Download USAGE-BASED THEORY AND EXEMPLAR ... and more Exercises Construction in PDF only on Docsity!
Chapter 4
USAGE-BASED THEORY
AND EXEMPLAR
REPRESENTATIONS OF
CONSTRUCTIONS
joan l. bybee
4.1. Introduction: Usage-based
Theory
The basic premise of Usage-based Theory is that experience with language creates and impacts the cognitive representations for language (Langacker 1987, 2000b; Kemmer and Barlow 2000). Cognitive representations are built up as language users encode utterances and categorize them on the basis of phonetic form, mean- ing, and context. As incoming utterances are sorted and matched by similarity to existing representations, units such as syllable, word, and construction emerge. Thus, grammar can be viewed as the cognitive organization of one’s experience with language (Bybee 2006). It is accepted within this approach that it is wiser to begin the search for the cognitive processes involved in language by first considering domain-general cognitive processes—that is, those known to function in domains other than language, such as vision or neuromotor processing—rather than to assume a priori that language requires special adaptations of cognitive functions (Elman and Bates 1997; Tomasello 2003; Bybee and Beckner 2009). Some of these domain-general processes are categorization, cross-modal association,
50 principles and methods
and neuromotor automation. Consider how these processes apply to language. Categorization applies at every level of form and meaning: phones and their combinations are categorized based on existing representations, as are features of context and meaning (Langacker 2000b; Bybee 2010). Cross-modal associa- tion allows humans to match up the phonetic (or manual) form experienced with properties of the context and meaning. Automation allows the units of language to be combined in conventional ways that lead to fluency in both pro- duction and perception (Bybee 2002). An important characteristic of human language is that the individual units and sequences of units are subject to high levels of repetition. It is repetition that leads to conventionalization of categories and associations, as well as to the automation of sequences. Because some units and sequences are repeated more than others, it has been possible to identify the properties of cognitive representations that depend upon the extent to which they have been accessed for production or perception. Thus, within Usage-Based Theory the study of frequency effects of various sorts has contributed to the understanding of the nature of grammatical organization (Bybee 2007). It is recognized that languages are constantly changing, and this change is gradual and takes place as language is used. Change is attributed to the way partic- ular cognitive processes apply in language use, thus change provides an important window into the understanding of the cognitive processes underlying language. As change is gradual, the categories and units of language are variable and they form gradient rather than strictly bounded categories. Thus, linguistic structure is viewed as emergent—governed by certain regular processes, but always chang- ing as it is re-created in the individual and in specific usage situations (Hopper 1987; Kemmer and Barlow 2000; Ellis and Larsen-Freeman 2006b). Thus, rather than a fixed, static set of representations, language is viewed as being affected by experience in an ongoing way even in adults. It also follows that we should not expect linguistic constructs such as segment, syllable, morpheme, word, or construction to have strict definitions, nor do we expect all the manifestations of these constructs in languages to exhibit exactly the same behavior (Bybee 2010: chapter 1). As the term implies, the object of study in Usage-based Theory is not only the native user’s competence, which resides in the cognitive representation, but all the perception and production processes that are brought to the task of using language (Kemmer and Barlow 2000). Thus, the data considered applicable to formulating and testing hypotheses can be very broadly drawn from experiments, child language acquisition, language change, and large corpora representing natural usage. The lat- ter source of data (now much more available than in the past) partially supplants the use of native speaker intuitions. While such intuitions are interesting and important, research with natural usage indicates that language users are often unaware of the nature and frequency of certain structures that they use. For this reason, it is impor-
52 principles and methods
4.3. Exemplar Representation for
Constructions
