segunda-feira, 4 de março de 2013
Language is a complex adaptative system
Language has a fundamental social function. Processes of human interaction along with domain-general cognitive processes shape the structure and knowledge of language. Recent research in cognitive sciences has demonstrated that patterns of use strongly affect how language is acquired, is used, and changes. These processes are not independent of one another but are facets of the same complex adaptive system (CAS). Language as a CAS involves the following key features: The system consists of multiple agents (the speakers in the speech community) interacting with one another. The system is adaptive; that is, speakers' behavior is based on their past interactions, and current and past interactions together feed forward into future behavior. A speaker's behavior is the consequence of compering factors ranging from perceptual constraints to social motivations The structure of language emerge from interrelated patterns of experience, social interaction, and cognitive mechanisms. The CAS approach reveals commonalities in many areas of language research, including first and second language acquisition, historical linguistics, psycholinguistics, language evolution, and computational modeling.
Introduction: Shared Assumptions
Language has a fundamentally social function. Processes of human interaction along with domain-general cognitive processes shape the structure and knowledge of language. Recent research across a variety of disciplines in the cognitive sciences has demonstrated that patterns of use strongly affect how language is acquired, is structured, is organized in cognition, and changes over time. However, there is mounting evidence that processes of language acquisition, use, and change are not independent of one another but are facets of the same system. We argue that this system is best constructed as a complex adaptive system (CAS). This system is radically different from the static system of grammatical principles characteristic of the widely held generativist approach. Instead, language as a CAS of dynamic usage and its experience involves the following key features: (a) The system consists of multiple agents (the speakers in the speech community) interacting with one another. (b) The system is adaptive; that is, speaker' behavior is based on their past interactions, and current and past interactions together feed forward into future behavior. (c) A speaker's behavior is the consequence of competing factors ranging from perceptual mechanics to social motivations. (d) The structure of language emerge from interrelated patterns of experience, social interaction, and cognitive processes.
The advantage of viewing language as a CAS is that it allows us to provide a unified account of seemingly unrelated linguistic phenomena. These phenomena include the following: variation at all level of linguistic organization; the probabilistic nature of linguistic behavior; continuous change within agents and across speech communities; the emergence of grammatical regularities from the interaction of agents in language use; and stage like transitions due to underlying nonlinear processes.
Language and Social Interaction
Language is shaped by human cognitive abilities such as categorization, sequential processing, and planning. However, it is more that their simple product. Such cognitive abilities do not require language; if we had only those abilities, we would not need to talk. Language is used for human social interaction, and so its origins and capacities are dependent on its role in our social life. To understand how language has evolved in the human lineage and why it has the properties we can observe today, we need to look at the combined effect of many interacting constraints, including the structure of thought processes, perceptual and motor biases, cognitive limitations, and socio-pragmatic factors.
Primates... reformed over time.
We adopt here a usage-based theory of grammar in which the cognitive organization of language is based directly on experience with language. Rather than being an abstract set of rules or structures that are only indirectly related to experience with language, we see grammar as a network built up from the categorized instances of language use (Bybee, 2006; Hopper, 1987). The basic units of grammar are constructions, which are direct form-meaning pairings that range from the very specific (words or idioms) to the more general (passive construction, ditransitive construction), and from very small units (words with affixes, walked) to clause-level or even discourse-level units (Croft, 2001; Goldberg, 2003, 2006).
Because grammar is based on usage, it contains many details of cooccurrence as well as a record of the probabilities of occurrence and cooccurrence. The evidence for the impact of usage on cognitive organization includes the fact that language users are aware of specific instances of constructions that are conventionalized and the multiple ways in which frequency of use has an impact on structure. The latter include speed of access related to token frequency and resistance to regularization of high-frequency forms (Bybee, 1995, 2001, 2007); it also includes the role of probability in syntactic and lexical processing (Ellis, 2002; Jurafsky, 2003; MacDonald & Christiansen, 2002) and the strong role played by frequency of use in grammaticalization (Bybee, 2003).
A number of recent experimental studies (Saffran, Aslin, & Newport, 1996; Saffran, Johnson, Aslin, & Newport, 1999; Saffran & Wilson, 2003) show that both infants and adults track co-occurrence patterns and statistical regularities in artificial grammars. Such studies indicate that subjects learn patterns even when the utterance corresponds to no meaning or communicative intentions. Thus, it is not surprising that in actual communicative settings, the co-occurrence of words has an impact on cognitive representation. Evidence from multiple sources demonstrates that cognitive changes occur in response to usage and contribute to the shape of grammar. Consider the following three phenomena:
1. Speakers do not choose randomly from among all conceivable combinatorial possibilities when producing utterances. Rather there are conventional ways of expressing certain ideas (Sinclair, 1991). Pawley and Syder (1983) observed that “nativelike selection” in a language requires knowledge of expected speech patterns, rather than mere generative rules. A native English speaker might say I want to marry you, but would not say I want marriage with you or I desire you to become married to me, although these latter utterances do get the point across. Corpus analyses in fact verify that communication largely consists of prefabricated sequences, rather than an “open choice” among all available words (Erman & Warren, 2000). Such patterns could only exist if speakers were registering instances of co-occurring words, and tracking the contexts in which certain patterns are used.
