where the complete meaning, whether all constraints in a the distinction between genotype and phenotype be- grammatical rule were satisfied, whether the recon- comes important. The genotype contains the nec- structed meaning is compatible with the scene, and essary ‘instructions’ to steer the developmental pro- so on. Diagnostics produce a failure message when cesses that builds the phenotype. Molecular biolo- some sort of problem arises. gists and geneticists are now able to unravel which If there has been a failure, repair strategies be- genes get expressed in the formation of a particular come active to work out a way to deal with each trait and how they selectively operate during devel- failure, for example, by expanding the lexicon with opment and the maintenance of the organism. Of a new entry, or by stretching a grammatical rule to course, the details remain to be worked out for each deal with new cases. A repair might only be possible specific case and may involve quite a bit of detective after all processing is finished and additional inter-

16 actions have clarified meaning and intention. Or it by the speaker. might be immediate, in which case normal language Research into language strategies is still in its processing can resume after the repair action. Diag- infancy. Although there has been a lot of research nostics and repairs implement the learning and ex- on general purpose learning methods, less is known pansion functions of a language strategy. At the end about the highly efficient heuristic learning meth- of an interaction, a third set of meta-level processes ods that human language users appear to employ as must become active, in order to align the inventories part of language strategies. Second language learn- of concepts, words, and grammatical constructions ers typically start out by applying the inventory of based on the outcome of the game. language strategies of their own native language to In addition to this two-level reflective compu- the new language, and so their characteristic er- tational architecture, the operationalisation of lan- rors are often instructive of what language strategies guage strategies requires the ability to engage in re- their native language requires. Data from creolisa- entrance [75]. Re-entrance means that the speaker tion is also highly instructive, because creoles often attempts to comprehend himself the utterance he show the impact of the language strategies coming has produced, before actually transmitting it to the from one source language on the lexical material of hearer. The speaker thus uses himself as a model of another language. the hearer and can detect possible ambiguities, ex- How can we test whether a language strategy has plosions in the search space, risky constructions, and been properly formalized and operationalized? We possibly reproduce a variant of the utterance that can proceed in three steps. The first step is to test might have a higher communicative success. The the learning function. We have an operational imple- speaker could even take the opportunity to expand mentation of a language game and its micro-ecology his language subsystem, for example by introduc- from stage 2, and we have an implementation of a ing an additional syntactic marker, or by lexicalis- language user modeled after an existing human lan- ing a new combination of categories. The diagnostics guage from stage 1. So if we have properly captured and repair strategies of the speaker operate not only the learning function of a strategy we should be able while he is producing the utterance but even more to set up simulated language game experiments be- so when he is re-entering and thus self-monitoring tween agents acting as tutors, who have been pro- his own speech. grammed with the existing system, and agents act- The hearer can also engage in re-entrance. Given ing as novices, who are only endowed with the learn- an utterance, he can first go through the set of steps ing component of a language strategy. We can then needed for comprehending and interpreting it, but test in repeated experiments whether the novice is then turn the process around and similate what he indeed able to reach the same level of performance would have said himself. This is particularly useful as the tutor, namely by measuring the communica- in language learning. The hearer can often only par- tive success of the novice against the base line. We tially parse, but nevertheless arrive at a reasonable could in principle also set up experiments where a interpretation based on understanding the context human plays the role of tutor and an artificial agent and by using additional inference based on common the role of novice, so that we can test whether the ar- sense. Once he has understood the utterance he can tificial agent is able to acquire an existing language make an educated guess about those words or gram- subsystem. And we can set up experiments where matical constructions with which he is unfamiliar. a human plays the role of novice and an artificial Re-entrance is actually not that difficult to agent the role of tutor, which might be useful to de- achieve in computer simulations of language game velop applications in the domain of second language experiments as soon as agents have the capacity to learning. produce and comprehend language, which they need The second step is to test the expansion func- anyway in order to build up the competence for both. tion. We can do this by turning the clock back: The main technical problem to be solved is that the We can set up simulated language game experiments representation of lexicons and grammars must have in which none of the agents has already an existing the mirror property, so that a speaker is able to use language subsystem. The agents are only endowed these representations to re-parse and re-interpret his with an operationalisation of the learning and ex- own utterance, and the hearer has the ability to re- pansion components of a language strategy. If they conceptualise and re-produce an utterance produced adequately capture not only how an existing lan-

17 guage subsystem can be learned but also how lan- name Munsell color chips or which chip they con- guage users expand their language subsystems while sider the most representative example for each ba- maintaining the underlying systematicity, then we sic color term in their language [4]. For example, should see a language subsystem being bootstrapped the eleven basic color terms in English are: black, from scratch through a self-organising process. This white, red, green, blue, yellow, pink, purple, brown, language subsystem will of course not be the same orange. There are known to be universal trends in as observed in the human language we used to re- the basic color terms of human languages but there construct the strategy, given that there are so many is also a lot of variation, not only in terms of the contingent factors impinging on ‘natural’ human lan- names that are chosen for colors but also in terms of guage evolution, but it should use a comparable ap- the perceptually grounded color categories that they proach for conceptualising what to say and for ex- express. As evolutionary linguists, we want to un- pressing these conceptualisations in terms of lexical derstand why and how such a system of basic color or grammatical constructions. terms could arise. Finally, the adequacy of the alignment function Several models of the processing needed for ba- of a strategy can be tested by keeping track of the co- sic color terms have already been discussed in the herence among the conceptual inventories, lexicons literature [29, 65]. Most of them center around the and grammars of different agents. Measures of co- notion of a 3-dimensional color space formed by the herence will depend on what aspect of the language red-green and yellow-blue opponent channels and subsystem we are interested in. For example, if we the brightness channel. Color prototypes which are want to track whether the perceptually grounded the most typical example of a color, are mapped as color categories of the agents become aligned, we points in this color space and color categorisation will have to apply a distance measure between these can be achieved using a nearest-neighbor computa- categories and plot them as agents acquire an exist- tion: a sample to be categorized is compared to all ing system or bootstrap a new language subsystem prototypes in the inventory and the prototype that from scratch. is nearest to the sample is considered to identify the Clearly, the technical challenge in operationalis- matching category. This kind of information pro- ing language strategies, language subsystems, and cessing can be operationalized easily by neural net- their application in situated embodied language works (for example radial basis function networks) games is enormously challenging. Nevertheless, the or by a straightforward computational implementa- state of the art in computer science and Artificial tion of nearest-neighbor computation [78]. Intelligence research is sufficiently advanced that it When we restrict ourselves to basic color terms, can be done. the linguistic component is straightforward as well. It consists of a bi-directional associative memory that associates color categories (or more precisely 5.2 Origins of a color language subsystem their prototypes) with color words. When the speaker needs to name a color he should first find Let me now very briefly introduce an example to the nearest color prototype and then look up the illustrate already these three methodological stages color term in his lexicon associated with this proto- for the domain of color terms, based on simulated type. When the listener wants to understand a color language game experiments reported in [78], [3], [9], term, he looks up which prototype corresponds to and [8]. Many other case studies, including for more this name in his own lexicon and then searches in complex grammatical phenomena such as case gram- the context for an object that matches the closest to mar [96], or tense and aspect marking [30] are cur- the prototype that was communicated. rently under development. I use the color example There has been a lot of empirical psychologi- because it is not overly complex. cal research in order to find out what color proto- types the speakers of a given language employ and these data can directly be plugged into an opera- 5.2.1 Identify a language subsystem tional model of color comprehension and production Most languages have a vocabulary of basic color processes. For example, Lillo and colleagues [48] terms. These have been studied intensively by an- have used the CIE L*u*v color space (where L is thropologists, often by asking human subjects to the brightness dimension and u and v are dimen-