Despite the more surface-oriented view of Construction Grammar in comparison to generative grammar in its various manifestations over the decades, abstraction away from particular tokens does take place, as evidenced by innovative uses of constructions. One important question addressed in recent literature is the nature of this abstraction and how the experienced tokens of a construction contribute to the formation of a cognitive representation of that construction (Goldberg 2006a; Bybee 2010). This chapter discusses these questions taking an exemplar view of constructions. The nature of exemplar representations will be explained in more detail in the next section, but for present purposes we can say that exemplar models propose that memory for linguistic experience is like memory for other types of experience: each token of experienced linguistic behavior has an impact on cognitive represen- tation; when stored representations are accessed in either encoding or decoding, the representations themselves change. In addition, memory storage for linguis- tic experience includes detailed information about the tokens that have been pro- cessed, including their form and the contexts in which they were used. Exemplar representations contrast with the more abstract representations of structural or generative theories (at all levels—phonetic, morphosyntactic, and semantic/prag- matic), in that variation and features predictable from general principles have not been removed. In such a model, the general categories and units of grammar can emerge from the experience that is recorded in memory because exemplars are cat- egorized by similarity to one another and because contiguous experiences—such as meaning and acoustic shape—are recorded as linked to one another. The following sections will describe the general properties of exemplar models and discuss how they can be applied to constructions. The following arguments for choosing exemplar representations for constructions will be presented in this chapter:
- As Fillmore et al. (1988) argue, many constructions have idiosyncratic features of morphosyntax, semantics, pragmatics, and phonology, and much of what a speaker/hearer knows about his/her language is not predictable from the very general rules that have occupied the attention of most syntacticians in the past, but rather consists of specific information that must be associated with specific constructions. Specific information finds a natural expression in an exemplar model, where the storage and categorization of all detail both predictable and idiosyncratic is considered to be a basic response to linguistic input and applies to all constructions, specific or general (see sections 4.4 and 4.5).
- The stored representations in the form of exemplars respond to usage by allowing the representation of both token and type frequency; these
usage-based theory 53
frequency patterns are important for understanding the categories that are formed for the schematic slots in constructions (see section 4.6).
- Specific instances of constructions develop into new constructions, thus specific exemplars of constructions need to have cognitive representation (see section 4.7).
- Exemplar models allow for specific meaning from the context of use to impact cognitive representation, which then accounts for the way that words and constructions are affected by the meaning that occurs in the context. The impact of context is seen both in the development of special meanings and implications of constructions and also in developments such as the negative prosody identified in corpora for constructions with cause (see section 4.8).
- Exemplar models were not developed especially for language; rather they apply equally to linguistic and nonlinguistic categories. Such models then take the usage-based view that language is a part of general cognition and allow us to access explanations for linguistic phenomena outside of language.
4.4. Exemplars of Construction:
Constructions as Chunks
Exemplars are categories formed from tokens of experience that are judged to be the same (Pierrehumbert 2001). Linguistic exemplars come in a variety of sizes, ranging from a single segment, such as a vowel, to whole paragraphs, such as the Pledge of Allegiance. The exemplars themselves are grouped together by similar- ity. Thus, the vowels of hit, swim, and sip may be grouped together, the different phonetic realizations of a word, such as pretty will be grouped together, as well as exemplars for longer sequences, such as all of a sudden. These exemplar clouds, as they are called, constitute categories. Exemplar categories are structured by simi- larity and frequency (Nosofsky 1988) and often exhibit prototype effects. Because they are grouped together based on similarity, in theory any type of category could be represented in exemplars. However, it is common to find prototype effects emerging from categories because of differences in degrees of similarity. In addition, exemplars may differ in strength depending upon the number of tokens that comprise them. That is, exemplars built up from a large number of tokens will be represented more strongly than those built up from a smaller num- ber of tokens. The stronger exemplar or set of exemplars often forms the center of a category and other exemplars are more or less similar to the stronger exemplar or set of exemplars (Pierrehumbert 2001). As exemplars are based on perceptual stimuli, and exemplars are grouped together based on similarity, we distinguish exemplar categories formed by
usage-based theory 55
