The Competition Model: Implications for Language Processing, Language Development and Language Breakdown! Elizabeth Bates*, Beverly Wulfeck**, Arturo Hernandez*, Elena Andonova*** * University of California at San Diego, USA. **San Diego State University, USA. *** New Bulgarian University BASIC TENETS OF THE MODEL ‘The Competition Model is a framework for the crosslinguistic study of language use. Itis designed to capture facts about the com- prehension, production, and acquisition of language by real human beings, across a variety of qualitatively and quantitatively distinct language types. Our own work on the Competition Mode! (CM) has been il- juminated by the insights of a particular class of linguistic theories alternatively called "functional grammar” and/or “cognitive grammar” (more on this point below). How- ever, the goals of linguistic and psycholin- This paper is based on extracts from the follow- ing papers: Bates, E. & MacWhinney, B. (1989) Fune- tionalism and the Competition Model. In: MacWhinney, B. & Bates, E. (eds.) The Crosslinguistic Study of Sen- tence Processing, Cambridge: Cambridge University Press (reprinted with the permission of Cambridge Uni- versity Press); Von Berger, E. Wulfeck, B., Bates, E. & Fink, N. (1996) Developmental Changes in Real-Time Sentence Processing, In: First Language, vol. 16, Part 2 June 1996) (reprinted with the permission of Alpha Ac- ademic); Hernandez, A., Bates, E., & Avila, L. (1994) ‘On-line Sentence Interpretation in Spanish-English Bi- linguals: What does it Mean to be "In Between”? In: Applied Psycholinguistics 15 (1994), 417-446 (reprinted with the permission of Cambridge University Press); Bates, E., Wulfeck, B,, & MacWhinney, B. (1991) Cross- Linguistic Research in Aphasia: An Overview. In: Brain and Language, 41, 123-148 (reprinted with the permis- sion of Academic Press); Bates, E., McDonald, J., MacWhinney, B. & Appelbaum, M. (1991) A Maximum Likelihood Procedure for the Analysis of Group and Individual Data in Aphasia Research. In: Brain and Language 40, 231-265 (reprinted with the permission of Academic Press) guistic research are often different, revolv- ing around the now-classic distinction be- tween competence and performance. Com- petence refers to the abstract knowldege of the language possessed by an ideal speak- er-listener, removed from the constraints and inconveniences of real-time language use. Performance refers to the actual process of language use by real people in real situa- tions. The Competition Model represents an effort to account for facts about performance across languages that vary dramatically in their structural and functional organization. There are three basic tenets of the Com- petition Model that we need to examine in some detail before reviewing results ob- tained to date within this framework: cross- linguistic variation, functionalism/ interac- tionism, and the probabilistic nature of lin- guistic knowledge. As noted above, the CM has grown out of a tradition in linguistics and psycholinguistics called functionalism. Linguistic functionalism can be defined as the belief that ”the forms of natural languag- es are created, governed, constrained, ac- quired and used in the service of communi- cative functions” (Bates & MacWhinney, 1982; MacWhinney, Bates, & Kliegl, 1984). Within linguistics and psycholinguistics, *functionalism” is clearly opposed to the principle of syntactic autonomy proposed within Chomskian generative grammar. Chomsky (1957; 1975) argues forcefully against attempts to relate sentence structure to aspects of the communicative environ-ment or to more general (non-linguistic) properties of the human mind. In the Chom- skian vision, grammar can be viewed as a kind of mathematical object, an abstract sys- tem that is unconstrained and unshaped by communicative purpose or function. Al- though the Competition Model does not pos- it the simple sorts of relations between communicative function and language pro- posed by Skinner (1957), neither does it ac- cept the total divorce of language from com- municative function embodied in Chomski- an linguistics. Instead, it considers the rela- tion between language form and language function as the major empirical phenome- non to be characterized by psycholinguistic theory. In terms of cross-linguistic variation, grammars can be viewed as a class of solu- tions to the problem of mapping hyper-di- mensional meanings onto a highly con- strained, low-dimensional medium whose only devices are word order, lexical mark- ing, and suprasegmentals, The universal and culture-specific contents of cognition inter- act with universal constraints on human in- formation processing, creating a complex multivectorial problem space (Bates & MacWhinney, 1982; Karmiloff-Smith, 1984). A central claim in the Competition Model is that forms ofterr express a variety of correlated functions. This "peaceful co. existence” of functions helps both the lis- tener and the hearer by allowing a small set of structures and markers to serve a myriad of related purposes. This then serves to build ina certain adaptive instability into language, since any particular coalition of meanings may eventually collapse and require rein- terpretation. Grammars are thus viewed as a set of partial solutions to the universal form-meaning mapping problem, each rep- resenting one pathway through the con- straints imposed by cognitive content and cognitive processing. They represent multi- ple solutions to this universal problem ex- hibiting differences of two kinds: (1) quali- tative differences (what concepts are cod- ed, i.e. meaning, and what coding options are selected, i.e. form); and (2) quantitative differences (in the strength of form-function mappings and the strength of form-form mappings). Our understanding of the structure of the language processing system can be expressed in terms of the following key concepts: two- level structure, direct mapping between these levels, cue validity, cue strength, cue cost. To begin with the first one, only two lev- els of informational structure are specified a priori in this model: a functional level (where all the meanings and intentions to be expressed in an utterance are represent- ed) and a formal level (where all the sur- face forms or expressive devices available in the language are represented). We assume that the mappings between these two levels are as direct as possible. Intervening layers will emerge only when they are essential for processing to take place in real time. In our work with Italian, Hungarian, and Dutch, we have encountered a variety of compound cues to sentence interpretation that seem to defy a linearly separable representation. For example, a sentence like "Il gatto spinge il cane” ("The cat pushes the dog”) is usually interpreted as an SVO (Subject-Verb-Ob- ject) when it appears with neutral stre: however, stress on either of the two nouns (e.g. "THE CAT pushes the dog” or ”The cat pushes THE DOG”) greatly increases the probability that the same sentence will be interpreted as an OVS (Object-Verb- Subject). A completely different set of stress/ order regularities hold for legal but prag- matically marked non-canonical orders like NNV (’The cat the dog pushes”) and VNN ("Pushes the cat the dog”). This means that word order and stress cues in Italian are never evaluated separately; instead, parsing requires the joint interpretation of word or- der and prosody in complex configurations. To handle compound cues like this, the ide- alized two-layer approach has to be modi-fied to permit the emergence of intervening layers. There are a variety of ways in which we can complicate a simple two-layer model in order to represent such compound cues. Of these, the most attractive is the "hidden unit” model of Rumelhart, Hinton, and Williams (1986). The acknowledgement that interven- ing layers can emerge through learning is a new feature of the model (cf. Bates & MacWhinney, 1979, 1982, 1987; MacWhin- ney, 1978, 1982). In this regard, it is inter- esting that the above order/stress configura- tions are acquired late by Italian children (Bates et al., 1984). They begin with a lin- car interpretation of each variable (e.g., treating stress as a cue to the agent role re- gardless of word order environment), and reach the above nonlinear combinations sometime after the age of five. In the Competition Model, the mapping betiveen Tori and function is sated as dec ly as possible. Direct mapping, however, does “not mean that relationships between form and function are necessarily one to one; indeed, we assume that one-to-one mappings are rare in natural languages, which are instead com- posed primarily of many-to-many relationships (polysemy, homophony, syncretism). Second- ly, the principle of direct mapping is proposed as an alternative to certain forms of modular- ity. Modularity postulates the computational independence of data sources and computa tional differences between the various mod- ules that have evolved to deal with each data source. By contrast, the principle of direct mapping emphasizes (1) the mixed nature of input to the language processor, and (2) the homogeneity of processing across different data types. By mixed data types we mean that the language processor can make use of com- pound cues that cross traditional boundaries (eg., segmental phonology, suprasegmental phonology, morphology, the lexicon, and po- sitional frames). The example of word order/ stress configurations in Italian we used earlier illustrates this claim. By homogeneous pro- cessing we mean that different sources of in- formation (morphological, phonological, lex- ical, and syntactic) are processed in a similar fashion, via a common set of recognition and retrieval mechanisms. Although the constraints imposed by the mapping problem are heavy, and the class of possible solutions is finite, the number and range of language types that are possi- ble seems to be quite extensive. Languages differ qualitatively, in the presence or ab- sence of certain linguistic devices (e.g., word order constraints, case marking), but they also differ quantitatively, in the extent to which the ”same” linguistic device is used at all and in the range of functional roles that the ”same” linguistic device has come to serve. Thus, in Italian word order can be varied extensively for pragmatic purposes. This is in contrast with English where word order is a strict and highly valid cue to sen- tence interpretation and where it is usually Subject-Verb-Object (SVO). Consider, for example, the following imaginary restaurant dialogue in Italian taken from Bates, MacWhinney, and Smith (1983): 1. SVO: Io mangerei un primo. (I would eat a first course.) 2. OSV: La pastasciutta Franco la prende sempre qui. (Pasta Franco it orders always here.) 3. VSO: Allora, mangio anche io la pastas- ciutta. (Well then, am eating also I pasta.) 