18 sions modeling the human opponent channels) to to the correct object. identify the prototypes that Spanish speakers asso- The Color Naming Game can be played with ciate with the eleven basic color categories in Span- contexts consisting of Munsell chips so that we can ish: blanco, negro, rojo, verde, amarillo, azul, mar- approach anthropological test conditions. But they ron, rosa, naranja, morado, gris. For example, the can also be played with real world objects, and in- prototype for “verde” (green) is located at the point deed there have already been experiments where the hL = 44.85, u = 38.42, v = 29.15i in the L*u*v color whole interaction is implemented with autonomous space. These values can be used to simulate in a re- robots that play the game about the colorful objects alistic fashion the color naming of Spanish speakers. they encounter in the room [9] (Figure 4). There have also been experiments that use pictures taken from natural scenes [3], [2]. The experimental re- 5.2.2 Identify its function in communication sults shown in the remainder of this section come An inventory of perceptually grounded color proto- from experiments with Munsell chips carried out by types for Spanish and a lexicon associating these Joris Bleys. color prototypes with their Spanish names constituts Figure 5 shows the result of a language game ex- the basic color language subsystem of Spanish, so we periment where two agents were endowed with the can now turn to the second stage of the methodology basic Spanish color language subsystem discussed and address the question: What is this language sub- earlier. Each game involves a randomly assembled system for? If color terms have no purpose whatso- set of Munsell chips. One Munsell chip is chosen as ever in human communication we would not expect topic by the speaker and its color named. The lis- to find them in human languages. tener needs to guess which chip was intended and There are in fact several functions. Here I will the game is a success if the listener was able to do just focus on one, namely reference: A speaker can this. We see that the agents are not always success- use basic colors to draw the attention of the hearer ful (average success rate is 90 %) because in some to an object in the world by naming its distinctive cases the color of the topic chosen by the speaker basic color. This kind of interaction can be captured is so close to that of another chip that they can no quite succinctly in a language game, known as the longer be distinguished through basic color proto- Color Naming Game [78]. types. This would be the case if the second or the The Color Naming Game assumes an open-ended fourth chip in the context shown as inset in Figure set of possible contexts consisting of objects of dif- 5, because they are both close to the prototype for ferent colors. To play the game, two agents are ran- “gris” (grey). In that case a more sophisticated color domly chosen from the population. One of them expression (such as “slightly light grey”) would need takes on the role of speaker and the other that of to be utilized which goes beyond the language strat- hearer. They then go through the following interac- egy under investigation here. On the other hand, if tion: the first chip is chosen, it matches distinctly with - 1. The speaker chooses one of the objects in the the prototype for “amarillo” (yellow) and a success- context as topic. ful game is possible. - 2. The speaker categorizes the color of the topic based on his internal inventory of color prototypes. - 3. The speaker looks up in his lexicon the color 5.2.3 Understand how the language subsystem gets term associated with the prototype and transmits built, given a strategy this to the hearer. Figure 5 shows that stage 1 and 2 could be success- - 4. The hearer looks up the prototype associated fully concluded: We were able to operationalise the with this color term in his own lexicon. language subsystem of interest (the Spanish basic - 5. He then selects the object whose color matches color terms) and show its utility in the context of a the closest with the prototype and points to this ob- particular language game (the Color Naming Game). ject. Next we investigate what kind of language strategy - 6. The speaker checks whether the object pointed is able to learn and bootstrap such a color system. at is the one he originally chose. If that is the case The first step is to identify and operationalise he signals success. functions for learning the basic color prototypes used - 7. If it is not the case he signals failure and points in Spanish, and for acquiring the words for these pro-

19 Figure 4: It is possible to carry out language game experiments with autonomous robots, which implies that the script of the game, the perception and action, the parsing and production, and the learning, expansion and alignment functions of language strategies are completely operationalised.

Figure 5: Graph showing the results of 2 simulated Spanish speakers playing a series of 500 Color Naming Games (x-axis) establishing the base-line performance for this micro-ecology. The running average of com- municative success is shown on the y-axis. The inset shows a typical example context. About 90 % success is reached with the Spanish basic color terms.

20 totypes. Learning color words is relatively straight- Figure 6 shows the results of an experiment to forward once the color categories have been mas- test both of these learning mechanisms. It is an tered: When the listener encounters a word not experiment involving two artificial agents. One is heard before or when he uses a word in a wrong acting as tutor and has been programmed with the way (from the viewpoint of the speaker), the game Spanish color language subsystem (from stage 1). is a failure and the speaker points to the topic he The other is a novice. He starts without an inven- chose. Because the listener can already categorise tory of perceptually grounded categories and with- the color of the topic, he can infer the possible mean- out a lexicon for expressing them. The novice has ing of the uttered word and store that in his lexicon. been programmed with the learning strategy dis- The association is only an hypothesis which has to cussed above. We see that the learning strategy is be confirmed by further interactions. Agents must entirely effective. The novice reaches quickly the therefore maintain a score between words and mean- same level of performance as the tutor. ings. They should prefer to use the association with We next turn to the alignment and expansion the largest score because it is their best hypotheses strategies. In this particular case, the alignment so far. If there is a successful game, then the score strategy is already part of the learning strategy dis- of the association that was used is increased and the cussed earlier. After every game, speaker and hearer competing associations (other words with the same adjust the scores of the associations in their lexicon meaning for the speaker or other meanings for the and they shift prototypes and keep track of their same word for the hearer) are decreased. If there is utility. Consequently their language subsystems be- an unsuccessful game, the score of the used associ- come more and more aligned. This is shown with ation is decreased. This lateral inhibition dynamics an additional graph in Figure 6 (interpretation vari- has now been widely employed and studied as an ance) which displays the average distance between adequate vehicle for modeling how a population set- the color prototypes of novice and tutor. The dis- tles a convention [1]. It can be operationalized using tance (measured in the L*u*v space) never becomes neural networks (such as bi-directional associative 0.0 but is small enough to support successful com- memories [41]) or straightforward computational im- munication. plementations. The expansion strategy needed by the speaker is How are color prototypes learned? When the lis- reminiscent of the learning strategy used by the lis- tener encounters a new word and gets feedback on tener. When the speaker is unable to find a distinc- the topic from the speaker, he categorizes the topic tive prototype, for example because several samples himself to guess the possible meaning. But it is pos- including the topic are equidistant from the same sible that there is no distinctive category yet. In that prototype, then the speaker should introduce a new case, the listener should introduce a new prototype, prototype by taking the topic as initial seed. And using the sample that acted as topic as a seed. As when the speaker has no word yet to name a distinc- in the case of words, the new prototype is only an tive prototype, he can invent a new name, for exam- hypothesis which needs to be confirmed by further ple by choosing a random combination of syllables, interactions. Agents must therefore maintain a score and adding a new association to his lexicon. Once about the utility of a prototype and they should shift introduced, the learning strategy and the lateral in- the prototype in the face of new evidence. For ex- hibition dynamics of the naming game do their work ample, they should shift the prototype that was suc- and the invention potentially spreads in the popula- cessful in the language game slightly in the direction tion. of the topic. When agents keep doing that over a Figure 7 shows that all this works beautifully. A series of games, their prototypes will not only be- population of five agents starts from scratch with come more similar but also they will become and zero communicative success but rapidly reaches a remain adaptive to the situations they effectively en- successrate comparable to the Spanish basic color counter in their world. The information processing system. There is no fixed limit on the number of required to operationalise this learning strategy can basic color prototypes and so the agents keep refin- be achieved with neural networks (for example ra- ing their ontologies and keep inventing new words, dial basis function networks [78]) or with a straight- so that they actually reach a higher level of suc- forward computational implementation of the same cess than the Spanish system which has only 11 ba- functions. sic color terms. The most common words in one