component words can lose their association with exemplars of the etymologically same word (Bybee 2003; Beckner and Bybee 2009; Bybee 2010). In grammaticaliza- tion this process is referred to as decategorialization (Hopper 1991). For instance, in the phrase in spite of, the erstwhile noun, spite, has lost its ability to function as a noun, as in this phrase it is not modifiable. It has also lost its earlier meaning of ‘in defiance’ and adds nothing to the general meaning of concessive that the phrase now has. This loss has taken place gradually, which means that there are degrees of linking between words in different contexts (Beckner and Bybee 2009). Note that complex prepositions such as in spite of are not extremely frequent. As Hoffmann (2004, 2005) has pointed out, loss of analyzability can come about without high lev- els of frequency, encouraged by changes in distribution, meaning, and function. Another aspect of chunking that is important for the understanding of how constructions emerge is that both nonlinguistic and linguistic chunks tend to have meaning assigned to them at the highest level possible (Ellis 1996; Bybee 2010). For instance, one might witness the repetition of a sequence of events such as a human throwing a ball, a dog running after it, catching it, and returning to the human, setting the ball at the human’s feet. The human picks up the ball and throws it again. It is a human propensity to try to understand such sequences as chunks of behavior and to assign labels, such as ‘playing fetch’ to such chunks. Similarly, sequences of linguistic units that occur together repeatedly tend to be assigned meanings as a whole rather than simply as a sum of the parts, as can be seen in many expressions such as in spite of, here and there, or all of a sudden, which no longer are fully compositional. It is this chunking and labeling phenomenon that provides language with constructions.
4.5. Details in Exemplar Representation
4.5.1 Details of Semantics and Pragmatics
The original motivation for proposing constructions as part of grammar, as expressed in Fillmore et al. (1988), was that constructions have idiosyncratic structure and meaning. As mentioned above, this attribute is taken by some to be criterial in defining constructions, though in the usage-based view con- struction can be defined simply as frequently used and thus conventional word sequences. As mentioned above, details of form and usage are automatically reg- istered in exemplar representation and through entrenchment can become an inherent part of the construction. In this section, based on Bybee (2006), I dem- onstrate that an exemplar model is essential to explaining how constructions acquire idiosyncratic semantic/pragmatic and phonetic properties. I also argue that redundant detail may turn out to be important in forming the bases of new uses of constructions.
56 principles and methods
In discussing the What’s X doing Y? (WXDY) construction, Fillmore and Kay (1999) note that one reason for regarding this as a construction is its special prag- matic implications, which in some cases occur without any accompanying literal meaning, as in (1).
(1) What’s that box doing up there?
As for the origins of this sense for the construction, they say: While the WXDY construction may have had its origin in conversational implicatures—through situations in which an individual A is clearly up to no good and B asks what A is doing—the semantics of incongruity is now conventionally associated with the special morphosyntax of WXDY constructs. (Fillmore and Kay 1995: 5; emphasis original)
The conventionalization of implicature (or from the hearer’s point of view, infer- ence) is also well-known from grammaticalization research (Bybee 1988b; Traugott 1989; Bybee et al. 1994; Traugott and Dasher 2002). It is thought that the frequent co-occurrence of an inference with a particular construction can lead to that infer- ence being taken as part of the meaning of the construction. The originally infer- ential meaning can even replace the earlier meaning. If we consider how this conventionalization occurs, we see that we need an exemplar model to account for it. In a model in which semantic representations contain only a core or abstract meaning and inferences are calculated on the fly in each context, there is no way for an implication/inference to stick to a construction. However, an exemplar model would record the inferences made in each instance of use, and if the same inference is made on multiple occasions, the strength of that inference increases. With sufficient strength among the semantic exemplars for a construction, the inference can become conventionalized as part of the meaning of the construction. This can happen over diachronic time and it can also happen over the course of language acquisition, as the learner establishes what the implica- tions of acquired constructions are. Note that the speaker/hearer cannot wait until s/he has experienced the inference more than once to store it in memory, because there is no way to know if the inference has been experienced before if it is not reg- istered in memory from the very first experience (Bybee 2006).