2. Articulatory patterns in speech indicate that as words co-occur in speech, they gradually come to be retrieved as chunks. As one example, Gregory, Raymond, Bell, Fossler-Lussier, & Jurafsky (1999) find that the degree of reduction in speech sounds, such as word-final “flapping” of English [t], correlates with the “mutual information” between successive words (i.e., the probability that two words will occur together in contrast with a chance distribution) (see also Bush, 2001; Jurafsky, Bell, Gregory, & Raymond, 2001). A similar phenomenon happens at the syntactic level, where frequent word combinations become encoded as chunks that influence how we process sentences on-line (Ellis, 2008b; Ellis, Simpson-Vlach, & Maynard, 2008; Kapatsinski & Radicke, 2009; Reali & Christiansen, 2007a, 2007b).
3. Historical changes in language point toward a model in which patterns of co-occurrence must be taken into account. In sum, “items that are used together fuse together” (Bybee, 2002). For example, the English contracted forms (I’m, they’ll) originate from the fusion of co-occurring forms (Krug, 1998). Auxiliaries become bound to their more frequent collocate, namely the preceding pronoun, even though such developments run counter to a traditional, syntactic constituent analysis.
In the usage-based framework, we are interested in emergent generalizations across languages, specific patterns of use as contributors to change and as indicators of linguistic representations, and the cognitive underpinnings of language processing and change. Given these perspectives, the sources of data for usage-based grammar are greatly expanded over that of structuralist or generative grammar: Corpus-based studies of either synchrony or diachrony as well as experimental and modeling studies are considered to produce valid data for our understanding of the cognitive representation of language.
The Development of Grammar out of Language Use
The mechanisms that create grammar over time in languages have been identified as the result of intense study over the last 20 years (Bybee et al., 1994; Heine, Claudi, & H ̈ nnemeyer, 1991; Hopper & Traugott, 2003). In the history of well-documented languages it can be seen that lexical items within constructions can become grammatical items and loosely organized elements within and across clauses come to be more tightly joined. Designated “grammaticalization,” this process is the result of repetition across many speech events, during which sequences of elements come to be automatized as neuromotor routines, which leads to their phonetic reduction and certain changes in meaning (Bybee, 2003; Haiman, 1994). Meaning changes result from the habituation that follows from repetition, as well as from the effects of context. The major contextual effect comes from co-occurring elements and from frequently made inferences that become part of the meaning of the construction.
For example, the recently grammaticalized future expression in English be going to started out as an ordinary expression indicating that the subject is going somewhere to do something. In Shakespeare’s English, the construction had no special properties and occurred in all of the plays of the Bard (850,000 words) only six times. In current English, it is quite frequent, occurring in one small corpus of British English (350,000 words) 744 times. The frequency increase is made possible by changes in function, but repetition is also a factor in the changes that occur. For instance, it loses its sense of movement in space and takes on the meaning of “intention to do something,” which was earlier only inferred. With repetition also comes phonetic fusion and reduction, as the most usual present-day pronunciation of this phrase is (be) gonna. The component parts are no longer easily accessible.
The evidence that the process is essentially the same in all languages comes from a crosslinguistic survey of verbal markers and their diachronic sources in 76 unrelated languages (Bybee et al., 1994). (...)
First and Second Language Acquisition
Usage-based theories of language acquisition (Barlow & Kemmer, 2000) hold that we learn constructions while engaging in communication, through the “interpersonal communicative and cognitive processes that everywhere and always shape language” (Slobin, 1997). They have become increasingly influential in the study of child language acquisition (Goldberg, 2006; Tomasello, 2003). They have turned upside down the traditional generative assumptions of innate language acquisition devices, the continuity hypothesis, and top-down, rule-governed processing, replacing these with data-driven, emergent accounts of linguistic systematicities.