21 Figure 6: Graph showing baseline communicative success (left y-axis) with 2 simulated Spanish speakers as in Figure 5 and a sequence of language games (y-axis) between a novice and a tutor, using the same contexts as for the baseline. We see that the novice progressively reaches the same level of communicative success as the base line. the same number of eleven basic color terms as the simulated Spanish speakers (right y-axis), and interpretation variance drops showing that the categories of the agents become similar.

1 50

45

0.8 40

35

0.6 30

25

0.4 20 communicative success 15

0.2 10 lexicon size/ontology size/interpretation variance 5

0 0 0 100 200 300 400 500 600 700 800 900 1000 number of interactions/agent communicative success ontology size lexicon size interpretation variance

Figure 7: Experiment in which a population of 5 agents self-organizes a color language system from scratch, given a strategy. The graph shows communicative success (left y-axis), average lexicon size (right y-axis) and interpretation variance. After about 500 language games (per agent) the agents reach a similar level of communicative success as the (simulated) Spanish speakers.

22 run of the experiment are: vamasi (greenish), fidate 6.1.1 Reconstruct the origins of a language strategy (brownish), bamoru (black), bamova (bluish grey), Grammaticalisation studies already provide a lot of riveke (purple), and kenafo (brownish grey). Every empirical data on possible evolutionary pathways for time the experiment is run another color language, language subsystems [37] and creole formation often including another set of perceptually grounded color shows quite clearly how a particular language fea- categories, emerges. ture that eroded away from a source language may Figure 8 shows the color prototypes for these re-emerge in a new form [53]. These studies suggest words after 1000 games (left) and after 2500 games that there are basically two ways in which a language (right). We see that the prototypes of the 5 agents user may come to a new language strategy. start to look more and more similar. This emerging Some language strategies appear to be exapta- coherence is remarkable because there is no central tions from already existing strategies. For example, supervisor nor controlling authority, no prior knowl- there is linguistic evidence that language strategies edge of the categories or lexicon, and no telepathic for temporal relations are often exapted from strate- relation between the agents. The color categories in gies that are effective for spatial relations: past is this artificial language are not identical to those of ‘before’ and future is ’after’ [35]. Strategies for spa- Spanish speakers and that cannot be expected. In tial language are often exapted from body language, fact, it is still an open question what additional con- before is “front” (from the front of the body) and af- straints need to be imposed on the micro-ecology of ter is “back” (from the back of the body) [59]. Exap- the agents or their perceptual and cognitive appara- tations also occur at a linguistic level, for example, tus in order to see color language subsystems that demonstratives often give rise to determiners [26]. exhibit the kinds of trends seen in human color lan- A language strategy may also originate de novo guage subsystems [3], but at least we now have a through a combinatorial search process, whereby very clear framework to investigate this. cognitive functions are configured and tried out until a combination is found that achieves the communica- tive goal of the speaker. This combination can then 6 Origins of Language Strategies be stored as a ‘chunk’, to be reused if a similar com- municative goal arises, and it can become conven- So far we looked at how one can identify a concrete tionalized through a grammatical construction that language subsystem (i.c. the Spanish basic color signals this chunk and provides slots for the infor- terms), isolate its communicative function (through mation that is needed to use it. The hearer in turn the Color Naming Game), and deconstruct the lan- can also engage in a combinatorial search process guage strategy that gives rise to such a system. Go- in order to reconstruct the possible meaning com- ing through the first three stages of the methodol- bination that may have been used by the speaker, ogy is not simple, particularly when more complex using constraints coming from partially understand- language phenomena like case grammar or subordi- ing what the speaker said and using the context and nation are being tackled, but it is doable and we pragmatic feedback. He can then try to abduct the can objectively test each stage. Once we are able meaning and function of unaccounted for features to deconstruct a language strategy, we can demon- found in the utterance in order to formulate reason- strate in repeatable simulated language game exper- able hypotheses to be tested later [77]. iments how a population of individuals is able to How can we develop and validate models for the learn and self-organize a language subsystem based origins of a particular language strategy? The first on this strategy. step is to set up language game experiments in which a particular language strategy is needed or would be beneficial, but the strategy is not implemented 6.1 Stages II explicitly as part of the initial state of the agent, Now we turn to the two remaining stages: (4) Re- in contrast to experiments in stage 3. Instead, the construct the possible origins of a language strategy, agent is given other strategies and operations to and (5) show that language subsystems built with a exapt strategies for new purposes, or he is supplied strategy have a selective advantage, so that we can with operationalisations of cognitive functions with explain why a particular strategy is adopted by a which strategies can be assembled and tried out. language community. Each agent has to discover himself the new strat-