4.5.2. Phonetic Detail
Similarly, phonetic change can accrue to the words of a construction through instances of use in reducing contexts. Grammaticalizing constructions undergo reduction as frequency of use increases. Words in grammaticalizing constructions often have a wide range of variation, as with the phonetic variants of English going to, which range from the full form to variants spelled gonna to extreme reduction as in I’m gonna [aimn] The entrenchment of the reduction in the construction helps to distinguish one construction from another, as in the case of used to as
58 principles and methods
For instance, it is often found that the set of lexical items that can occupy a slot in a construction may be constituted of two or more clusters of closely related items (Goldberg 1995; Israel 1996). An example is the well-known way construction which is used with verbs that indicate either manner of motion (4) or means (5) (see Jackendoff 1990).
(4) Romana Kryzanowska weaves her way through Drago’s Gym in midtown Manhattan like a mother hen... (Time Magazine Corpus, 2004) (5) His weakness, which Holmes unflinchingly describes, was an inability to resist the financial and sexual rewards that came along as he clawed his way to the top. (Time Magazine Corpus, 2000)
However, one also finds within these groups more specific clusters of verbs such as those indicating winding motion (pick, thread, wind, wend, worm, weave, etc.) or laborious motion (plod, crawl, grind, slog, stumble, etc.) and so on (Goldberg 1995; Israel 1996). Such clusters of highly related items are the results of item-based anal- ogy (Skousen 1989; Eddington 2000; Bybee 2010: chapter 4). In other words, rather than making reference to a general semantic feature when using a construction, the speaker may very well reference a particular lexical item that has already been used in the construction and stored in memory. Exemplar-based representations of constructions will include a list of all the words experienced in a certain slot in a construction. This list of words (organized in clusters by similarity of mean- ing) is the basis for new extensions of the construction. Barðdal (2011a) shows that the majority of verbs borrowed into Icelandic take the case construction of syn- onymous verbs. Boas (2003) has argued for resultative constructions in English that many micro-constructions are available for reference when extending existing constructions to new lexical items. Item-based extension leads to a family resemblance structure among the lexical items that constitute a schematic category in a construction. For instance, the set of adjectivals (including prepositional phrases) that occur in the Spanish ‘become’ construction using the verb quedarse can be analyzed into several clusters based on semantic similarity (Bybee and Eddington 2006). Some of these clusters are large enough that they include adjectivals that have no semantic relation to one another, but rather share properties with more central members in family resem- blance fashion. Thus, our corpus study turned up quedarse with the adjective qui- eto, which means ‘still.’ It is, of course, related to inmóvil ‘immobile’ but also to an adjective with very different implications, tranquilo ‘tranquil, peaceful.’ This could in turn be related closely to adjectives expressing the more positive side of ‘tran- quility, satisfaction’ with conforme ‘satisfied’ and this in turn to a gusto ‘pleased’ as shown in (6).
(6) inmóvil quieto tranquilo conforme a gusto ‘motionless’ ‘still’ ‘tranquil’ ‘satisfied’ ‘pleased’
As with other family resemblance chains, the ends of the chain may have very little to do with one another; inmóvil ‘motionless’ and a gusto ‘pleased’ share few features.