4, VOS: Ha consigliato la lasagna qui Fran- co, no? (Has recommended the la- sagna here Franco, no?) 5, OVS: No, la lasagna ’ha consigliata Eliz~ abeth. (No, the lasagna it has rec- ommended Elizabeth.) 6. SOV: Allora, io gli spaghetti prendo. (In that case, I the spaghetti am hav- ing.) This short but plausible conversation contains all possible orders of Subject, Verb, and Object. At one level, the sen- tences given above simply serve to illustratewell-known qualitative differences between languages: Italian has word order options that do not exist in English at all. Howev- er, this qualitative variation also has quan- titative implications. We have demonstrat- ed in several different experiments that Ital- ian listeners “trust” word order less than their English counterparts. Given a sen- tence like “The pencil hits the cow,” En- glish listeners from ages 2 to 80 have a strong tendency to pick the pencil as the agent/subject. Given the Italian equivalent ”La matita colpisce la vacca,” Italians are much more likely to choose the cow as the agent/subject. Hence a qualitative differ- ence in the availability of word order types has a quantitative effect on that subset of grammatical structures that both languag- es share (e.g., SVO order). As stated above, an important aspect of the Competition Model is its emphasis on quantitative variation. Within this model, linguistic knowledge is characterized not as a set of rules but as a complex network of weighted form-function mappings. Every language must provide cues (lexical, syntac- tic, morphologic, or prosodic) that signal the presence of universal meanings (e.g., the agent role). These mappings can vary in strength within and across languages, from -1 (eg. a noun marked by Suffix A never takes the agent role) to +1 (e.g., a noun marked by Suffix B always takes the agent role). In contrast with competence-based models (which predict yes-or-no mappings), the Competition Mode! ailows for a full range of probabilistic values in between these two extremes, Thus sentence compre- hension is viewed as a process of constraint satisfaction and conflict resolution. Many different sources of information are exam- ined and weighed together across the course of the sentence, until the system “settles” into the best available fit between meaning and form. From this point of view, language development, for example, involves more than the acquisition of rules or mappings; it also involves a gradual process of “tuning” these mappings to fit the linguistic iaput. PREDICTIVE CONSTRUCTS The major predictive construct in the Competition Model is cue validity, which refers to the information value of a given source of information (e.g., preverbal posi- tion) for a particular communicative func- tion or meaning (c.g., the agent role). Vz lidity is an objective property of the cue it- self, a property of the perceptual environ- ment relative to some organismic state. It can be measured directly in samples of spo~ ken and written language, and used to de~ tive predictions concerning language pro- cessing by adults and/or tanguage acquisi tion by children. Cue validity has been anz lyzed into two components: cue availability (how often is a particular cue available when we need it to assign a sentence role?) and cue reliability (when the cuc is available, how often does it lead to the right answer?). All other things being equal, the order in which form-function mappings are acquired will reflect the strength or validity of that map- ping, with the most valid cues acquired first. Once these mappings are in place, children must adjust them until they provide an opti- mal fit to the processing environment (i.c., cue strengths reflect cue validity). Conflict validity is yet another way in which the validity of a cue can be measured (when two or more cues conflict, how often does a cue "win”, i.e. lead to a correct inter- pretation?). The distinction between over- all validity and conflict validity constitutes one of the most important discoveries in our crosslinguistic research. Overall cue validi- ty (defined as the product of availability times reliability) can explain many phenom- ena, but some puzzling exceptions remain that can only be explained by considering the way that cues behave in conflict sita- tions. Conflict validity has been particularly helpfat in explaining certain late and/or U-shaped developments in children, and in explaining how relatively infrequent struc- tures can influence adult performance. To model the organism’s knowledge about the validity of information, we postulate a subjective property of the organism called ewe strength. This is a quintessentially connection- ist notion, referring to the probability of weight that the organism attaches to a given piece of information relative to some goal or meaning with which it is associated. In other words, cue strength is the weight on the con- nection between two units. We view devel- ‘opment as the process whereby, under ideal conditions, the value of cue strength converg- es on the value of cue validity. As a result, the order of importance of cues to meaning for adult speakers ought to closely reflect cue validity estimates. This simple prediction has been confirmed repeatedly in our own crosslinguistic studies, and in work by our colleagues around the world. ‘The cue validity principle leads to a num- ber of predictions regarding the timing of sentence interpretation in adults and children. According to Bates and MacWhinney (1987), "In order to control the real-time interaction of the various cues participating in the com- petition, we believe that the parsing system engages in an ongoing updating of assign- ments of nouns to case roles” (p. 170). For example, when parsing a sentence, such as ”The dogs are chasing the cat,” the assign- ment of "dogs” as the agent is first promoted by its appearance as the initial noun. Then, the fact that ”are chasing” agrees with "dogs” in number further supports this assignment. Finally, when the singular noun cat” appears post-verbally, its binding to the object case role further supports the candidacy of ”dogs” as the agent. Thus, at each point in sentence processing, the mapping from the lexical item dogs” to the agent role is updated, a pro- cess that continues until a winning interpre- tation” has emerged. If the cue validity prin- ciple is correct, then this process of ongoing updating ought to reflect the following four predictions (Li, Bates & MacWhinney, 1993; Liu, Bates & Li, 1992; Mimica, Sullivan & Smith, 1994): (1) There should be a monotonic relation- ship between cue strength and speed of response (i.¢., stronger cues will lead to faster reaction times; weaker cues will be associated with slower reaction times). Converging cues should facilitate tence interpretation and thus lead to faster response times. Competing cues should inhibit imme- diate interpretation and thus slow reac- tion times down. Prediction 1 will interact with predic- tions 2 and 3, so that a very strong cue may still result in relatively fast reac- tion times, despite competition from weaker sources of information. @ 3 (4 These four predictions for reaction time parallel predictions that are made in Com- petition Model experiments using choice response data: (1) Stronger cues should result in more con- sistent decisions. (2) Decisions should be more uniform un- der cue convergence but less uniform under cue competition. (3) Stronger cues may win despite a con- spiracy from weaker sources of infor- mation. Although these predictions are straight- forward, they are constrained by the set of factors that are referred to in the Competi- tion Model under the rubric ewe cost (Kail, 1989; Kail & Charvillat, 1988). In princi- ple, cue validity is a property of the linguis- tic environment, one that can be quantified in a variety of ways (see McDonald & MacWhinney, 1989; MacDonald et al., 1994) from samples of input to the language learner. By contrast, cue weights are a prop- erty of the language user, a psychological construct that can only be inferred from the subject’s behavior under various experimen- tal conditions. Under ideal conditions, lan-guage acquisition is complete when the learner has a set of cue weights that are iso- morphic with cue validity (e.g., when form- function mappings are “tuned” in an opti- mal way). However, these ideal conditions are rarely met in real life. Some cues are harder to use than others. Thus one inflec- tion may be perceptually salient (ic., easy to hear) while another is easily lost in rapid conversation. For example, MacWhinney, Pleh & Bates (1985) have shown that use of case Cues is reduced in Hungarian listeners when those case markers are difficult to perceive (e.g. the accusative marker ‘-’ at the end of a consonant cluster), compared with the overwhelming reliance on case that is observed when the marker is easy to per- ceive (e.g. the ‘t? morpheme following a strong vowel). One grammatical device (e.g. long-distance” agreement between the sub- ject and the verb) may place heavy demands on memory before it can be used, while an- other (c.g., ”local” case marking on the noun) may require very little memory at all. For example, some studies have shown that the memory demands required to use long- distance agreement cues (subject-verb agree- ment and/or object-clitic agreement) have an effect on the development of sentence interpretation strategies. In particular, chil- dren who are acquiring French, Italian, Spanish, Serbo-Croatian, or German do not make consistent use of agreement cues in this sentence interpretation task unti] 5 - 6 years of age (Kail, 1989; Bates ct al., 1984). This contrasts markedly with the develop- ment of “local” case cues, which are used consistently by children before 3 - 4 years of age (Slobin & Bever, 1982; Bates et al., 1984; Mac-Whinney, Pleh & Bates, 1985). Hence cue validity and cost interact to de- termine cross-linguistic differences in the use and development of sentence interpretation strategies. ‘These examples illustrate two major types of cue cost or processing limitations, per- ceivability and assignability. The most ob- vious limitation on processing is the low-level limitation dictated by the perceivability of the stimulus, as exemplified in the case of Hungarian above. The second one, assigna- bility, refers to the amount of material that must be held in memory before a meaning assignment can be made. This is manifest in the distinction between “loca{” and “long- distance” (“topological”) cues described in the previous paragraph. Because two soure- es of information with the same cue validity can vary in cue cost, mature speakers/lis- teners occasionally come to rely more on the low-cost cue, a fact that can sometimes cancel out or reduce the predictions of the Competition Model based on cue validity. Indeed, predicted effects on reaction time do not always hold up (Kail, 1989, for French and Spanish; Li et al., 1993 for Chi- nese; Mimica et al., 1994 for Serbo-Croat- ian). These authors all report only partial fits