23 Figure 8: Basic color prototypes of each agent (from top to bottom) for the most dominant words. Left: after 1000 games. Right: after 2500 games. The alignment strategy causes not only the words but also the prototypes to become similar and this will increase the chance of communicative success. egy and self-organisation based on the selectionist in the micro-ecology in which they are used, proto- impact of communicative fitness on the usage of a types are aligned, and words are introduced and then strategy then ensures that the strategy is adopted also aligned using the lateral inhibition dynamics. by the language community. This sort of experi- Figure 9 (taken from [8]) shows the outcome of a ments are in a very early stage at this point, but Color Naming Game experiment similar to the one there are already a few examples. One example in discussed earlier where such a brightness based strat- line with the Color Naming Game discussed earlier egy has been employed. Contexts with Munsell chips goes as follows (see [8] for more details). are generated and agents play a Color Naming Game There are languages which do not have a worked in which the hearer has to guess the chosen topic out hue-based basic color vocabulary but utilise in- but now based on its brightness instead of its full stead a rich set of distinctions based on light reflec- color. Despite the fact that the Munsell chips have tion (brightness), with words like dark (black), light a hue component, agents manage to evolve a lexicon (white), pale, shiny, etc. Moreover a common trend of brightness distinctions and have almost the same for these brightness-based languages is to start lexi- performance as for the full color space. This explains calising categories into the color space formed with perhaps why color-blind subjects have hardly any the brightness and hue channels, particularly if a dye problem to cope with color language, except when industry develops or hue becomes culturally impor- it really matters, i.e. when there is almost equal tant [50]. Often there will be a shift in meaning brightness but different hue. of existing terms as brightness based terms become How can a full-color strategy emerge by exapta- reused for full-color meanings. For example, “yel- tion from a brightness-based strategy? Agents only low” has shifted from meaning “shiny” to the hue need to expand the dimensions of the color space, so distinction that we know today. as to incorporate the two opponent channels. For the A brightness-based strategy is from an informa- rest they can keep all other components the same. tion processing point of view not that different from This kind of exaptation mechanism appears to be a hue-based strategy. It builds up and aligns proto- quite common: A language strategy that works with types in the brightness color space as opposed to the a particular conceptual space is re-used by plugging full-color space but is otherwise entirely the same: in another space which is an expansion or contrac- New prototypes are introduced when discrimination tion of the earlier used space. Often not only the fails, prototypes are shifted to become more relevant strategy but also a large part of the lexical and

24 1 80

70

0.8 60

50 0.6

40

0.4 30 communicative success

20 0.2

10 lexicon size/ontology size/interpretation variance

0 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 number of interactions/agent communicative success ontology size lexicon size interpretation variance

Figure 9: In this experiment, a population of 5 agents bootstraps a language subsystem that names bright- ness distinctions. A similar prototype-based strategy is used as for the full-color space (Figure 7) but now only the brightness dimension is employed. grammatical system already developed are exapted, 6.1.2 Show the selectionist advantage of a language which explains why color terms originally used for subsystem brightness start to shift towards the expression of Novel genes or gene networks may originate by ge- full-color. nomic operations or recombinations, but they will not thrive in the gene pool unless there is a selective Here is another example of the same exaptation advantage for the phenotypic traits that they help to mechanism but now applied to an entirely differ- build and maintain. Similarly, novel language strate- ent domain. In many languages posture verbs are gies may originate by novel recombinations of cog- metaphorically reused for spatial positions (as in nitive functions or by variations made on existing “the bottle stands on the table”) [44]. [86] report lan- language strategies. But these novel strategies will guage game experiments showing how the origins of not spread or survive in a language community un- a strategy could be explained. In a first stage, agents less the language systems they help to build yield a build up a language for body postures by playing selectionist advantage for the population. action games, in which a speaker asks the hearer As explained earlier, in the case of language, to execute an action and the game is a success if selective advantage translates into various factors the hearer indeed carries out this action (see Figure which all show up through the success rate of ut- 10). To be successful in this game, agents not only terances. The first point is expressive power. A lan- have to develop body models of their own actions, guage subsystem must handle a communicative need for example by looking at their body or performing of the language community that was not handled be- actions before a mirror [84], but also of the visual fore. A communicative need includes both specific features that characterise these actions when they communicative goals as well as a semantic domain are carried out by others [85]. Once a vocabulary of in which these goals have to be reached. Given that body posture terms is in place, a strategy for spatial language is an adaptive system, we expect that when language can emerge by reusing the visual features the communicative needs of a language subsystem that are used to recognise body postures work for disappear the system erodes and will become forgot- categorising the postures of inanimate objects and ten, particularly in subsequent generations. On the by importing even the words used to lexicalise these other hand, when new needs arise we expect to see categories [86]. a surge of language creativity where novel strategies

25 0.12 blue block turquoise block 0.1

0.08

0.06 value 0.04

0.02

0

-0.02 mu20 mu11 mu02 mu30 mu21 mu12 mu03 moments

0.3 red block 0.25 yellow block

0.2

0.15 Values 0.1

0.05

0

-0.05 mu20 mu11 mu02 mu30 mu21 mu12 mu03 moments

0.3 pink trumpet

0.25

0.2

0.15

0.1

Coordinate Value 0.05

0

-0.05

-0.1 mu20 mu11 mu02 mu30 mu21 mu12 mu03 moments

Figure 10: Experiments for evolving language about body postures. Left: Robots play action games, asking each other to perform certain actions. They thus learn connections between a command, a visual image of that command, and their own body image. Right: The visual features acquired in action games and the words for them are then re-used for talking about the spatial posture of non-animate objects. appear and partial systems get built which then start to learn from and it is more efficient from a cognitive to compete with each other through the selectionist viewpoint because fewer variants must be stored. dynamics. The third factor is cognitive effort. Fluid speech Communicative success also depends on whether is extraordinarily fast and consequently language the language has precise enough ways to express the processing needs to be highly optimized. More- meanings that need to be expressed. Listeners are over it needs to be occasionally semi-automatic so assumed to be intelligent and hence language can be that the brain can focus on other tasks at the same an inferential coding system relying on the context time. Speakers and hearers can monitor internally and additional inference to fill in gaps [69]. Conse- the amount of cognitive effort they require, in the quently natural language can be much more efficient sense of tracking the size of the search spaces needed than formal languages (like computer programming for parsing and production or for semantic concep- languages) in which everything needs to be coded ex- tualisation and interpretation, and the amount of plicitly. However this brings a risk of possible misun- memory they need for storing intermediary stages derstandings. A language strategy that introduces in language processing. The diagnostics that come more precision for important meanings may there- with a language strategy can take these data into fore give a significant communicative advantage. account and expand or adapt the language inven- The second factor contributing to selective ad- tory to avoid excessive search or memory usage in vantage is sharing. If the choices within a language the future. Those language strategy that are better subsystem are shared between speaker and hearer, in doing this should be preferred. then they will obviously have a higher chance of com- Learnability also affects selectionist advantage municative success. Often there is no other advan- but in an indirect way. Although it is remarkable tage between competing strategies or competing lin- how very complicated language subsystems with lots guistic choices given a strategy, then the simple fact of idiosyncracies can nevertheless be acquired by na- that they are widely adopted. Absolute identity of tive speakers, it is nevertheless the case that a lan- language subsystems is an unattainable goal because guage subsystem which is more easy to learn will all speakers build up their own language subsystems have a higher chance to survive in future genera- by local interactions with others. They do not have tions. This factor is different in the sense that lan- a global overview and there is no telepathy. Neverth- guage users cannot deliberately ‘design’ their lan- less the alignment mechanisms that come with each guage subsystems to be learnable, it is just that lan- language strategy should be such that they lead to guage strategies and language subsystems that are progressively increased coherence. Moreover when more easy to learn will spread more easily in the the population shares the same system, it is easier population. In the case of communicative success, for new members to acquire the language subsystem expressive power, and cognitive effort, individuals that is already present because they get clearer data have ways to track these properties and adjust their