usage-based theory 59
But they both share some features with tranquilo ‘tranquil.’ While tranquilo might have originally been used in this construction because of its similarity to quieto, it brought along with it the ‘peacefulness’ sense, which then could be referenced for further extensions. This type of structure, then, points to local extensions by item-based analogy, rather than global summations of abstract features. In order for item-based analogy or local similarity to apply, the cognitive representation of this construction must contain the specific adjectivals that have been experienced in it (see also Bybee 2010: chapter 5). Local extensions may also be based on form, showing that lexical slots in con- structions are not purely semantic, but contain information registered in memory for the form of items. The history of the Spanish ‘become’ construction with que- darse, as reported in Wilson (2009), shows that it first appeared with the adjec- tive solo ‘alone’ in the thirteenth century, as well as with a few other adjectivals. The ‘alone’ sense gave rise in later centuries to uses with the adjectives huérfano ‘orphaned’ and viudo ‘widowed,’ as well as some phrases such as sin heredero ‘with- out an heir’ and sin padre ‘without a father.’ By the fifteenth century, there are many examples with sin ‘without’ but they extend well beyond the original semantics related to solo ‘alone.’ They include phrases such as sin armas ‘without weapons,’ sin pluma ‘without a pen,’ sin deuda ‘without debt,’ sin pena ‘without grief,’ and sin quexa ‘without complaint.’ The semantics of these phrases do not fit with the ‘alone’ semantics, nor do these phrases share abstract semantic features. All they have in common is the use of the preposition sin ‘without.’ This pattern suggests that features of form may also influence the choice of items occurring in a lexical slot in a construction. An exemplar representation would include these properties of the form, as all tokens and all their properties potentially have an impact on the representation. In contrast, if only semantic features are recorded in the construc- tion’s representation, there would be no way to account for this extension pattern.
4.6. Token and Type Frequency
4.6.1. Token Frequency
Another important argument for an exemplar model of linguistic representations is that exemplars, by their very nature, provide a record of the frequency of occur- rence of tokens in linguistic experience. As mentioned above, each token of expe- rience has an impact on the memory for linguistic items (Nader et al. 2000). In Bybee (1985), I proposed that representations are strengthened by repetition, mak- ing them easier to access. This proposal applies primarily to token frequency—the number of times a particular string occurs in a text or corpus. Thus, every time a construction is used, the constant parts of it are strengthened. But construc- tions also have schematic slots which will be represented by different exemplars
usage-based theory 61
What do these patterns of occurrence mean for cognitive representation? If each token of experience has an effect on representation, then each occurrence of the construction maps onto the exemplar cloud for the construction. The figure in (10) schematizes this exemplar cloud. Note that the particular items in the schema, such as her or mad are represented by clouds of exemplars themselves.
her wild him mad
(10) SUBJECT [DRIVE] me crazy
[send] you up the wall [make] them nuts NP batty
.. ..
In (10) larger font is used to represent items that have a higher token frequency in the construction, that is, DRIVE, me, and crazy. We have shown in Bybee and Eddington (2006) that items with higher token frequency within the construction serve as the central members of the categories that form for schematic slots within the construction. That is, drive is the central verb for the construction and crazy (for American English) is the central adjective. Extensions of the construction will be based on these central members. In addition, these items of higher frequency will also be easier to access, which will increase their frequency even more. As mentioned in section 4.4, constructions are formed by the domain-gen- eral process of chunking, by which repeated sequences of experience come to be remembered and processed as units. These processes operate gradually in language use, so that the creation of a construction as a unit does not mean that analyzability is lost immediately when constructions are formed. However, extensive repetition can lead to autonomy, as we will see in the next section. For the moment, note that a frequent instance of the construction, it drives me crazy, while it maps onto the schema in (10), also has its own sequential relations which may be stronger than other instances of the construction, such as Lindsay nearly drove him wild with anxiety.