between sentence interpretation results (which do follow the predictions of the Com- petition Model) and reaction times (which may diverge from initial predictions of the model), For example, although results sup- port predictions showing that competition between cues can slow down or inhibit pro- cessing, and convergence between cues (also called coalitions) can speed up or facilitate processing, there are clear exceptions to this generalization (Kail, 1989; Mimica et al., in press). The contribuiton of cue cost may be even greater for incompetent listeners, e.g. small children (with more serious process- ing limitations than those encountered by normal adults) and brain-injured tisteners (c.g, adult aphasics). However, in every case we must provide independent evidence for the nature and existence of a processing cost, including experiments in which cue cost is varied directly as a factor in the design. Oth- erwise, the cue cost principle can reduce to nothing more than a “fudge factor,” i.c. an ad hoc device whose only function is to save the model from disconfirmation. The four predictions on processing times described earlier have now been tested in several on-line studies of sentence interpre-tation within the CM (Kail, 1989; Li, Bates, Liu, & MacWhinney, 1992; MacWhinney, 1976, 1985; MacWhinney & Pleh, 1988; Kilborn, 1987; Von Berger, Wulfeck, Bates, & Fink, 1996; Mimica et al., in press). For the most part, results are compatible with the CM in its original form, but there are also some interesting exceptions. For ex- ample, convergence does not always lead to faster reaction times if a very strong cue is available. Instead, listeners sometimes operate as if ”enough is enough”, so that weak converging cues are either ignored (ie. have no further effect on reaction times) or actually result in a slight delay compared with items in which the strong cue acts alone. This effect is reminiscent of Simon’s principle of ”satisficing” (1960), an heretical contribution to economic the- ory based on the finding that people do not always act to maximize gain. Instead, they operate to maximize gain up to but not be- yond a certain degree of effort. We will return to this idea later, to help resolve some discrepancies between sentence in- terpretation and reaction time in our data for bilingual listeners. For present purpos- es, the point is that processing limitations (cue cost) seem to play a particularly im- portant role in on-line studies of sentence interpretation, constraining the predicted effects of cue validity on reaction time. To summarize, the model begins with the assumption that all listeners must deal with two important but occasionally conflicting tasks. On the one hand, listeners must know in advance which pieces of information in the input language carry valuable informa- tion, and merit attentional priority. On the other hand, they must be sensitive to the pro- cessing costs and timing parameters of a par- ticular language, in order to deploy resources in the most efficient way. We have referred to these two dimensions of language pro- cessing with the terms cue validity (the in- formation value associated with particular linguistic forms) and cue cost (the proc ing costs involved in using those forms, in- cluding demands on perception and memo- ry). Sentence comprehension is viewed as a process of interactive activation, a form of constraint satisfaction in which linguistic forms or ”cues” compete and converge in order to lead to a particular interpretation, i.e. the interpretation that provides the best fit to this particular configuration of inputs. Within this framework, languages can vary not only in the presence or absence of spe- cific form types (e.g. presence or absence of case-marking on nouns), but also in the relative strength of form-function mappings. In other words, there are quantitative as well as qualitative differences among language types. The Competition Model has demonstrat- ed important heuristic value. Its predictions have been tested experimentally in a num- ber of fields. Here we will examine data from research conducted in language acquisition, bilingualism, and aphasia in this order. DEVELOPMENTAL CHANGES IN REAL-TIME SENTENCE PROCESSING Language development involves much more than the acquisition of linguistic knowl- edge (c.g., words and grammatical rules). To become fluent speakers and listeners in their native language, children must also Iearn to process language rapidly and effi- ciently. Here we summarize a study which examined the sentence interpretation strat- egies used by English-speaking children between 7 and 12 years of age compared with young adults, to uncover developmen- tal changes in the cues that children use to assign agent roles (with a focus on the inter- play of syntactic, semantic, and morpholog- ic cues) and the processing times ass\ ed with use of these cues (Von Berger, E., Wulfeck, B., Bates, E., & Fink, N. Devel- opmental changes in real-time sentence pro- cessing. (1996) In: First Language, vol. 16, Part 2 (June).An important aspect of the Competition Model is its emphasis on quantitative varia- tion. Although some researchers have ac- knowledged the role of processing factors (Bloom, Miller & Hood, 1975) and varia- tion (Labov, 1969; Cedergren & Sankoff, 1974) in language development, most theo- ries of language have traditionally focused on linguistic competence, described in terms of rules or regularities that are either present or absent, acquired by children in discrete steps and applied by adults in an all-or-noth- ing fashion at various points across the course of a sentence (ef. Fodor, Bever & Garrett, 1974). In such models, statistical variations are treated as noise or as the product of some unspecified performance factor. By contrast, the Competition Model is designed to account for the probabilistic variations that characterize linguistic perfor- mance by adults and for the statistical chang- es in performance that are observed over time in children. Within this model, linguis- tic knowledge is characterized not as a set of rules but as a complex network of weight- ed form-function mappings. All other things being equal, the order in which form-func- tion mappings are acquired will reflect the strength or validity of that mapping, with the most valid cues acquired first. Once these mappings are in place, children must adjust them until they provide an optimal fit to the processing environment (ie., cue strengths reflect cue validity). We will not repeat here the major predictions of the CM already described in the previous section but to pro- vide more background for this study, we start with a brief review of the reaction time task that will be used to test developmental pre- dictions based on the model. The "Whodunit” Task The original procedure, upon which the present and many previous sentence inter- pretation studies have been based, can be thought of as a ’whodunit” task, in that the listener serves as a detective. whose iob it is to detect the perpetrator of the action. For example, subjects are asked to interpret sim- ple sentences by acting them out with small tay objects, (1) "Show me “The cow is kicking the pen- cil” 2) "Show me “The pencil is kicking the cow.” The sentence stimuli represent converg- ing and competing combinations of seman- tic cues (i.e, the contrast between animate and inanimate objects), syntactic cues (i.e., canonical and non-canonical word orders), and morphological cues (ie., presence/ab- sence of subject-verb agreement with the first or second noun). For example, in English Noun-Verb-Noun (NVN) sentences, the first noun is more likely to be the agent of the sentence than the second noun, and animate objects are more likely to be agents com- pared to inanimate objects. This design per- mits us to assess the hierarchy of importance of syntactic, semantic, and morphological cues to agent/object relations and serves as a direct test of the cue validity predictions of the Competition Model. Performance lev- els can be compared between sentences that have converging cues (e.g. Sentence 1 - an- imate first noun vs inanimate second noun) and sentences where cues compete (c.g. Sentence 2 - inanimate first noun vs animate second noun). Because the notion of per- cent correct is meaningless in such a com- petition design, the dependent variable is percent choice of the first noun as agent. English is a language in which word or- der tends to be rigidly preserved, compared, for example, with languages like Italian or Hungarian in which the order of sentence constituents can be varied in a number of ways to reflect topicalization or emphasis. Hence, it is not surprising that in studies of sentence interpretation, the main effect of word order in English accounts for up to 51% of the experimental variance (e.g., Bates, McNew, MacWhinney, Devescavi & Smith, 1982: MacWhinney, Bates & Kliegl, 1984).‘As we might expect, this effect is especially pronounced for sentence stimuli that follow canonical word order (ic, NVN (noun verb noun) -> SVO (subject verb object). The strong "second noun strategies” associated with non-canonical word orders that have also been observed are somewhat more surpris- ing (i¢., VNN (verb noun noun) > VOS (verb object subject); NNV (noun-noun-verb) > OSV (object-subject-verb)), since they do not correspond to any single legal word or- der in English. However, these findings are consistent with what is known about pockets of natural word order variation in English. For example, in English, we do occasionally find left- dislocated OSV structures, such as John I like, but Roger I can’t stand” as well as right-dislocated VOS constructions, such as "Really gets on my nerves, old Roger.” ‘These relatively rare structures may be the basis for the OSV and VOS strategies report- cd for English-speaking adults. As pointed cout by Bates and MacWhinney (1989), the second noun strategies may also reflect a partial or fragmentary fit to other, more com- mon legal options in the language. For ex- ample, imperative constructions like ”Hit the ball, John” could support a VOS interpreta- tion of VNN strings. Relative clause construc tions like ”The ball John hit went out of the park” contain an OSV fragment (literally [O[SV]V)). There is, in fact, no reason why we have to choose among these accounts for the robust second noun strategies observed in English- speaking adults. Within an inter- active activation model, like the one proposed by MacWhinney and Bates, it may be the case that all possible word order “templates” are activated in parallel at the beginning of a sen- tence, and those candidates that bear even a partial resemblance to the whole string may compete and converge in parallel to deter- mine the listener’s final interpretation (see also MacDonald et al., 1994). If this is the case, it leads to some inter- esting predictions for sentence interpreta- tion in young children. Specifically, compre- hension