26 Figure 11: A spatial language game in which agents embodied as autonomous robots play a game of reference about objects in their environment using spatial positions and relations. Left is a typical scene for a game. The right shows the internal world model of each agent, top is their 3d reconstruction from their observer’s vantage point and bottom their bird’s eye reconstruction. language subsystems or language strategies accord- The goal of this experiment was to show that a strat- ingly. egy of using perspective reversal and of expressing perspective reversal gave a selective advantage to the How can the selective advantage of strategies be agents. investigated using formal models and language game experiments? There are two kinds of experiments The experiment proceeded in four stages: (1) that can be done. The first type of experiment tests First an experiment was done where agents see the in controlled conditions when a particular language world from the same angle, which means that a sin- strategy would be beneficial, by comparing the com- gle robot body was used and the agent animating municative success of agents as they learn or self- this body acted both as speaker and hearer. They organise a language without a particular strategy then used a language strategy, similar to the one for and then one with the strategy, given a particular the Color Naming Game, to self-organise an ontol- language game and a particular micro-ecology. This ogy of spatial relations and the language to express is similar to the approach that evolutionary biolo- them. This provided an indication of the base-line gists take when they go out in the field and measure performance, even though the experiment was en- the actual survival and reproductive success in dif- tirely unrealistic with respect to real world agents. ferent ecological circumstances, either for different (2) Then a language game experiment was conducted species which are phylogenetically related (as in the where the speaking and hearing agent see the world case of species of cichlid fish which have eggspots from their own private position. This lead to a dras- versus those who do not [66]) or different variants tic reduction in communicative success because spa- of the same species which is currently undergoing tial relations are often perspective-dependent. For speciation [56]. example, if the speaker says “the yellow block to the An example of such an experiment is reported left”, this could be to the left of the speaker or to in [87] for the domain of spatial language. The the left of the hearer. So the language game and the agents find themselves in a micro-ecology with vari- micro-ecology being used here are effective in isolat- ous blocks of the same color, boxes that have a front ing the expressive need that we want to investigate. and back as marked by bar-code like signs pasted on (3) Next the agents recruited perspective reversal as them, and global landmarks such as the big sign at- part of their conceptualisation of the scene and as tached to the wall as shown in figure 11. The agents part of their interpretation of what was said, however play a game of reference similar to the Color Nam- they did not express the perspective explicitly yet. ing Game, but now they try to draw attention using This is like saying “the yellow block to the left” and spatial expressions, such as “the block to your left”. leaving it up to the hearer to figure out from which

27 perspective this is valid. A significant improvement nicative success. Those conceptualisations and lin- of performance could be observed even though the guistic constructions that lead to higher communica- hearer had to engage in more cognitive effort as he al- tive success, greater expressive power, or less cogni- ways had to consider the different views from which tive effort, are amplified in the population and the the scene could be conceptualized (see Figure 12). language strategies with which they are built are pre- (4) In a final experiment, the agents used an ex- ferred as they have a higher ‘communicative fitness’. panded strategy, in which they made the viewpoint Languages thus become and remain adaptive to the from which they conceptualized reality explicit as needs and ecologies of their users. well, so that they now could come up with phrases The methodology rests on a series of stages in like “the block to your left”. Only with this strategy which an existing language subsystem and the strat- did we see consistent communicative success as well egy underlying it is deconstructed in terms of formal, as a minimisation of cognitive effort [87]. operationalized models that can be tested through The second type of experiments investigates the language game experiments that capture the eco- selectionist dynamics at the level of strategies. Dif- logical conditions in which the language subsystem ferent language strategies for similar ecological con- is relevant. Once it can be shown how a language ditions and communicative goals are made available subsystem may self-organise in a population given to the agents in a population at the start of the a language strategy, attention shifts to how the un- experiment and the question is whether and how derlying strategy could have originated and achieved they will be able to select the strategy that is most enough communicative fitness to thrive in the pop- adapted to their micro-ecology. If the selectionist ulation. logic has been captured properly, we should see that There are a number of topics which I have not long term communicative success with language sub- discussed in this paper, although they certainly de- systems built by one strategy should give that strat- serve more attention. First of all, I did not focus egy a selectionist advantage so that it comes to dom- on the question of the origins of language itself but inate in the population. An example of such an ex- rather on the origins of language subsystems and periment for the domain of color is reported in [8], the strategies underlying them. The two topics may using the brightness-based and full-color language however not be so far apart as often assumed. Once strategies discussed earlier. Results are shown in we understand the origins of language strategies by figure 13. We see that in a first phase the brightness- the recruitment of generic cognitive functions, we based strategy is winning, in the sense that it was should be able to wind the clock back entirely to used more often by the agents to invent or interpret a pre-language stage and do language game experi- new words. Progressively the full-color space strat- ments where the agents start with no language at all egy overtakes this initial advantage. Interestingly, and progressively bootstrap more and more complex the same word can temporarily be used by one agent symbolic communication systems. Of course there using the brightness-based strategy and by another are many other factors that had to be in place for agent using the full-color strategy (as happened in the ‘first language’ to arise, such as social coopera- the evolution of English with the word “yellow” for tion or high brain plasticity. They are being studied example). in the context of how communication may emerge in general and under what conditions symbolic com- munication might be viable and needed [25], [40]. Second, I did not discuss the role of genetics in 7 Outlook language, but emphasized instead cultural evolution. The paper has proposed a methodology for the I did not want to imply that there is no such role. field of Evolutionary Linguistics. The methodology The cognitive functions that are recruited for as- assumes a Darwinian selectionist logic but trans- sembling language strategies are relevant for a wide posed to the linguistic domain. The distinction be- range of tasks and genetics and evolution by nat- tween phenotype and genotype is mapped onto the ural selection must have played a role to establish distinction between language subsystems and lan- them. Also the overall plasticity and computational guage strategies. Natural selection, based on suc- architecture needed for cognition in general, includ- cess in survival and reproduction in an ecosystem, is ing the capacity to recruit functions and to retain mapped onto linguistic selection based on commu- those configurations that are beneficial for the tasks

28 1.0 50 1.0 50 1.0 50 0.9 0.9 0.9 A A 0.8 40 0.8 40 0.8 40 0.7 0.7 0.7 0.6 30 0.6 30 0.6 30 C 0.5 0.5 0.5 0.4 20 0.4 20 0.4 20 l e x i c o n s z ( B ) l e x i c o n s z ( B ) l e x i c o n s z ( B ) 0.3 0.3 0.3 B B c o m u n i a t v e s ( A ) B c o m u n i a t v e s ( A ) 0.2 10 0.2 10 0.2 10 0.1 0.1 0.1

A c o m u n i a t v e s ( A ) / f r h C 0.0 0 0.0 0 0.0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 number of games number of games number of games

Figure 12: Three experimental situations to prove the selective advantage of using perspective reversal. The y-axis shows the total number of games in the population (only two agents per game). Case 1 (left) : Agents see the world from the same perspective. They self-organise a spatial language subsystem from scratch. Communicative success (A) reaches 90 % success. B is average lexicon size. Case 2 (middle): Agents are embodied and see the world from different perspectives. Communicative success drops dramatically. Case 3 (right): Speakers use perspective reversal to select the ‘surest’ viewpoint and hearers use perspective reversal to make sense of the utterance. Communicative success (A) now reaches almost 90 % again. Graph (C) shows an internal measure of cognitive effort which drops again if perspective reversal is also expressed.