4.6.2. Type Frequency
Type frequency is measured in the context of a construction and counts how many different items occur in the schematic slots of constructions. Thus, we might count the number of adjectives that occur in the drive X crazy construction in a corpus, as Boas did for the BNC, and find that there were eight (provided that nuts and crackers in this construction are considered adjectives). Counting the prepositional phrases would increase the type frequency of this slot. Constructional slots all have their own type frequency: the verb slot in (10) has a very low type frequency while
62 principles and methods
the subject slot is virtually unlimited within the domain of noun phrases and thus has a very high type frequency. The importance of representing type frequency in the representation of a construction is that type frequency relates directly to productivity. In general, the higher the type frequency of a construction, the more likely it is to occur with a novel item (Bybee 1985; Baayen 1993; Barðdal 2008, 2011a). The effect of type fre- quency on productivity is constrained by two factors. First, items with very high token frequency may have formed a more autonomous chunk and do not activate the construction’s exemplar cluster; thus such items will not contribute to produc- tivity (Bybee 1985). Second, the semantic (and in some cases formal) properties of the schematic slot may restrict productivity. Thus, the complement to drive in (10) may include adjectives synonymous with crazy or phrases such as over the edge, to distraction, to suicide, but we do not find uses such as drive X to pneumonia or drive X to fear. In terms of formal properties, there is a construction have the noun to, where many of the nouns end in the suffix -ity or -ability, such as have the capa- bility to. This perhaps accounts for the extension have the availability to (which seems only marginally acceptable to me, but was written on a sheet provided by an optometrist). The semantic or formal properties of a slot determine its schematicity. A highly schematic slot, such as the object position in (10) has a broad set of properties, such as ‘animate noun.’ A less schematic slot might have a narrower definition, such as ‘synonymous with crazy.’ Schematicity and productivity are independent of one another: the ‘crazy’ slot in (10) is fairly productive within this semantic domain. Thus, drove me bananas, bonkers, berserk are all possible extensions of this slot. Barðdal (2008) notes the inverse correlation between type frequency and seman- tic coherence, in that both properties encourage productivity, even though a very high type frequency often correlates with high schematicity and low schematicity (greater coherence) also gives productivity in the domain delimited. Despite the trade-off, the most productive categories have both high schematicity and high type frequency (e.g., the English regular verb formation with the suffix -ed). Both pro- ductivity and schematicity can be determined directly from an exemplar represen- tation where all the types that have occurred in the construction are represented. At least two factors contribute to the relation between high type frequency and productivity. First, the greater the number of types in a construction, the more bases there are for the item-based analogy that creates novel instances of the con- struction. Second, as Baayen (1993) points out, the processing of low frequency instances of constructions requires parsing or activation of the representation of the construction, which strengthens the construction. Processing high frequency instances of constructions, as we have already seen, can occur without activating the construction and therefore does not strengthen the constructions. Another approach to frequency in constructions is Collostructional Analysis (Stefanowitsch and Gries 2003; Stefanowitsch, this volume). This approach rolls token and type frequency into one measure along with controls for overall fre- quency of the lexical item and a count of constructions in the language. In this
64 principles and methods
just one instance of a general construction for expressing movement in space for a purpose, but over the last few centuries it has lost its association with the purpose construction, as well as with the movement verb go. It has become an autonomous construction used for a distinct function, as in the example A deal this big is going to carry a price tag (Time Magazine Corpus, 2000), where the speaker is making a prediction and no movement in space is expressed. In addition, a grammaticaliz- ing construction increases in type frequency as it comes to be used with more and more lexical items (Bybee 2003; Himmelmann 2004), as for instance, when the be going to construction comes to be used with inanimate subjects and stative verbs (see also Fried, this volume). Thus, an exemplar model provides the representational basis upon which a particular instance of a construction can be established as a new construction and can become autonomous from its source.