strategies may change over time as the pool of competing and converging sen- tence types shifts and expands. In fact, there is some reason to believe that the second noun strategies emerge relatively late in English-speaking children. In a study by Bates, MacWhinney, Caselli, Devescovi, Natale and Venza (1984), it was shown that English-speaking children rely on word or- der by 2 years of age, a tendency that in- creases markedly by age 5. However, most of the variance comes from growing use of an SVO strategy. Although there are some differences as to the age of the first appear- ance of word order strategies, these results are compatible with several earlier studies (Bever, 1970; Strohner & Nelson, 1974; Chapman & Kohn, 1978). In fact, there is little evidence for the emergence of VOS and OSV strategies in English-speaking children before age 6. This has been confirmed by extensive preliminary data analyses by Marchman and Bates (1994) of 7-12 year old, native English- speaking children, performing an off-line sentence interpretation task. While a major- ity of agent role interpretations of English- speaking children are governed by word or- der cues, the second noun strategies of the non- canonical word orders (e.g., VOS and OSV) are noi observed until the later school years. To explain these findings, MacWhin- ney and Bates suggest that the late emer- gence of VOS/OSV strategies is ticd to a *second wave” of language acquisition, char- acterized by the reorganization of sentenc level grammar in the service of higher dis course functions. This may result in an in- creased attention to and use of discourse- based left- and right- dislocated structures and complex sentences with an embedded relative clause a phenomenon which should lead, in turn, to the gradual emergence of second noun strategies in English-speaking children. However, although these second noun strategies take time to develop, there is good reason to think that the developmental tra- jectories between these two non-canonicalorders may differ. In previous off-line stud- ies of sentence interpretation in adults (Bates, Friederici, & Wulfeck, 1987; MacWhinney, Bates & Kliegl, 1984) we have reported somewhat less choice consis- tency with NNV sentences compared to VNN sentences. In addition, we have ob- served far greater preservation of VOS strat- egies compared to OSV strategies for Bro- ca’s aphasics (Bates et al., 1987 see below). We have speculated that these differences may be due to the fact that imperative sen- tences, which require a VOS interpretation, appear more frequently in conversation com- pared to more complex structures such as relative clauses which take an OSV inter- pretation. Also, there may be accessibility. differences. VNN structures occur more of- ten in “stand alone” contexts in informal conversations relative to NNV structures which tend to occur in more complex (and less frequent) contexts such as cleft construc- tions which may make them somewhat hard- er to extract. In short, modest sentence in- terpretation differences between VNN and NNV in healthy adults and striking differ- ences in adults with aphasia may be related to accessibility and frequency differences. Such factors might contribute to develop- ment contrasts between these two non-ca- nonical orders in the present study. Real-time language processing and the present study Until recently, most English-language studies of processing in adults and children have involved tasks which make use of off- line procedures (ic., methods that do not tap the real-time properties of natural lan- guage processing but instead focus on post- sentence performance). However, language processing occurs rapidly and people only rarely have time for reflection in the pro- cess of communicating. If we are to under- stand the complex mechanisms which un- derlie language representation and process- ing, then we need to study language during as well as after processing, We know from previous off-line studies of English-speaking children that sentence interpretation strategies change during the course of development. Far less is known about the relationship between the devel- opment of these strategies and the process- ing times associated with these cues across the school years. In order to fully capture the developmental time course of seatence interpretation strategies, one needs to ex- amine the protracted development of ca- nonical, first noun strategies as well as the later development of non-canonical second noun strategies. Moreover, only a real-time paradigm can reveal information about the speed with which children arrive at their decisions. However, previous studies make clear that tests of cue validity must be con- sidered in relation to cue costs and there is good reason to assume that cue cost fac- tors may change with development. Very young children approach the problem of sentence interpretation with relatively lim- ited processing resources a fact that may change over time in ways that are at least partially independent of the processing advantages that accrue from learning itself. The cue validity principles of the Compe- tition Model predict that children learn the most valid form-function mappings first. Therefore, our subjects should rely prima- rily on the most valid cues to interpret sen- tences in real time. These predictions, how- ever, may be offset by serious limitations * in information processing that raise the costs associated with otherwise valid cues to the point where children cannot use them, or prefer not to use them, or use them only in conjunction with other sources of information that make the whole problem easier to handle. Furthermore, we may also expect to find transition points across de- velopment, where children "give vp” their immature avoidance of highly valid but very costly cues, and opt for the best sources of information. At these transition points, we may see a U-shaped function in processing efficiency. That is, children may have toslow down their sentence interpretations temporarily (“consolidation”) while they start to profit from the most valid sources of information in their language ("integra- tion”). In sum, the present study was designed to examine the emergence of second noun strategies in English-speaking children and young adults and the integration of these with other sources of information. We ex- pected to replicate previous off-line findings showing a developmental progression in the reliance of young children on canonical word order cues (NVN) for thematic role assign- ment followed by the later emerging non- canonical word order cues (VNN, NNV). However, by employing a reaction time par- adigm, we hoped that response latencies would reveal insights into the time course of sentence interpretation and the ways in which strong (word order) and weak (ani- macy and agreement) cues compete and converge during development. Nore ON SUBJECTS, DESIGN AND MATERIALS: The subjects were divided into four age groups, as follows: twenty-four 7-8 year olds (mean age of 7 years; 7 months), six- teen 9-10 year olds (mean age of 10 years; 1 month), sixteen 11-12 year olds (mean age of 12 years; 0 months), and twenty- four adults (mean age of 19 years; 10 months). Subjects were asked to interpret 162 simple grammatical and semigram- matical auditorily presented sentences, administered via computer. The test sen- tences (mean length of 2053 ms) were con- structed from a pool of 12 animate nouns (animals), 7 inanimate nouns (common objects), and 15 present progressive, tran- sitive action verbs. There were six sentenc- es at each level of a 3 x 3 x 3 design, rep- resenting orthogonal combinations of word order (NVN, NNV, VNN), noun- verb agreement (AGO, where both nouns agree with the verb, and AG1 and AG2, where first and second noun, respectively, agrees with the verb), and animacy (AA, where both nouns are animate, AI and IA, where first and second noun, respectively, is an- imate and the other noun is inanimate) configurations. Examples of converging cue combina tions would include NVN/AG1/AI sentenc- es in which word order, agreement and an- imacy promote the first noun as agent or ‘VNN/AG2/IA sentences in which ali cues promote the second noun as agent. Compe- tition configurations would include NNV/ AGO/AI sentences, for example, in which word order promotes the second noun, ani- macy promotes the first noun and agreement is neutral. Conspiracy items (a subset of the competition class) would include NVN/ AG2/IA sentences in which word order (first noun) is pitted against weaker agreement and animacy (second noun) cues. Table 1 shows sample sentences from each of the 27 cell combinations. WO AGR ANI SAMPLE NVN AG) AA The elephant iskising the cow. ‘AL The bearis scratching the pencil TA Therockiseatingthe cow. AGL AA Thebear is graibingthe aebrs, Al Thezebrais patting the rocks IA Thepenciisgrabbing the elephants. AG2 AA Thezebrasis patting the cow ‘Al Thezebrasishitting the pencil TA The locks grabbing the elephant NNV AGO AA Theearthe horses biting Al Thepigthe blocks licking, TA Theballthezebrais licking AG] AA The pigthe elephants is grabbing. ‘Al Thebearthe halls is kissing 1A The pencil the zebras is scratching AG2 AA The pigsthe elephants kissing, AL Thehorses the rock s hitting JA The balls the beat is pushing. VNN AG AA. Isscratchingthe zebra the cow. AL Islicking the cow the rock TA Tspushing the pencil the cow. AGL AA _Isbitingthe pigthe elephants ‘AT licking the elephant the rocks. TA [shitting the rock the bears. AG2 AA Ispatingthe bears the zebra. AL Ischasingthe bears the block 1A __Iseatingthe blocks the horse. Table 1: Sample sentence stimuli from each of the 27 word order x agreement x animacy cell combinationsRESULTS: PERCENT CHOICE FIRST NOUN We first examined choice data (percent choice first noun) to determine which cues subjects rely upon during sentence interpre- tation and whether or not changes in cue preference occur with development. Collapsed across age groups, word order ac- counted for 83% of the experimental variance. More importantly, there was a significant group main effect as well as a significant group by word order interaction. As Figure 1 illustrates, the three basic word order preferences displayed by En- glish language listeners in previous studies were replicated here: NVN as SVO; VNN as VOS; NNV as OSV. As expected, while all groups showed a very strong preference for SVO with NVNsentences, developmental differences were observed with the non-canonical orders. There is evidence of the emergence of VOS interpre- tation (second noun choice) with VNN sentenc- es in the youngest age group, although it is not as well developed as it is in the older child groups whose performance mifrors adult levels. How- ever, while the OSV interpretation of NNV sen- tences is just below adult levels in the two oldest child groups, the 7-8 year old children show no such second noun bias as indicated by chance performance. Figure 