1

0.8

0.6

0.4

0.2 communicative success/coherence/fitness/usage

0 0 200 400 600 800 1000 number of interactions/agent communicative success usage (brightness) strategy coherence fitness (full colour space) fitness (brightness) usage (full colour space)

Figure 13: This figure shows the results of an experiment where a brightness-based strategy competes with a full-color strategy in a population of agents. Communicative success hovers around 85 % even though there is a shift where brightness first dominates and is then overtaken (after about 410 games per agent) by the full-color strategy. The graphs show the communicative fitness (the running average tracked by each agent) as well as the frequency of choice for each strategy. The evolution of strategy coherence, which is how far the different agents share the same language strategy, is shown as well. It starts from low values to reach 70 %. the individual has to engage in, have without doubt orientation, etc. The language game experiments dis- a genetic basis. I only argued that there are no cussed here help to show that all of this could in language-specific genes, in other words genes that principle emerge. A process of co-evolution between encode the distinction between nouns and verbs, language and meaning is not only feasible but the common sentence patterns (like SVO) in sentences, most effective way in which a population can de- innate ways to express argument structure, innate velop a communication system that is adapted to color categories, spatial distinctions based on body- their ecological needs.

29 Third, I did not discuss the equivalent of spe- logic applies to language. The wave of language ciation in language, i.e. the formation of different game experiments that have become operational re- languages from a common proto-language, or the in- cently is showing that the methodology outlined here fluence of one language on another due to language is feasible and can be put into practice with today’s contact. But it is clear that the cultural selectionist technologies. But we still have decades of exciting re- framework as summarized in Figure 2 explains this search before us to even begin confront the incredible as well. Similar to biological speciation, we should variation and richness of human natural languages. expect that a language subcommunity can start to diverge, partly due to the inherent ‘language drift’ that is a side effect of the cultural dynamics that 8 Acknowledgements sustains language, and partly because the subcom- This article is based on research conducted by teams at munity optimizes other aspects of their language in the University of Brussels (VUB Artificial Intelligence order to deal with the communicative challenges oc- Laboratory) and the Sony Computer Science Laboratory curring in their ecology. For example, the global in Paris, with grants from the Sony Computer Science landmark scheme for spatial relations might work Laboratory, the Belgian National Science Foundation, well if everybody is aware of such global landmarks IWT, and additional grants from the European Com- (like the mountain and the sea) but if the population mission through the ECAgents and ALEAR projects. I moves to a continent where such landmarks are no thank all current members of the team (Joris Bleys, Ka- trien Beuls, Thomas Cederborg, Joachim De Beule, Carl longer present, this solution would not be adequate. Jacobs, Martin Loetzsch, and Pieter Wellens in Brus- Evolutionary Linguistics is only in its early sels; and Katya Gerasymova, Vanessa Micelli, Damien phases. There is a large amount of data both about Munch, Simon Pauw, Michael Spranger, Remi van Trijp language evolution, going back centuries of histor- in Paris), as well as past members (Tony Belpaeme, Bart ical linguistics research, and about the semantics de Boer, Bart de Vylder, Frederic Kaplan, Angus McIn- and grammar of individual languages, thanks to tyre, Pierre-Yves Oudeyer, Joris Van Looveren, Paul Vogt, Jelle Zuidema) for their technical contributions, the meticulous work by many descriptive linguists. inspiration, and discussions. The Color Naming Game The methodology proposed here attempts to connect experiments used as an example in this paper were car- this empirical work with formal models of the in- ried out by Joris Bleys as part of his Ph.d thesis research. formation processing that ‘linguistic agents’ need to This paper would not have happened without the gener- produce, comprehend, learn, and self-organise lan- ous support of the Wissenschaftskolleg in Berlin and the guages, and to study how a Darwinian selectionist discussions with many colleagues at this institute.

References 1. Baronchelli, A., V. Loreto and L. Steels (2008) In-depth analysis of the Naming Game dynamics: the homogeneous mixing case. Int. J. of Mod. Phys. C, 19(5), 785-812. 2. Puglisi, A., A. Baronchelli, and V. Loreto (2008) Cultural route to the emergence of linguistic categories. Proceedings of the National Academy of Sciences, 105(23), 7936-7940. 3. Belpaeme, T. and Bleys, J. (2005) Explaining Universal Color Categories Through a Constrained Acqui- sition Process. Adaptive Behavior, 13(4), 293-310. 4. Berlin, B. and P. Kay (1969) Basic color terms: Their universality and evolution. University of California Press, Berkeley Ca. 5. De Beule, J. and L. Steels (2005) Hierarchy in Fluid Construction Grammar. In: Furbach, U. (eds) (2005) Proceedings of KI-2005. Lecture Notes in AI 3698. Springer-Verlag, Berlin. p.1-15.

6. Bickerton, D. (1984) The language bioprogram hypothesis. Behavioral and Brain Sciences, 7, 17-388. 7. Bickerton, D. and E. Szathmary (2009) (eds.) Biological Foundations and Origin of Syntax. Strungmann Forum Reports, vol. 3. The MIT Press, Cambridge Ma.

30 8. Bleys, J. and L. Steels (2009) Linguistic selection of language strategies. A case study for colour. Pro- ceedings of European Conference on Artificial Life. Springer Verlag, Berlin. 9. Bleys, J., M. Loetzsch, M. Spranger, and L. Steels (2009) The Grounded Color Naming Game. Workshop Proceedings of 18th IEEE International Symposium on Robot and Human Interactive Communication. (Ro-man 2009). 10. Brakefield, P., J. Gates, D.Keys, F. Kesbeke, P. Wijngaarden, A. Monteiro, V. French, S. Carroll.(1995) Development, plasticity and evolution of butterfly eyespot patterns. Nature, 384(6606), 236-242. 11. Brighton, H. (2005) Linguistic Evolution and Induction by Minimum Description Length. In Werning, M. and Machery, E. (eds.) The Compositionality of Concepts and Meanings: Applications to Linguistics, Psychology and Neuroscience. Frankfurt: Ontos Verlag.