4.8. Exemplar Semantics and
Pragmatics for Constructions
The rich memory representation that is characteristic of exemplar models is com- patible with the ideas expressed in cognitive linguistics that linguistic meaning is highly dependent upon and not separable from encyclopedic knowledge, or knowl- edge of one’s world (Fillmore 1982, 1985b; Langacker 1987; Lakoff 1987; Croft and Cruse 2004). The representation of encyclopedic knowledge along with linguistic knowledge would emerge naturally from an exemplar model in which tokens of words or constructions are represented in memory along with the situations they have been associated with in experience. Given the vast array of detail that memory can record, a pressing theoretical issue is how semantic categories are abstracted from the representation of experi- ence and how abstract these categories really are. Note that this is not a question created in any way by exemplar theory. It is a question we must answer in any theory unless we want to postulate that linguistic categories are innately speci- fied, for any theory must eventually address the nature of the abstraction process that takes place in language development in the child and the resulting relation- ship between exemplars derived from experience and abstract categories (see, for instance, Abbot-Smith and Tomasello 2006). This question has been addressed for lexical meaning through research in pro- totype categorization, which has shown that the natural categories that humans construct are not based on highly abstract necessary and sufficient conditions, but rather consist of members that are more or less similar to a central member or prototype. The question concerning such categories is how the features on which similarity is based are chosen. Rosch and Mervis argue that prototype “categories form to maximize the information-rich clusters of attributes in the environment”
usage-based theory 65
(Rosch and Mervis 1975a: 458). To this statement one might add, “as the environ- ment is experienced by a human agent.” Grammatical meaning, as more abstract, is much less dependent upon ency- clopedic knowledge, such as that represented in semantic frames (Fillmore 1982, 1985 b). Grammatical meaning seems to make reference to the speech situation itself (as with deictic elements) or to internal relations within the utterance, by means of constructions. Yet, the meaning of constructions and grammatical mor- phemes must also be built up through specific tokens of use, so the question of the relation between the specific and the abstract still arises. Just as we must be able to identify attributes residing in different entities in the environment, and just as we are able to find similarities in the form of linguistic elements, so we are also able to find attributes that are shared by situations that we encounter in association with particular constructions. When shared attributes are identified, their repre- sentations are strengthened while the nonshared properties of situations are not reinforced. Thus, abstraction occurs because only certain features are associated with a construction. The extent and nature of abstraction for grammatical meaning is still some- what controversial, as it has been traditional to attempt to identify only one abstract feature for each grammatical morpheme (Jakobson 1971 and many others). Such analyses often run into difficulty, as we will see below. As exemplar models do not insist on the reduction of a category to only one feature, and as the research on natural categorization turns up prototype effects, it is important to examine the question of just how abstract grammatical meaning is. This question can be addressed through the natural experiment of language change, especially when viewed as ongoing. In the following I review what language change can tell us about grammatical meaning and argue that exemplar representation is necessary for our understanding of how these changes occur. Just as recent research on idioms, prefabs, and other types of formulaic lan- guage (Nunberg et al. 1994; Wray 2002a; Corrigan et al. 2009) has emphasized the diversity and specificity of constructions and collocations in language use, so studies of grammaticalization have turned up many cases of uses of grammatical morphemes that are not as abstract as we were led to expect. Under the heading of ‘retention’ (Bybee and Pagliuca 1987) or ‘persistence’ (Hopper 1991), we find examples in which the meaning or distribution of gram- matical morphemes reflects their earlier patterns of use and meaning. For instance, while there is no question that English will expresses the abstract meaning of future, as in example (11), there are contexts in which a meaning of ‘willingness’ comes through as in (12) and (13), reflecting the earlier meaning of will as ‘desire.’ The second will in (13) can be interpreted as either prediction or willingness (negated to ‘refusal’).