1. Interaction of Group and Word Order on percent choice first noun. The main effects of agreement and an- imacy were also significant. In contrast with the main effect for word order, how- ever, these main effects were small, cach accounting for only 3% of the experimen- tal variance. The two-way interaction be- tween agreement and animacy also was significant leading to more consistent choices when these cues converged to pro- mote the first noun (60% first noun choice for AG1/AI) or the second noun (33% first noun choice for AG2/IA). Competition be- tween these two weaker cues resulted in chance performance (AG2/AI = 46%, AGI/IA = 47%). Additionally, both these variables were involved in significant two- way interactions with word order. Figures 2 and 3 illustrate these two-way interac- tions with word order, reflecting the com- petition and convergence effects predict- ed by the Competition Model (i.c., more consistent interpretations when cues point in the same direction). When agreement or animacy cues converge with word or- der cues (e.g. NVN/AG1, NVN/AI, NNV/ AG2, NNV/IA, VNN/AG2, VNN/IA) choice responses are more consistent. However, although there is some soften- ing of choice consistency when agreement or animacy cues compete with word or- der, it is quite clear from these graphs that word order cues are dominant in English (ie., word order always ”wins” when there is a competition with animacy or agree- ment). eeoeeaes Figure 2. Interaction of Word Order and Agreement on Percent choice first noun.Figure 3. Interaction of Word Order and Animacy on percent choice first noun. As expected, the interaction between group and agreement reached significance. This interaction is illustrated in Figure 4, which shows that the effect of noun-verb agreement is restricted almost entirely to the adult group. This fact is also evident in the significant three-way interaction of group, word order, and agreement where although word order predominates in any competition with agreement, adults are af- fected most by this competing source of information (Table 2). For example, for adults, first noun choice drops from 98% for converging cue sentences (NVN/AG1) to 65% for competition sentences (NVN/ AG2). Similarly, when cues converge to promote the second noun in VNN/AG2 sentences, first noun choice is very weak (3%) but rises to 32% for competition sen- tences (VNN/AG1). With respect to the children, the oldest are beginning to show some limited sensitivity to competing cues, however, in general there is very little evi- dence of cue competition in the choice per- formance of the child groups. RESULTS: ADIUSTED REACTION TIME. As expected, analysis of the baseline RT data revealed a significant difference be- tween the four age groups, reflecting a de- crease across age levels. To examine the temporal processing of these syntactic, se- mantic, and morphological cues, we turn to the analysis of variance conducted on the adjusted reaction times. Collapsed across groups, the word order contrast 2 counted for 70.6% of the experimental vari- ance within the reaction time data and the results were in the predicted direction. Sub- jects were fastest overall with canonical NVN stimuli (1881 ms, SD = 478), fol- lowed by VNN sentences (2068 ms, SD = 671), and slowest on NNV sentences (2492 ms, SD = 616). This processing speed hi- erarchy is similar to the acquisition hierar- chy we observed with the choice data (NVN > VNN > NNV). The interaction between, group and word order was also significant and is illustrated in Figure 5. The first strik- ing element in this graph is the non-mono- tonic relationship between age and reac- tion time. Adults are actually slower than any of the child groups on NVN and VNN. As we saw in the choice data and shall see again shortly, this is due to the fact that adults are more affected by competition Figure 4. Interaction of Group and Agreement on percent choice first noun. ‘TByearols —$10yearohs1/2yearolds College students ‘AGL AGH AG? AGL AGO AG2 AGI AGO AGD AG) AGO AG2 NN HH 8 8 3 7% GB OB Wi 3 eu 29 6 Bu B93 AW 56 3 2S U BHM HHS Table 2: Percent choice first noun for group x word order x agreement interactionfrom other sources of information than chil- dren. The second striking element in this graph is performance of the 9-10 year old children. This group displays the slowest processing of NNV items slower than the youngest children! This result seems anom- alous until considered in the context of the choice data. Recall that the 7/8 year old children showed no evidence of a second noun strategy for NNV sentences with per- formance at chance levels. Taking this into account, it is not surprising that the young- est children respond quickly since they sim- ply are pushing response buttons at random. On the other hand, an OSV strategy was observed in the 9/10 year old children. We attribute the slower processing times as- sociated with these NNV sentences to the fact that the OSV strategy is relatively new and processing efficiency is still develop- ing in this group. Hence, this pattern is an example of the costs associated with inte- grating and consolidating new sources of information as predicted by the Competi- tion Model. Although the main effect of agreement did not reach significance, the main effect of animacy did. Subjects were faster when there was an animacy cue (AI or IA) than when no animacy contrast was available (AA). Agreement and animacy were both involved in significant two-way interac- tions with word order. As illustrated in Figure 6, these interactions reflect the pre- dicted effects of cue competition and con- vergence. For example in the canonical NVN order, subjects were faster when word order and agreement cues converged (NVN/AG1 - 1784 ms, SD = 438) than when they competed (NVN/AG2 - 1971 ms, SD = 503). In the non-canonical NNV order, the story is much the same, subjects were faster under cue convergence (NNV/ AG2 - 2440 ms, SD = 620) than competi- tion (NNV/AG1 - 2564 ms, SD = 631). Although in the predicted direction, there was only a 30 ms advantage for VNN/AG2 over VNN/AG1. re = hw es Figure 5. Interaction of Group and Word Order on adjusted reaction time. Te _ é 0 Ee: Figure 6. Interaction of Word Order and Agreement on adjusted reaction time. We also obtained a significant three-way interaction of group by word order by agree- ment. Although this interaction is complex, careful examination reveals some interest- ing developmental differences in the pro- cessing of cues under conditions of compe- tition and convergence. From Figures 7a - 7c, which illustrate the three-way interac- tion of group by word order by agreement, we can examine the relative effects of com- petition and convergence by comparing re- action times for each word order type when the preferred response is compatible or in- compatible with subject-verb agreement. For all four age groups, faster responses were obtained on NVN convergence items than on NVN competition items. While this dif- ference increased across the four age levels(from a 116 ms difference at 7-8 years, to 126 ms at 9-10 years, to 249 ms at 11-12 years, and 254 ms for adults), the difference between NVN convergence (AG1) and com- petition (AG2) items was significant only for the two oldest age groups. 12 oes aes as in SS HE ys = iz is an Figure 7 a) Interaction of Group and b) Interaction of Group and ¢) Interaction of Group and Agreement on adjusted reaction time Agreement on adjusted reaction time Agreement on adjusted reaction time ‘on NUN items only on VNN items only ‘on NNV items only For adult listeners, we can see some ef- fect of competition and convergence be- tween agreement and non-canonical word orders, The same conclusion does not hold for children. Across child groups, process- ing times did not differ whether or not agree- ment cues converged or competed with the two non-canonical word orders or provided no contrast whatsoever. Taken together, these data indicate that interactions between morphologic and word order cues seen in adult subjects take time to develop, extend- ing well into the school years for canonical SVO and into adolescence for non-canoni- cal orders. We will now turn to a final set of analyses directly comparing choice behav- ior and RT, to determine how this relation- ship changes with development. RESULTS: CHOICE CONSISTENCY AND ADIUSTED REACTION TIME, In order to focus on cue cost factors and the relationship between choice and process- ing times, we assessed the choice data with a statistic called choice consistency. Choice consistency reflects the degree to which choice performance in a given cell deviates from chance. For example, if the first noun was chosen 89% of the time in a given cell, the choice consistency statistic for that cell wonld he 300% (29% above the 50% chance level). On the other hand, if the first noun was chosen only 11% of the time, the choice consistency statistic for that cell would also be 39% (50% chance level minus 11%). Hence this statistic gives us an estimate of the relative certainty” associated with ac- tor- assignment in a given cell, without re- gard for the direction of that estimate {i.c., toward the first noun or toward the second noun). Large consistency scores mean a great degree of certainty; small consistency scores mean that choices hovered around chance. Choice, adjusted reaction times, and choice consistency scores by group for the 27 data cells (See appendix) were entered into a series of correlational analyses within and across age groups. First we correlated the choice consisten- cy scores for each child group with the adult choice consistency profile, across the 27 cells in the design. These analyses indicate that as children get older, their choice strategies become more consistent with the patterns displayed by adults (r = .22, p < .15 for the 7-8 year olds; r = .51, p < .005 for the 9-10 year olds; r = .57, p < .001 for the 11-12 year olds). Second, we were interested in the rela- tionship between speed and choice consis- tency within age levels. The simplest pre- dictian wanld he an ineresce puer timethe correlation between speed and choice, on the assumption that children are moving toward an optimally efficient system. Results of this analysis were somewhat surprising. Within each of the four groups, we did find a robust negative correlation between choice and speed (i.