12. Briscoe, T. (ed.) Linguistic Evolution through Language Acquisition: Formal and Computational Mod- els. Cambridge University Press, Cambridge, UK. 13. Cangelosi, A. and D. Parisi (eds.) (2001) Simulating the Evolution of Language. Springer-Verlag, Berlin. 14. Camazine S., J.L. Deneubourg, N. Franks, J. Sneyd, E. Bonabeau and G. Theraulaz (2001). Self- Organization in Biological Systems. Princeton University Press, Princeton NJ.

15. Cangelosi, A. (2005) The emergence of language: neural and adaptive agent models. Connection Science, 17(3-4), 185-190. 16. Casson, R. (1997) Color shift: evolution of English color terms from brightness to hue. In: Hardin, C. and L. Maffi (eds.) Color categories in thought and language. Cambridge University Press, Cambridge. p. 224-239 17. Chomsky, N. (1965). Aspects of the Theory of Syntax. MIT Press, Cambridge Ma. 18. Christiansen, M., and S. Kirby (2003). Language evolution. Oxford University Press: Oxford. 19. Clark, H. (1996) Using Language, Cambridge University Press, Cambridge.

20. Croft, W. (1998) Explaining language change. Pearson Education. London 21. Croft, W. (2008) Evolutionary Linguistics. Annual Review of Anthropology Vol. 37: 219-234 22. Daelemans, W. and A. van den Bosch (2005) Memory-Based Language Processing. Cambridge University Press, Cambridge.

23. Dawkins, R. (1976) The Selfish Gene. Oxford University Press, Oxford. Chapter 11. 24. Deacon, T. D. (2006) Evolution of language subsystems in the human brain. In J. Kaas (ed.), Evolution of Nervous Systems. Vol. 5. The Evolution of Primate Nervous Systems. 25. Dessalles, J-L (2007) Why we talk – The evolutionary origins of language. Oxford University Press, Oxford. 26. Diessel, H. (1999) Demonstratives. Form, function, and grammaticalization. John Benjamins, Amster- dam. 27. Dik, S. (1997) The Theory of Functional Grammar, Vol. 1 and 2 Mouton-de Gruyter, Berlin.

28. Dobzhansky, T. (1973) Nothing in Biology Makes Sense Except in the Light of Evolution. The American Biology Teacher, 35, 125-129.

31 29. Fairchild, D. (1997) Color appearance models. Addison Wesley Longman, Reading.

30. Gerasymova, K., L. Steels and R. van Trijp (2009) Aspectual Morphology of Russian Verbs in Fluid Construction Grammar. Cognitive Science Conference, Amsterdam. p. 1370-1375. 31. Goldberg, A.E. (1995) Constructions.: A Construction Grammar Approach to Argument Structure. University of Chicago Press, Chicago

32. Halliday, M.A.K. (1994) Introduction to functional grammar. Edward Arnold, London. 33. Haspelmath, M., M. S. Dryer, D. Gil, and B. Comrie (eds.) (2004) The World Atlas of Language Structures Online. Max Planck Digital Library, Munich. (http://wals.info/) 34. Hauk, O., Johnsrude, I., and Pulvermuller, F. (2004). Somato-topic representation of action words in human motor and pre-motor cortex. Neuron, 41, 301-307. 35. Haspelmath, M. (1997) From Space to Time: Temporal Adverbials in the World’s Languages. Lincom Europa, Muenchen. 36. Heine, (1997) The cognitive foundations of grammar. Oxford University Press, Oxford.

37. Heine, B. and T. Kuteva (2008) The Genesis of Grammar. A reconstruction. Oxford University Press, Oxford. 38. Hockett, C. F. (1960) The origin of speech. Scientific American, 203:88-96. 39. Hopper, (1991) Emergent Grammar. In: Traugott, E. and Heine, B. (1991) Approaches to Grammati- calization. Volume I and II. John Benjamins Publishing Company, Amsterdam.

40. Knight, C. (2007) Language co-evolved with the rule of law. Mind and Society. 41. Kosko, B. (1988) Bidirectional Associative Memories. IEEE Transactions on Systems, Man, and Cyber- netics, 18(1), 49-60. 42. Labov, W. (1994) Principles of Linguistic Change. Volume 1 and 2. Basil Blackwell, Oxford.

43. Langacker, R.W. (1991) Foundations of cognitive grammar. Stanford University Press, Stanford Ca. 44. Lemmens, M. (2002) The semantic network of Dutch posture verbs. In J. Newman (Ed.) (2002) The linguistics of sitting, standing and lying. John Benjamins, Amsterdam. 45. Loreto, V. and L. Steels (2007) Emergence of language. Nature Physics, 3(11), 758-760.

46. Hurford, J. R., Studdert-Kennedy, M. and C. Knight (eds), Approaches to the Evolution of Language - Social and Cognitive Bases, Cambridge University Press, Cambridge. 47. Levinson, S. (2003) Space in Language and Cognition: Explorations in Cognitive Diversity. Cambridge University Press, Cambridge.

48. Lillo, J., Moreira, H., Vitini, I., and J. Martın (2007) Locating basic Spanish colour categories in CIE L*u*v* space: identification, lightness segregation and correspondance with English equivalents. Psicolog- ica 28 (2007) 21–54 49. Lyon, C., C. Nehaniv, and A. Cangelosi (eds) Emergence of Communication and Language. Springer Verlag, Berlin.

50. MacLaury, R. (1992). From Brightness to Hue. An explanatatory model of color-category evolution. Current Anthropology, 33(2), 137-186.

32 51. Mallet, J. (2008). Mayr’s view of Darwin: was Darwin wrong about speciation? Biological Journal of the Linnean Society, 95, 3-16. 52. Minett, J. W. and Wang, W. S-Y. (2005) Language Acquisition, Change and Emergence: Essays in Evolutionary Linguistics. City University of Hong Kong Press: Hong Kong. 53. Mufwene, S. (2001) Competition and Selection in Language Evolution. Selection, 3(1).

54. Nishitani, N., M. Schuermann, K. Amunts, R. Hari (2005) Broca’s region from Action to Language. Physiology, 20, 60-69. 55. Nyogi, P. and R. Berwick (1997) Evolutionary consequences of language learning. Linguistics and Phi- losophy, 20, 697-719.

56. Nosil, P., L. Harmon, and O. Seehausen (2008) Ecological explanations for (incomplete) speciation. Trends in Ecological Evolution. 57. Nowak, M.A. and D.C. Krakauer (1999) The evolution of language. Proc.Natl.Acad.Scie. USA., 96, 8028-8033. 58. Nowak, M. (2006) Evolutionary Dynamics . The Belknap Press of Harvard University Press, Cambridge Ma. 59. N´u˜nez,R. and E. Sweetser (2006) With the Future Behind Them: Convergent Evidence From Aymara Language and Gesture in the Crosslinguistic Comparison of Spatial Construals of Time. Cognitive Science, 30 (3), 401-450.