In summary, we have the following: (a) Usage leads to change: High-frequency use of grammatical functors causes their phonological erosion and homonymy. (b) Change affects perception: Phonologically reduced cues are hard to perceive. (c) Perception affects learning: Low-salience cues are difficult to learn, as are homonymous/polysemous constructions because of the low contingency of their form-function association. (d) Learning affects usage: (i) Where language is predominantly learned naturalistically by adults without any form-focus, a typical result is a Basic Variety of interlanguage, low in grammatical complexity but communicatively effective. Because usage leads to change, in cases in which the target language is not available from the mouths of L1 speakers, maximum contact languages learned naturalistically can thus simplify and lose grammatical intricacies. Alternatively, (ii) where there are efforts promoting formal accuracy, the attractor state of the Basic Variety can be escaped by means of dialectic forces, socially recruited, involving the dynamics of learner consciousness, form-focused attention, and explicit learning. Such influences promote language maintenance.
Modeling Usage-Based Acquisition and Change
In the various aspects of language considered here, it is always the case that form, user, and use are inextricably linked. However, such complex interactions are difficult to investigate in vivo. Detailed, dense longitudinal studies of language use and acquisition are rare enough for single individuals over a time course of months. Extending the scope to cover the community of language users, and the timescale to that for language evolution and change, is clearly not feasible. Thus, our corpus studies and psycholinguistic investigations try to sample and focus on times of most change and interactions of most significance. However, there are other ways to investigate how language might emerge and evolve as a CAS (complex adaptive system). A valuable tool featuring strongly in our methodology is mathematical or computational modeling.
Given the paucity of relevant data, one might imagine this to be of only limited use. We contend that this is not the case. Because we believe that many properties of language are emergent, modeling allows one to prove, at least in principle, that specific fundamental mechanisms can combine to produce some observed effect (Holland, 1995, 1998, 2006a, 2006b; Holland et al., 2005). Although this may also be possible through an entirely verbal argument, modeling provides additional quantitative information that can be used to locate and revise shortcomings. For example, a mathematical model constructed by Baxter et al. (2009) within a usage-based theory for new-dialect formation (Trudgill, 2004) was taken in conjunction with empirical data (Gordon et al., 2004) to show that although the model predicted a realistic dialect, its formation time was much longer than that observed. Another example comes from the work of Reali and Christiansen (2009), who demonstrated how the impact of cognitive constraints on sequential learning across many generations of learners could give rise to consistent word order regularities.
Modeling can also be informative about which mechanisms most strongly affect the emergent behavior and which have little consequence. To illustrate, let us examine our view that prior experience is a crucial factor affecting an individual speaker’s linguistic behavior. It is then natural to pursue this idea within an agent-based framework, in which different speakers may exhibit different linguistic behavior and may interact with different members of the community (as happens in reality). Even in simple models of imitation, the probability that a cultural innovation is adopted as a community norm, and the time taken to do so, is very strongly affected by the social network structure (Castellano, Fortunato, & Loreto, 2007, give a good overview of these models and their properties). This formal result thus provides impetus for the collection of high-quality social network data, as their empirical properties appear as yet poorly established. The few cases that have been discussed in the literature—for example, networks of movie co-stars (Watts & Strogatz, 1998), scientific collaborators (Newman, 2001), and sexually-active high school teens (Bearman, Moody, & Stovel, 2004)—do not have a clear relevance to language. We thus envisage a future in which formal modeling and empirical data collection mutually guide one another.
Above all, a usage-based model should provide insight into the frequencies of variants within the speech community. The rules for producing utterances should then be inducted from this information by general mechanisms. This approach contrasts with an approach that has speakers equipped with fixed, preexisting grammars.
Despite these observations, many details of the linguistic interactions remain unconstrained and one can ask whether having a model reproduce observed phenomena proves the specific set of assumptions that went into it. The answer is, of course, negative. However, greater confidence in the assumptions can be gained if a model based on existing data and theories makes new, testable predictions. In the event that a model contains ad hoc rules, one must, to be consistent with the view of language as a CAS, be able to show that these are emergent properties of more fundamental, general processes for which there is independent support.
Characteristics of Language as a Complex Adaptive System
We now highlight seven major characteristics of language as a CAS, which are consistent with studies in language change, language use, language acquisition, and computer modeling of these aspects.
Distributed Control and Collective Emergence
Language exists both in individuals (as idiolect) and in the community of users (as communal language). Language is emergent at these two distinctive but interdependent levels: An idiolect is emergent from an individual’s language use through social interactions with other individuals in the communal language, whereas a communal language is emergent as the result of the interaction of the idiolects. Distinction and connection between these two levels is a common feature in a CAS. Patterns at the collective level (such as bird flocks, fish schools, or economies) cannot be attributed to global coordination among individuals; the global pattern is emergent, resulting from long-term local interactions between individuals. Therefore, we need to identify the level of existence of a particular language phenomenon of interest. For example, language change is a phenomenon observable at the communal level; the mechanisms driving language change, such as production economy and frequency effects that result in phonetic reduction, may not be at work in every individual in the same way or at the same time. Moreover, functional or social mechanisms that lead to innovation in the early stages of language change need not be at work in later stages, as individuals later may acquire the innovation purely due to frequency when the innovation is established as the majority in the communal language. The actual process of language change is complicated and interwoven with a myriad of factors, and computer modeling provides a possible venue to look into the emergent dynamics (see, e.g., Christiansen & Chater, 2008, for further discussion).