¢., more consistent choice/faster RTs), but the magnitude of this correlation did not increase monotonically with age: 7, Pp < .0001 for the 7-8 year olds; -.86, p < .0001 for the 9-10 year olds; 3, p < .0001 for the 11-12 year olds; 9, p < .005 for adults, In fact, the relationship between speed and choice con- sistency appears to be weaker in adults than it is in children. We suggest that adults are paying a price for their highly consistent per- formance, a price that is reflected in slower Teaction times and a less direct relationship between speed and certainty of response. We have shown that children gradually converge on an adult-like pattern of choice behavior across these 27 cells. Do they also converge on an aduit-like pattern of reac- tion times? To ask this question, we corre- fated the 27 mean RT scores for each child group with the 27 mean RTs observed in adults. The resulting correlations between child RT and adult RT were very high at every age level: r = .90, p < .0001 for the 7- 8 year olds; r = .91, p < .0001 for the 9-10 year olds; r = 93, p < .0001 for the 11-12 year olds. In other words, children and adults tend to find the same items difficult or easy to resolve even though they do not always provide the same interpretation for those items, Finally, we correlated the choice consis- tency scores for each child group (over the 27 cells of the design) with RT scores for adults. The point of this analysis was to de- termine the extent to which item difficulty (indexed by RTs in the adults) predicts con- sistency in children. This analysis also yield- ed very high correlations, though slightly lower than the ones summarized above. In other words, the patterns of sentence inter- pretation displayed by school-age children are predicted surprisingly wel! by the reac- tion times that adults display to the same sentence types (i.e., more consistent choice in children/faster RTs in adults). Adults appear to have developed a kind of "delay of gratification”. They confront difficult sen- tence types and come up with a consistent interpretation, even though that interpreta- tion may be relatively expensive to compute. By contrast, children seem to engage in a kind of "satisficing”, relying on the cues that are easiest to use (even for adults) and avoid- ing or ignoring cues that are hard to pro- cess, but would (if they were used) bring themi closer into alignment with the cue va- lidity structure of the language. In the terms provided by the Competition Model, we may conclude that sentence interpretation strat- egies displayed by school age children are jointly determined by cue validity (strong correlations with adult interpretations for children 9 years and up) and cue cost (re- flected in adult reaction times), but cue cost is more important for children than it is for adults. Discussion The results from the present study, ex- amining sentence interpretation strategies in _ English-speaking children and young adults, are in line with Competition Model predic- tions that (1) stronger cues in a language (cg., word order for English) will result in more consistent choices, (2) decisions will be more uniform when cues converge, and (3) stronger cues may prevail despite con- spiracies from weaker cues (c.g. animacy and agreement). Also, all groups showed clear preference for the second noun as agent in non- canonical VNN sentences. Developmental differences were revealed, however, with the least consistent VOS in- terpretation pattern shown by the youngest group. For NNV sentences, the developmen- tal differences were even more striking, The 7/8 year old children showed no evidence of having a second noun strategy whereas the older child groups, although below adultlevels, showed clear evidence of an OSV strategy. Our study reveals a systematic de- cline with age in young subjects in the ten- dency to generalize a first noun strategy for NNV sentences. These results suggest that this drop continues through 7 years of age and it is not until around 9 years of age that children show signs of adopting the second noun OSV strategy that is so clearly evi- denced by adults. By contrast, development of an VOS strategy with VNN sentences occurs earlier. As we noted above, there are structural and frequency differences be- tween these two word orders that may give rise to the contrasting patterns we’ve seen in children and in the greater preservation of VOS observed in aphasia. Turning to the reaction time results, we observed that all age groups were fastest overall processing NVN sentences, followed by slightly slower processing times for VNN. Slowest reaction times were observed in all four groups for NNV sentences. Notably, the 7/8 year old group was fastest overall with NNV sentences and the 9/10 year old chil- dren were slowest. However, recall from the choice results that the 7/8 year old group showed no evidence of a second noun strat- egy for NNV sentences. Taken together, their fast” response times are consistent with random choice performance. A very different picture emerges for the 9/10 year old children. This was the youngest age group in whom we observed an OSV strate- gy with NNV sentences. We believe their slower processing times reflect a newly ac- quired OSV strategy. In short, while the 9/ 10 year old group’s choice performance sug- gests a fairly stable second noun choice for the weaker non-canonical word order (NNV), the reaction time measure illustrates the cost of a developing OSV interpretation strategy. Finally for the adults, the RT dif- ferential we observed between VNN and NNV sentences may be indexing frequency and accessibility differences at the process- ing level, long after choice strategies have stabilized. In the present study, choice and RT re- sults were compatible with Competition Model predictions that weaker cues for En- glish (animacy and agreement) will play only a minor role in sentence interpretation strat- egies and have their greatest influence when cues converge. For example, weaker effects were seen with animacy and agreement cue processing, and again in the predicted di- rection. When agreement or animacy cues converged with word order, faster process- ing times generally resulted and when they competed, slower times were observed. Finally, our analyses comparing choice consistency and reaction time at various points in development underscore the con- stant trade-off between cue validity (which leads to adult patterns of choice behavior) and cue cost (which affects reaction times in a similar way at every age level). Taken together, our results illustrate developmen- tal differences in the degree to which listen- ers rely upon syntactic, semantic, and mor- phologic cues to assign agent roles. ‘There is some debate about whether this technique can be viewed as an ”on-line” measure of sentence processing. Until re- cently, most on-line research has concentrat- ed on processing at the word level. And most current research at the grammatical level has used ”punctate” single-event probes (usual- ly words) to assess notions of grammatical difficulty and/or goodness-of- coreference with reaction times in the 300 - 1000 milli- second range. By contrast, our studies of sentence-level processing typically involve much longer reaction times, somewhere in the 2000 - 2500 millisecond range. We are not using single-word probes and we are not looking ‘at a single-word process. Instead, we are examining events that by definition transcend the boundaries of a single word. In our study, RTs were measured from the onset of each sentence, not from an event midway through the sentence. So, it must be the case that a response decision is made after sufficient information is available tothe subject on which to base a response. In this study then, RTs are upper bounds on the time subjects take to make decisions. Taken together with other studies of sen- tence interpretation by ourselves and col- leagues, we anticipate that our data can be used as a starting point to "work backwards” to identify places in the stimuli where suffi- cient information has accumulated to make parsing decisions. Conclusions The successful application of a real-time processing paradigm to the study of language development opens the door to a whole new area in the study of the development of lan- guage knowledge as well as the information processing mechanisms that access this knowledge. Overall, our findings are large- ly compatible with Competition Model pre- dictions. However, the model in its original form does not, for example, predict the U- shaped increase in reaction times that indi- cates development and integration of the NNV strategy in 9-10 year old children that we discovered in the present study. Also, as mentioned earlier, several recent studies of sentence interpretation report only a partial fit to the model (Li et al., 1993; Kail, 1989; Mimica et al., 1994). We are not discour- aged by this as we always have viewed the Competition Model as a developing frame- work, not a completely predictive theory (yet). We view our experiments as data in- puts which lead to adjustment of the model, not (dis)confirmation of a finished theory. ‘The important point for present purposes is that this new round of real-time studies will lead to some important revisions in the Com- petition Model. Scientists working with a model like ours have two choices: revise the model to deal with discrepant findings, or abandon it when it becomes weighed down with ad hoc assumptions and find a better model. We think the Competition Model has considerable heuristic value, and it is one of the few frameworke that are currently avail. able for the study of real-time sentence pro- cessing, Investigators working in more than a dozen languages have found the model useful, We shall see what happens in the next few years when the Competition Model is put to more severe tests, in new develop- mental studies of sentence processing in normal and impaired populations. REAL-TIME SENTENCE PROCESSING IN BILINGUALS As Grosjean (1992) and Paradis (1987) have pointed out, bilinguals do not behave like two monolingual speaker/listeners housed in a single brain. Instead, the evi- dence to date suggests that bilinguals dis: play a qualitatively different form of Jan- guage processing, based on a system that is in some very real sense ”in between” the individual’s two codes (Cutler & Mehler, 1992; Kilborn, 1989; Kilborn & Ito, 1989; Liu, Bates, & Li, 1992; Vaid & Pandit, 1991). This is true not onty for late bilin- guals, ic. individuals who have acquired a second language after puberty. It is also true for many carly bilinguals who have achieved what appears to be native-like fluency in both of their codes (Liu, Bates, & Li, 1992; Kutas, Kluender, Bates & Van Petten,’ 1993; Vaid & Pandit, 1991), The Compe- tition Model has also been applied for many years to the study of second language ac- quisition and bilingualism. In this section, we will summarize results from a recent on- line (reaction time) study of sentence in- terpretation in Spanish- English bilinguals who have grown up in a bilingual commu~ nity, compared with monolingual controls in both language groups (Hernandez, A., Bates, E., & Avila, L. On-line sentence in- terpretation in Spanish-English bilinguals: What does it mean to be “in between”? Applied Psycholinguistics, 1994, 15, 417- 446). The primary purpose of this study was to provide information about what it means to be in between, in real-time processing tema.