60. Oudeyer, P-Y. and Kaplan, F. (2007) Language Evolution as a Darwinian Process: Computational Studies, Cognitive Processing, 8(1), pp. 21-35. 61. Palmer, F.R. (1994) Grammatical Roles and Relations. Cambridge University Press, Cambridge. 62. Pfister, K. (1999) Behavioral Strategies for Coping with Poisonous Plants. In: Launchbaugh, K., K. Sanders, J. Mosley (eds.). Grazing behavior of livestock and wildlife. Idaho Forest, Wildlife and Range Exp. Sta. Bull 70, University of Idaho. pp. 45-59. 63. Pinker, S. and P. Bloom (1990) Natural language and natural selection. Behavioral and Brain Sciences, 13, 707-784. 64. Roberson, D., I. Davies, J. Davidoff (2002) Color categories are not universal: replications and new evidence. In: Saunders, B., and J. van Brakel (eds) Theories, technologies, instrumentalities of color. University Press of America, Lanham, MD pp. 22-35 65. Rosch, E. (1975) Cognitive representations of semantic categories. Journal of Experi- mental Psychology, 104(3), 192–233. 66. Salzburger, W., T. Mack, E. Verheyen and A. Meyer (2005) Out of Tanganyika: Genesis, explosive speciation, key-innovations and phylogeography of the haplochromine cichlid fishes. BMC Evolutionary Biology 5(17). 67. Salzburger, W., I. Braasch and A. Meyer (2007) Adaptive sequence evolution in a color gene involved in the formation of the characteristic egg-dummies of male haplochromine cichlid fishes. BMC Biology 5(51).

68. Sick, S. (2006) Der Dativ ist dem Genitiv sein Tod. Kiepenheuer und Witsch, Koeln. 69. Sperber, D. and Wilson, D. (1995) Relevance communication and cognition. Blackwell Pub., London.

33 70. Steels, L. (1988) Meaning in knowledge representation. In: P. Maes and D. Nardi (eds.) Meta-Level Architectures and Reflection. North Holland, Amsterdam. 71. Steels, L. (2000) Language as a Complex Adaptive System. In Schoenauer, M., (ed.) Proceedings of PPSN VI, Lecture Notes in Computer Science, 1917. Springer-Verlag, Berlin. pp. 17-26. 72. Steels, L. (2000b) The Emergence of Grammar in Communicating Autonomous Robotic Agents. In Horn, W. (ed.) Proceedings of ECAI 2000, IOS Publishing, Amsterdam. pp. 764-769.

73. Steels, L. (2001) Language games for autonomous robots. IEEE Intelligent systems, 16 (5), 16-22. 74. Steels, L. (2002) Grounding symbols through evolutionary language games. In Cangelosi, A. and Parisi, D., editor, Simulating the evolution of language. Springer Verlag, Berlin. pp. 211-226. 75. Steels, L. (2003) Language-reentrance and the Inner Voice’. Journal of Consciousness Studies, 10(4- 5),173-185. 76. Steels, L. (2004) Analogies between Genome and Language Evolution. In: Pollack, J. (ed.) Artificial Life IX. The MIT Press. p.200-206. 77. Steels, L. (2004) Constructivist Development of Grounded Construction Grammars. In D. Scott, W. Daelemans and M. Walker (Eds.), Proceedings Annual Meeting of Association for Computational Linguis- tics Conference. Barcelona: ACL. pp. 9-16. 78. Steels, L. and T. Belpaeme (2005) Coordinating Perceptually Grounded Categories through Language. A Case Study for Colour. Target article. Behavioral and Brain Sciences. 24(6). 79. Steels, L. (2005) The Emergence and Evolution of Linguistic Structure: From Lexical to Grammatical Communication Systems. Connection Science. 17(3), 213-230. 80. Steels, L. (2007) Is symbolic inheritance similar to genetic inheritance? Behavioral and Brain Sciences, 2007. 81. Steels, L. (2007) The Recruitment Theory of Language Origins. In: Lyon, C., C. Nehaniv, and A. Cangelosi (eds) Emergence of Communication and Language. Springer Verlag, Berlin. pp. 129-151. 82. Steels, L. and Beule, J. D. (2006) Unify and Merge in Fluid Construction Grammar. In P. Vogt and et al., editors, Symbol Grounding and Beyond: Proceedings of the Third International Workshop on the Emergence and Evolution of Linguistic Communication, pp. 197-223. Springer-Verlag, Berlin. 83. Steels, L. and McIntyre, A. (1999) Spatially Distributed Naming Games. Advances in complex systems, 1(4), January 1999.pp. 301-323. 84. Steels, L. and M. Spranger (2008) The Robot in the Mirror. Connection Science, 20(4):337-358. 85. Steels, L. and M. Spranger (2008) Can body language shape body image? In: Proceedings of Artificial Life XI. The MIT Press, Cambridge Ma. pp. 577-584.

86. Steels, L. and M. Spranger (2009) How experience of the body shapes language about space. IJCAI-09. Pasadena, Ca. AAAI Press. 87. Steels, L. and M. Loetzsch (2009) Perspective Alignment in Spatial Language. In: Coventry, K., J. Bateman and T. Tenbrink (2009) Spatial Language in Dialogue. Oxford University Press. Oxford.

88. Steels, L. and P. Wellens (2006) How grammar emerges to dampen combinatorial search in parsing. In: Vogt, P. , Y. Suga, E. Tuci, C. Nehaniv (2006) (eds.) Symbol Grounding and beyond. LNAI 4211. Springer-Verlag, Berlin. pp. 76-88.

34 89. Stoll, S. (1998) The role of Aktionsart in the acquisition of Russian aspect. First Language, 18(54), 351-376. 90. Stromswold, K. (2001) The Heritability of Language: a Review and Metaanalysis of Twin, Adoption, and Linkage Studies. Language, 77(4), 647-723. 91. Szathmary, E. (2006) The origin of replicators and reproducers. Phil. Trans. R. Soc. B (2006) 361, 1761–1776.

92. Talmy, L. (2000) Toward a Cognitive Semantics: Concept Structuring Systems (Language, Speech, and Communication) The MIT Press, Cambridge Ma. 93. Tomasello, M. (1999). The cultural origins of human cognition. Harvard University Press, Cambridge, MA.

94. Van Kemenade, A (1987) Syntactic Case and Morphological Case in the History of English. Forist Publications, Dordrecht. 95. Van Langendonck, W. and W. Van Belle (eds) (1998) The Dative. Volume 1. Descriptive Studies. Volume 2. Theoretical and Contrastive Studies. John Benjamins, Amsterdam.

96. Van Trijp, R. (2008) The emergence of semantic roles in Fluid Construction Grammar. In: Smith, A, K. Smith and R. Ferrer i Cancho (eds.), Proceedings of the 7th EVOLANG. World Scientific Publishing, Singapore. pp. 346-353. 97. Van Trijp, R., J. Valenzuela, J. Hilferty, M. Garachana (2009) Contemporary strategy evolution in Spanish pronouns. in preparation.

98. Verschueren, J. (1999) Understanding pragmatics. A Hodder Arnold Publication, London.

35