- Off of public lands in Nevada alone over the next five years, more than $ billion worth of gold is going to be removed. And you and I will not get a penny for that. (COCA, 1990)
usage-based theory 67
The semantic prosodies found in corpus work (Sinclair 1996b) also demon- strate that exemplars of words or constructions are associated with contexts that can at once be very specific and cover long stretches of text. Smith and Nordquist (2012) argue that the association of negative affect with the cause to construction came about over time as the percentage of negative collocates increased. In Early Modern English, cause was as likely to occur with positive or neutral collocates as with negative ones, though the neutral collocates greatly outnumbered the positive ones, with the result that cause was associated with negative or neutral collocates. Smith and Nordquist (2012) find that in Present Day English (PDE), the negative collocates have come to dominate, representing 81% of the tokens they extracted from a PDE corpus. Note, however, that it is not the lexical item cause itself that has absorbed the negative affect, but rather the particular construction it occurs in. In the transitive construction, cause + NP (as in cause an accident, cause damage), a negative collocate is found in 90% of the cases, while in the infinitival construction (as in cause the boat to sink), the collocate is negative 60% of the time. In both cases, the other uses are neutral rather than positive. As might be expected, the general meaning of the construction is highly affected by the lexical types that occur in it, and it is the constructional unit that affects the meaning of its component parts. These examples demonstrate that even though grammatical and construc- tional meaning can be very abstract in certain contexts, it is also the case that use in specific contexts can affect the meaning of grammatical and lexical morphemes, sometimes providing them with more specific meaning not found in all contexts. These specific meanings can be added in by inference, as we discussed in section 4.7, they can be retained from earlier uses in context, they can be absorbed from the overarching meaning of the construction they occur in, or they can be added by association with certain lexical items. All of these scenarios require cognitive rep- resentations in which a great deal of context is associated with individual tokens, which determines how they are mapped onto exemplars.
4.9. The Limits of Exemplar Models
An exemplar model is not, on its own, a theory of language; it is a model of memory representation. It addresses only the mechanism of memory storage and the result- ing structures or categories and it says nothing about the content of linguistic cat- egories, such as why languages tend to have past or future markers or categories such as ‘subject.’ Exemplar models speak to the degree of abstractness of mem- ory representations, positing a rich, detailed memory for experience that spans many domains—linguistic form (phonetics), context, affect, and meaning. Much of the complexity needed for a complete model of language is not provided by the exemplar models that have developed for other domains. The levels of abstraction for constructions argued for in Croft (2001, 2003a) and Barðdal (2008, 2011a), for
68 principles and methods
instance, are not given by exemplar theory, but must be discovered empirically for linguistic constructions. Also, while exemplar models would allow exemplar clouds containing a word or other element that occurs in many different contexts, the factors that determine whether a word is analyzable in a certain context, such as spite in in spite of—that is, semantic transparency and frequency of use—are outside of exemplar theory. At the same time one might argue that exemplar models are too unconstrained, as they allow any type of category to form and any property to provide the basis of a similarity matching. I do not consider this a problem, however, because it allows the substance of linguistic experience (as filtered through human cognition and attention) to be the prime determinant of the nature of the categories. Thus, it leaves many linguistic questions to be addressed empirically, through the compari- son of languages and the study of language change.
4.10. Conclusion
Usage-based theory covers a wide range of research questions arising from the premise that use of language figures critically in determining the nature of cognitive representations of language, or put another way, usage events create linguistic structure. The premise of usage-based grammar leads directly to con- structions with their direct association of meaning with form as the basic units of grammar. Exemplar representation is central to a usage-based approach to language, since it reflects the way that linguistic structure emerges when human cognition deals with tokens of experience with the world. I have argued here that exemplar representation helps us explain many of the dynamic properties of constructions, such as how they arise from other constructions and how they change over time. A further important advantage of exemplar models is their ability to register token and type frequency for constructions. Frequency effects raise many interest- ing questions that are still under investigation, such as questions concerning the interaction of token frequency with type frequency, the gradual development of autonomy, the interaction of schematicity and type frequency in determining pro- ductivity, and the question of the effects of different levels of token frequency. Two properties of the usage-based approach are worth emphasizing. First, rep- resentations are dynamic and change with usage events, not just across generations but within the individual as usage patterns change. The ability to represent change provides investigators with explanations for how and why constructions develop. Second, the processes we have examined here that create constructions and there- fore grammar are domain-general. The processes that create constructions— cross-modal association (for linking sound and meaning or context), chunking, and categorization—are all processes that apply in other cognitive domains, such