‘A number of linguistic theories have been offered to account for the in-between status of bilingual speakers, with special emphasis on adult learners of a second language (Hyams, 1986). Most of these are what have been called competence models”, which stress the set of rules that a bilingual speak- er applies within each language at various stages of the learning process . One of the best known proposals of this type is the In- terlanguage Hypothesis of Selinker and col- Ieagues (see Sclinker, 1972; Selinker, Swain, & Dumas, 1975), a model which makes spe- cific predictions about the direction of trans- fer and the compatibility or incompatibility of particular rule types at successive stages in second- language learning. Competence models have proven useful for many purpos- es, but they provide no mechanisms to ac- count for the quantitative nature of linguis- tic performance in bilingual speaker/listen- ers. The relevant performance facts include variations in the degree to which bilinguals rely on a given rule type, and variations in the timing of rule application. These quanti- tative dimensions are critical for a full defi- nition of fluency (and lack thereof) in late learners of a second language, and they are also necessary to explain the differences in processing that have been observed between early "balanced” bilinguals and monolingual speakers of the same language types. The Competition Model attempts to pro- vide a more dynamic account of monolin- gual and bilingual language processing. Within this framework, languages can vary not only in the presence or absence of spe- cific form types (e.g. presence or absence of case-marking on nouns), but also in the relative strength of form-function mappings. In other words, there are quantitative as well as qualitative differences among language types. To illustrate, consider some well- known contrasts between English and Span- ish. Spanish offers a rich set of markings for subject- verb agreement, and as a result, subject-verb agreement is also a strong cue to agent-object relations. At the same time, Spanish is a pro-drop language (i.e. it per- mits omission of the subject in fres ind- ing declarative sentences), and it permits a great deal of word order variation. As a re- sult, word order is a relatively weak and unreliable cue to semantic roles. English behaves quite differently: there are very few contrasts in verb morphology to mark the subject role (”I eat....you eat....they eat...”), but subjects cannot be omitted in free- stand- ing declarative sentences, and word order is rigidly preserved in most sentence types. Hence subject-verb agreement is a weak cue but word order is a very strong cue to agent- object roles in English. Note that Spanish and English have the same basic word or- der (Subject-Verb-Object, or SVO, in prag- matically neutral sentences), and both lan- guages have at least some forms of subject- verb agreement. So the primary difference here is a matter of degree: which cues to meaning should the listener trust in assign- ing semantic roles? The CM predicts that Spanish listeners will rely primarily on morphological cues in sentence interpretation, ignoring word or- der if the two sources of information do not agree. By contrast, English listeners should rely primarily on word order cues, at the expense of morphological information. Two previous studies have confirmed this predic- tion for Spanish and English. (Wulfeck, Jua- rez, Bates & Kilborn, 1986; Kail, 1989). What will happen when Spanish and English are processed by the same brain? Cross-linguistic variation and bilingualism Inthe cross-linguistic studies that we have described so far, language is treated as a between-subjects variable. In the study of bilinguals, cross-linguistic variation must be treated as a within-subjects variable. Bilin- gualism presents a particularly interesting challenge to the CM, because the same in- dividual has to develop two different sets of mappings between form and meaning, map- pings that can coalesce or diverge in a num-ber of ways. At the same time, the CM of- fers some useful mechanisms to describe patterns of transfer and dominance in dif- ferent bilingual situations. Four distinct pat- terns of dominance are theoretically possi- ble, and could be described with the quanti- tative mechanisms provided by the CM: (1) Differentiation refers to the use of sep- arate strategies for each language, iden- tical to the strategies shown by mono- linguals. Although many theories of bi- lingualism assume that differentiation is the desired end-point of second-lan- guage learning, note that this is also equivalent to the claim that bilinguals ought to behave like two monolinguals housed in a single brain. (2) Forward Transfer is defined as the use of L1 strategies in processing L2. This is the pattern that we would expect to find in late bilinguals with relatively lit- tle experience in their second language, and it is an effect that should diminish over time. (3) Backward Transfer refers to the use of L2 strategies in processing L1. Back- ward transfer represents a feedback process in which new learning influenc- es old strategies. In the extreme case, backward transfer could result in the elimination or replacement of L1 pro- cessing strategies by L2, a possible step in the direction of language loss (Wong- Fillmore, 1991). (4) Amalgamation is our term for a situa- tion in which bilingual listeners apply a single set of strategies to both their lan- guages, derived by merging the two cue hierarchies used by monolinguals. In the limit, amalgamation could be viewed as a combination of forward and backward transfer. All four patterns of dominance and trans- fer have been reported in studies of sentence interpretation in bilingual listeners within the CM, including studies of Dutch-English, Span- ish-English, Chinese-English, German-En- glish, Japanese-English and Hindi-English speakers (for reviews, see Kilborn and Ito, 1989; McDonald, 1989; Vaid and Pandit, 1991). In late bilinguals, forward transfer is the most common pattern, and in some indi- viduals it seems to persist more than 30 years after immersion in an L2 environment (Bates & MacWhinney, 1981). At the same time, backward transfer has been observed in adult learners with no more than 2 - 3 years of ex- posure to L2 (Liu et al., 1992). In early bilin- guals, the two most common patterns that have been observed so far are forward trans- fer (L1 dominance) and amalgamation. Of the four logically possible patterns of transfer and dominance, full differentiation is the least common although many individuals do dis- play partially differentiated patterns of amal- gamation (c.g. Spanish- English bilinguals who rely primarily on agreement in both of their languages, but make greater use of word or- der in English). In short, we find a wide array of in-between patterns in bilingual listeners, suggesting that "bilingualism is a matter of degree” (Kilborn, 1987). These data constitute a combination of * good news and bad news for the CM. The good news is that the quantitative apparatus of the CM can be used to describe all of these in-between patterns. This is not true for traditional competence models, where in-between states can only be described in terms of a presence or absence of rules. The bad news is that the CM cannot (as current- ly formulated) predict or explain this vari- ability in sentence- processing strategies. Forward transfer does follow from the learn- ing principles of the CM. Indeed, McDonald (1989) has shown that forward transfer is gradually reduced with years of experience in a second language, in line with the pre- dicted effects of cue validity on learning. However, this is simply a quantitative vari- ant of the predictions made by virtually all models of second-language learning (c.g. the Inter-Language Hypothesis). If we take the learning principles of the CM at face value, then we should expect bilinguals to movetoward optimal mappings of form and mean- ing for each of their two languages, with cue strength (a property of the learner) converg- ing gradvally but inevitably on cue validity (a property of the linguistic envionment). Instead, we find that (a) some individuals are remarkably resistant to the cue validity structure of their second language (i.e. per- sistent forward transfer in both early and late bilinguals), and (b) others use an amalgam- ated form of processing that falls somewhere in between their two language types, a solu- tion that is less than optimal if processing were determined solely by cue validity. In the past, failures of cue validity have led to an increased understanding of cue cost. Pre- sumably, the bilingual listener persists in the use of in-between strategies because these strategies are a good solution to the special problems posed by the bilingual condition. To investigate the interaction of cue validi- ty and cue cost in bilingualism, we need to move toward on-line methods that tap into the temporal course of sentence processing. To our knowledge, Kilborn’s investiga- tion of sentence interpretation in German- English bilinguals was the first on-line study of sentence interpretation in bilingual listen- ers (Kilborn, 1987 and 1989). This was also the first on-line version of the CM sentence interpretation task. Bilingual subjects lis- tened to digitized auditory sentence stimuli in German and English (in separate sessions, administered in a counterbalanced order). ‘Their task was to say the name of "the one who did the action” as quickly as possible without making a mistake, even if the sen- tence was not finished. The sentences were all simple active declarative forms with two nouns (animate or inanimate) and a transi- tive action verb. Variables included three levels of word order (NVN, VNN, NNV), four levels of subject- verb agreement (first noun agrees, second noun agrees, both agree, neither agree), and four levels of se- mantic reversibility (animate noun first, in- animate noun first, both animate, both inan- imate), English monolingual controls in Kil- born’s study replicated the sentence inter- pretation patterns reported in previous off- line investigations: a very strong effect of word order for all three word order types (interpreted as SVO, VOS, and OSV, re- spectively), with significant but very weak effects of animacy and agreement. The re- action time findings were largely compati- ble with these interpretations, reflecting the four predictions described earlier. As a group, Kilborn’s German-English bilingual subjects displayed strong patterns of forward transfer in their assignment of agent-object roles. In both languages, they showed the sentence interpretation profile that had been reported in earlier offsine studies of Ger- man: strong effects of agreement and ani- macy, with relatively weak effects of word order. The reaction time findings were in general accord with these results, although Kilborn did not pursue them in detail. Inthe Hernandez et al. study, we investi- gated sentence interpretation