(...) Each idiolect is the product of the individual’s unique exposure and experiences of language use (Bybee, 2006). Sociolinguistics studies have revealed the large degree of orderly heterogeneity among idiolects (Weinreich, Labov, & Herzog, 1968), not only in their language use but also in their internal organization and representation (Dabrowska, 1997).
Both communal language and idiolects are in constant change and reorganization. Languages are in constant flux, and language change is ubiquitous (Hopper, 1987). At the individual level, every instance of language use changes an idiolect’s internal organization (Bybee, 2006).
Adaptation Through Amplification and Competition of Factors
Complex adaptive systems generally consist of multiple interacting elements, which may amplify and/or compete with one another’s effects. Structure in complex systems tends to arise via positive feedback, in which certain factors perpetuate themselves, in conjunction with negative feedback, in which some constraint is imposed—for instance, due to limited space or resources (Camazine et al., 2001; Steels, 2006). Likewise in language, all factors interact and feed into one another. (...)
Nonlinearity and Phase Transitions
In complex systems, small quantitative differences in certain parameters often lead to phase transitions (i.e., qualitative differences). (...)
Sensitivity to and Dependence on Network Structure
Network studies of complex systems have shown that real-world networks are not random, as was initially assumed (Barab ́ si, 2002; Barbar ́ si & Albert, 1999; Watts & Strogatz, 1998), and that the internal structure and connectivity of the system can have a profound impact on system dynamics (Newman, 2001; Newman, Barab ́ si, & Watts, 2006). Similarly, linguistic interactions are not via random contacts; they are constrained by social networks. The social structure of language use and interaction has a crucial effect in the process of language change (Milroy, 1980) and language variation (Eckert, 2000), and the social structure of early humans must also have played important roles in language origin and evolution. An understanding of the social network structures that underlie linguistic interaction remains an important goal for the study of language acquisition and change. The investigation of their effects through computer and mathematical modeling is equally important (Baxter
et al., 2009).
Change Is Local
Complexity arises in systems via incremental changes, based on locally available resources, rather than via top-down direction or deliberate movement toward some goal (see, e.g., Dawkins, 1985). Similarly, in a complex systems framework, language is viewed as an extension of numerous domain-general cognitive capacities such as shared attention, imitation, sequential learning, chunking, and categorization (Bybee, 1998b; Ellis, 1996). Language is emergent from ongoing human social interactions, and its structure is fundamentally molded by the preexisting cognitive abilities, processing idiosyncrasies and limitations, and general and specific conceptual circuitry of the human brain. Because this has been true in every generation of language users from its very origin, in some formulations, language is said to be a form of cultural adaptation to the human mind, rather than the result of the brain adapting to process natural language grammar (Christiansen, 1994; Christiansen & Chater, 2008; Deacon, 1997; Schoenemann, 2005). These perspectives have consequences for how language is processed in the brain. Specifically, language will depend heavily on brain areas fundamentally linked to various types of conceptual understanding, the processing of social interactions, and pattern recognition and memory. It also predicts that so-called “language areas” should have more general, prelinguistic processing functions even in modern humans and, further, that the homologous areas of our closest primate relatives should also process information in ways that makes them predictable substrates for incipient language. Further, it predicts that the complexity of communication is to some important extent a function of social complexity. Given that social complexity is, in turn, correlated with brain size across primates, brain size evolution in early humans should give us some general clues about the evolution of language (Schoenemann, 2006). Recognizing language as a CAS allows us to understand change at all levels.
Cognition, consciousness, experience, embodiment, brain, self, human interaction, society, culture, and history are all inextricably intertwined in rich, complex, and dynamic ways in language. Everything is connected. Yet despite this complexity, despite its lack of overt government, instead of anarchy and chaos, there are patterns everywhere. Linguistic patterns are not preordained by God, genes, school curriculum, or other human policy. Instead, they are emergent -- synchronic patterns of linguistic organization at numerous levels (phonology, lexis, syntax, semantics, pragmatics, discourse, genre, etc.), dynamic patterns of usage, diachronic patterns of language change (linguistic cycles of grammaticalization, pidginization, creolization, etc.), ontogenetic developmental patterns in child language acquisition, global geopolitical patterns of language growth and decline, dominance and loss, and so forth. We cannot understand these phenomena unless we understand their interplay. The individual focus articles that follow in this special issue illustrate such interactions across a broad range of language phenomena, and they show how a CAS framework can guide future research and theory.