strategies and associated reaction times in Spanish-English bilinguals living in the United States, com- pared with monolingual English and Span- ish-speaking controls. In contrast with the late bilinguals in Kilborn’s study, our sub- jects have spent most of their lives using both languages on a daily basis. In an off-line study of sentence interpretation in this pop- ulation, Wulfeck, Juarez, Bates, & Kilborn (1987) report a high proportion of individu- als who fit the amalgamation profile de- scribed above. It may be the case that amal- gamation, while not an ideal solution from a monolingual perspective, is a good solu- tion for bilingual speakers who have to deal with a number of real-time processing con- straints which are not encoun-tered by mono- lingual speakers (i.e. interference, switch- ing from one language to another). Assum- ing that these amalgamation profiles repli- cate in the present study, we will have an opportunity to investigate the time course and processing costs associated with a pat- tern of sentence processing that is genuine- ly "in between”.Metnop Our subjects were one hundred under- graduates or graduate students at the Uni- versity of California, San Diego or at the University of Baja California, Mexicali. Sub- jects were divided into three different groups: bilingual Spanish-English, monolingual En- glish and monolingual Spanish. The bilinguals had spent an average of 19.75 years speak- ing Spanish as compared to 14.35 years spent speaking English. The monolinguals were all native speakers of either English or Spanish. ‘A questionnaire of language history was used in the bilingual group to acquire information regarding experience and proficiency in En- glish and Spanish (Liu, et al., 1992). The experimental stimuli consisted of 162 sentences for both English and Spanish para- digms. These were generated by a random se- lection from a pool of twelve animate nouns (animals), seven inanimate nouns (common objects) and 15 transitive action verbs. The four dependent variables were language (whether auditory sentence was in Spanish or English), word order sequences (NVN: noun-verb-noun, ‘VNN: verb-noun-noun, or NNV: noun-noun- verb), agreement between noun(s) and verb (Ag0: ambiguous agreement, Ag!: verb agrees with the first noun, or Ag2: verb agrees with the second noun), and animacy (AA: animate first noun and animate second noun, AI: ani- mate first noun and inanimate second noun, or IA: inanimate first noun and animate sec- ‘ond noun). For each language, crossing the three levels of word order with the three levels of agreement with the three levels of animacy yielded 27 possible sentence types (ice. cells). An example of these can be seen in Table 3. Each cell contained six sentences, thus ac- counting for the total of 162 sentences. The six sentences within each cell varied in accordance with the agreement condition. In the Ag0 con- dition, three sentences utilized two singular nouns and the other three utilized two plural nouns. In the Agi and Ag? conditions three sentences were first noun singular and second noun plural, and vice-versa for the remaining three. The procedure used included instruct- ing subjects to determine which of the two nouns was responsible for doing the action in the sentence (cf. the "whodunit” task). Condon ngish Spanish NYNAZIAA — Thedogischaingthecms — Eipsroestconetandoles acs NVNABIAL —Thedogischsingthecyps _Elpenoesticorteao lists NYNARUIA —Thcopschasingthedogs—Lataestcoreaspees NYNAQAA —Thedopachaingtheoms Elpemoestinconceanb ls vacs NYNAQA —Thedopaechaingtbecups Elperbesinconeails tas NYNAQIA —Theaparechsngte dbs Latuasincartandlospets NYNAQAA —Thedngschaingticow —_Elpenbetcoreteann iva NYNAQAL —Thedipischaingliecup —Elpenbewicoreenin twa NYNAGIA —Thecopischsngthe dog Lateral pero VANARIAA — chaingtedgtecows Eticonteadoelperoisacs VANARIAL — ching dogtecus Ex4creeandel pe astaas VINARIIA —chaingtecupthecmss —Extécoreadolatalisvacs VINADAA —Arecaingthedoebeows Euinemecadel pels acs VINAGAL — Arelaingdedoecys Esineneteanoc pols tas VANAQIA —Acclusngtecptbecms Estnowrteaola tals acs VANAMAA —Schaingtedoeticany Escorted! peolaaca VANAGIAL —Ischingtedpthep ——Estcurteandnelerolatra VANAGIA —Ischsingeciptecow —-Escomteavoltalaaca NNWARLAA —Thedogtiecowsischsing —Elperolasracasstumetendo NNWAglAL —Thedogtecupsisctasng Elpenoas assumed NNVAgIIA — Thecpthedosisctasng _Lataalasvarasestcotendo NNWIQDAA —Thedogticowsarectaing Elperobisassstinaeteand NWWAQAL — Thedogtecupsarecasng Elpelastzaestincoreardo NNVAQIA — Thecupthedogsarecasng—Lataalsvacasestincoeando NAVAYLAA —Thedcgthecovischaing El prolsracasestcoreteanio NAVAGLAL —Thedopthecsischasing —Elperolstwasestéconeand NOVA Thecupedopsischasng _Lataals acess coeteando Table 3: Sample stimuli for each of the 27 conditions of the experiment Resutts: Cuoice Data Here we will describe three sets of analy- ses. The first will be a set of analyses looking at monolingual English- and Spanish-speak- ing groups. The second set of analyses will involve comparisons of bilingual groups to the monolingual group performing in the same language (ic. bilinguals in English compared to monolinguals in English; bilinguals in Span- ish compared to monolinguals in Spanish). The third set of analyses will be a within- group analysis of the bilingual subjects in both Spanish and English. English Monolinguals (EM) and Span- i olinguals (SM). The order of im-portance of cues in both these languages con- firms the pattern observed in off-line stud- ies, Monolinguals in English use word or- der primarily and some subject-verb agree- English esesseassss os8eseguess PERCENT FIRST-NOUN CHOICE NN! yn! NV’ ment in assigning agency to a noun. Mono- linguals in Spanish use agreement primarily and some animacy in determining the agent of a sentence. Spanish 4 AGL ° Aco Aga NVN' VNNT NNV Figure 8. Word order by verb agreement interaction for percent first-noun choice (monotinguals) With these two patterns of performance in mind let us turn to the bilingual results. Agent Choice with Bilinguals. Here we will begin with two separate between-group com- parisons, using the bilinguals and monolin- guals. Each of these comparisons will help to elucidate the similarity between monolinguals and bilinguals in Spanish (SM and SB) and English (EM and EB). Then we will look at within-group analysis of the bilingual subjects only (SB and EB) to determine the amount of differentiation between the two groups. First, a 2x3x3x3 (group by word order by verb agreement by animacy) ANOVA. compared the EB with the EM group. Re- sults (compare Figure 8 and 9) show that English 3 sssesgesse sssgguge PERCENT FIRST-NOUN CHOICE, NN! vNNT NNV’ bilinguals use verb agreement more than monolingual English speakers. The signif- icance of these group effects implies that bilinguals and monolinguals use a differ- ent hierarchy of cue weights to interpret En- glish sentence stimuli . The merged hierar- chy of cues displayed by bilinguals tested in English can be summarized as follows: Agreement > SVO >VOS > Animacy = SOV. In this case, the less prototypical SOV: interpretation is at chance when up against the animacy cue but VOS wins any compe- tition with animacy. This suggests that the most prototypical interpretation (SVO) re- tains its strength while the less prototypi- cal word order effects are being lost. Spanish 4 AGt © Aco = AG2 ‘NVNT VNNT NNV Figure 9, Word order by verb agreement interaction for percent first-noun choice (bilinguals).Next, a 2x3x3x3 (group by word order by verb agreement by animacy) ANOVA compared the SB with SM groups. Looking at the word order by agreement interaction for both groups (compare figures 8b and 10), it is clear that bilinguals use word order more consistently than Spanish monolinguais in choosing the agent of a Spanish sentence. In addition, the SB group relies very heavily on word order when there is an absence of agreement. This is not true of Spanish mono- linguals. It is also important to note that bi- linguals are much less sensitive to animacy than monolinguals. geeseess PERCENT FIRST-NOUN CHOICE ar AA ar Figure 10. Group by animacy interaction between ‘monolinguals and bilinguals in Spanish The higher-order three- and four-way in- teractions in this analysis all reflect the same contrast in hierarchy of importance of cues. Spanish monolinguals show the pattern Agreement > Animacy > Word Order. As a group, bilinguals tested in Spanish show a pattern that is Spanish-dominant in many ways, but differs in the relative weighting of animacy information, as follows: Agree- ment > SVO > Animacy = VOS = OVS. Notice that canonical word order is stron- ger than animacy but that the less prototyp- ical VOS and OVS interpretations are on equal ground with respect to animacy. We have clear evidence here for trans- fer and amalgamation in these Spanish-En- glish bilinguals. Let us turn now to a com- parison of the same bilinguals in Spanish (SB) and English (EB), to determine the extent to which these individuals behave in a differentiated fashion between their two languages. The relevant comparisons here can be found between Figures 8b and 9. The data from the bilinguals in both of their lan~ guages were placed into a 2x3x3x3 (language by word order by verb agreement by anima- cy) within-subjects ANOVA. The resulting interactions with language, and the corre- sponding graphs, show that there our bilin- guals do (as a group) display some degree of differentiation. In the word order by verb agreement interaction, subjects favor word order slightly more in English than in Span- ish. For agreement the opposite is true. Sub- jects use verb agreement as a cue to sub- jecthood more often in Spanish than in En- glish. In addition, in both languages the sub- jects use word order in the absence of agree- ment contrasts. Hence these bilinguals have clearly adopted an amalgamated set of strat- egies, with partial differentiation. RESULTS: REACTION TIME Reaction times in this experiment pre- sented a more formidable problem for a number of reasons, one of which is the fact that individual words are substantially long- er in Spanish. To avoid the difficulty of in- terpreting and comparing absolute reaction times in the two different languages, all re- action times (for a given individual, in a giv- en language) were converted to z-scores. This enabled us to directly compare the bi- linguals to themselves and to the two mono- lingual groups. Here we will report three sets of analyses. The first will examine and com- pare monolingual English- and Spanish- speaking groups. The second will involve comparisons of bilinguals to the monolin- gual group performing in the same language (ie. bilinguals in English compared to mono- linguals in English; bilinguals in Spanish compared with monolinguals in Spanish). The third set of analyses will be a within- group analysis of the bilingual subjects in both Spanish and English. Z-score reaction times for monolingual control groups: