Automaticity and Control in Language Processing

Antje Meyer, Linda Wheeldon, Andrea Krott, Antje Meyer, Linda Wheeldon, Andrea Krott

• Publisher: Psychology Press
• Year: 2007
• ISBN:
  9781135419660
• Edition: 1
• Topic: Psychology
• Subtopic: History & Theory in Psychology
• Index: Psychology

The use of language is a fundamental component of much of our day-to-day life. Language often co-occurs with other activities with which it must be coordinated. This raises the question of whether the cognitive processes involved in planning spoken utterances and in understanding them are autonomous or whether they are affected by, and perhaps affect, non-linguistic cognitive processes, with which they might share processing resources. This question is the central concern of Automaticity and Control in Language Processing.

The chapters address key issues concerning the relationship between linguistic and non-linguistic processes, including:

• How can the degree of automaticity of a component be defined?
• Which linguistic processes are truly automatic, and which require processing capacity?
• Through which mechanisms can control processes affect linguistic performance? How might these mechanisms be represented in the brain?
• How do limitations in working memory and executive control capacity affect linguistic performance and language re-learning in persons with brain damage?

This important collection from leading international researchers will be of great interest to researchers and students in the area.

1 Automaticity of language production in monologue and dialogue

Simon Garrod
University of Glasgow, UK

Martin J. Pickering
University of Edinburgh, UK

Over the last decade, there has been much discussion among social psychologists about automaticity in relation to the control of social behaviour. In this chapter, we discuss automaticity in the context of language production, a key aspect of social behaviour. We consider automaticity and control in the context of both monologue and dialogue. The discussion is motivated by a recent claim by Pickering and Garrod (2004) that the fundamental mechanism underlying dialogue is an automatic process known as interactive alignment. We review this claim, and then analyse in more detail automaticity and language production.

Pickering and Garrod (2004) argued that successful dialogue involves the alignment of situation models, and that this occurs via three processes: (1) an automatic mechanism of alignment involving priming at all levels of linguistic representation and percolation between these levels; (2) a mechanism that repairs alignment failure; (3) alignment via explicit “other modelling”, which is used as a last resort. One criticism of the model is that language comprehension and production are clearly not automatic, and that they involve different kinds of strategic processing, conscious guidance and monitoring, and so on. For example, Shintel and Nusbaum (2004) stated in their commentary that dialogue cannot be as automatic as we claim, because “automaticity implies a passive process in which the input is processed in an invariable, inflexible manner” (p. 210). In response, our claim is not that language processing is largely automatic (which would be ludicrous) but rather that alignment is largely automatic.

In this chapter, we focus on language production. We propose that when one interlocutor produces an utterance, the other interlocutor is likely to produce an utterance that reflects some of the linguistic properties of the first utterance. For example, if A produces chef, B is more likely to produce chef when he comes to speak; but if A produces cook, B is then more likely to produce cook (Brennan & Clark, 1996; Garrod & Anderson, 1987). Our claim is that a major reason for this alignment is that the comprehension of chef (or alternatively cook) activates representations corresponding to this stimulus in B’s mind (roughly corresponding to a lexical entry). These representations remain active, so that when B comes to speak, it is more likely that these representations will be employed, and hence it is more likely that he will utter chef (or cook). Whereas other factors of course affect whether B will actually utter a particular word, the tendency is enhanced by prior comprehension of that particular word. Our proposal is similar at other levels of linguistic representation, such as syntax (Branigan, Pickering, & Cleland, 2000) or phonology (Pardo, 2006). We assume that this tendency to alignment is automatic.

This account contrasts with an account in which alignment is a strategic process. If B wants to talk about cooking, he may have to select between cook and chef. If A has just uttered chef, then B is more likely to decide to use chef as well. There may be many reasons for this. For example, Giles and Powesland (1975) proposed that people speak in the same way as their interlocutors when they wish to affiliate to them. On their account, alignment is the result of a decision to behave like one’s interlocutor for social reasons. Somewhat similarly, Brennan and Clark (1996) proposed that interlocutors set up tacit “conceptual pacts” to use the same term to refer to the same thing. For instance, if A refers to a shoe as a pennyloafer and B does not query this use but rather responds to A’s instruction, then both A and B assume (1) that B has accepted this referential term, and (2) that both know (1). (In contrast, if B queried A’s expression, then the pact would not be formed.) On their account, alignment is therefore the result of a process of negotiation that is specialized to dialogue and involves inference.

Strategic accounts assume that alignment is entirely the result of additional social or inferential processes. In contrast, interactive alignment proposes that there is an automatic tendency to align underlying all dialogue. Strategic decisions to align particularly strongly or not to align at all occur “on top of” this basic tendency, which we argue is automatic. The goal of this chapter is to explicate this notion of automaticity in relation to more general claims about automaticity in language production.

The chapter falls into three sections. First, we explicate the notion of automaticity as a graded property of cognitive processes (Bargh, 1994; Cohen, Dunbar, & McClelland, 1990). We then apply this graded notion of automaticity to language production in relation to standard models whose motivation is primarily research in monologue (e.g., Levelt, 1989), and then show how they need to be modified in the context of dialogue. Finally, we address the question of the automaticity of alignment.

The decomposition of automaticity

In cognitive psychology, automaticity and control were traditionally treated as all-or-none phenomena (Posner & Snyder, 1975; Shiffrin & Schneider, 1977). Automatic processes were considered to be involuntary, not drawing on general resources, and resistant to interference from attended activities or other automated activities (Johnson & Hasher, 1987). Controlled processes were just the opposite: voluntary, interfering, and subject to interference. More recently, this view has been challenged. For instance, Cohen et al. (1990) proposed that processes exhibit different degrees of automaticity as a function of what they call their strength of processing. They defined strength of processing in relation to processing pathways within a connectionist network. Strong processing pathways have strong connections between units and modules, leading to fast and accurate transmission of information along the pathway. One aspect of such pathways, which we consider below, is that they offer less optional choices than weaker processing pathways. Cohen et al. (1990) argued that strength of processing determines the extent to which processes are open to interference from other processes that may share portions of the same pathway. So a strong process is likely to be considered automatic because it tends to be efficient and resistant to interference. A weaker process is likely to be considered more controlled because it is less efficient and more likely to suffer from interference from the stronger process.

Another challenge to the all-or-none view of automaticity comes from researchers in social cognition, notably Bargh (1994). He argued that any processes as complex as those studied by social psychologists are bound to be made up of different components, some automatic and some controlled. He identifies four criteria, what he calls the “four horseman of automaticity”. The first horseman is awareness. Automatic processes are likely to be those of which the subject is not aware. Examples of automaticity at this level include the effects of subliminally presented stimuli, as in subliminal priming of attitudes or activation of stereotypes (Bargh & Pietromonaco, 1982). The second horseman is intentionality: whether the subject needs voluntarily to instigate the process. For example, Stroop interference effects are considered automatic because they occur whether the person wants them to or not (Pratto & John, 1991). The third horseman is efficiency. Automatic processes are more efficient than controlled processes: they are faster, require less focal attention, and so on. Bargh’s final horseman is controllability. Automatic processes are those which a subject cannot easily control in the sense of stopping or modifying the process once it is under way. However, to avoid confusion with the more normal contrast between automatic and controlled processes, we will refer to this as interruptibility instead (notice that efficiency and interruptibility are also key properties of strong and weak processing pathways in Cohen et al.’s (1990) terms).

Bargh’s horsemen sometimes ride together, but sometimes do not. For example, evidence suggests that stereotypes are accessed unintentionally but that use of the stereotype to support a judgement is subject to some degree of control (Fiske & Neuberg, 1990). Hence, automaticity is a graded notion. Because language production is a complex activity with many distinct components, it is also appropriate to consider whether it, too, might exhibit graded automaticity.

The components of language production and their automaticity

This section is in two parts. We first discuss standard models of language production, principally that of Levelt (1989), in relation to the automaticity of their components. However, such models were developed largely to provide an account of production in monologue. Hence, we then discuss how such models need to be developed in order to account for language production in dialogue and how such models now relate to automaticity.

Automaticity in models of language production

It should be clear that language production involves both automatic and controlled processes. On the one hand, some decisions about what to talk about are clearly controlled, but some aspects of constructing the particular form of what one says are automatic. In order to determine the extent to which production is automatic, we need to decompose accounts of language production with respect to the four “horsemen”. To do this, we focus on Levelt’s account of production, making reference both to the general account (Levelt, 1989) and to the account of lexical encoding (Levelt, Roelofs, & Meyer, 1999).

Such accounts assume that language production involves converting a nonlinguistic representation of what a person wants to talk about into a sequence of speech sounds (or writing or signs). Levelt (1989) divides production into three broad stages: conceptualization (deciding what to talk about), formulation (constructing linguistic representations), and articulation. In his framework, speakers construct a series of intermediate representations as they move from message to sound, and also monitor this process. Each step in this process could be automatic to a greater or lesser degree with respect to Bargh’s four criteria, as we shall see. One important point is that most aspects of language production involve some degree of choice between alternatives. It may be that degree of automaticity is related to the extent to which the speaker has to make such choices because choice relates both to intentionality and strength of processing.

In Levelt’s first stage, the speaker conceptualizes the message that she wishes to convey. This seems to be controlled with respect to all four criteria. It is clearly amenable to introspection. It is generally assumed that people intend to convey the message they produce (Levelt’s book is subtitled Speaking: From Intention to Articulation). (Any exceptions, such as echoing a previous speaker without understanding her message, are presumably highly atypical.) It is not efficient, in the sense that people can put considerable effort into deciding what to talk about next (or, indeed, whether to talk at all). Finally, it is interruptible, because people can decide to abandon or change what they are planning to talk about. Note, however, that thought is of course not an entirely controlled process, with associations between ideas occurring spontaneously, it being impossible to suppress unintended thoughts (Wenzlaff & Wegner, 2000) and so on. Hence, even deciding what to talk about has some automatic components.

Of course, there is a considerable difference between situations where a speaker’s decision about what to say is not obviously driven by any external stimulus and situations where the speaker responds to a particular stimulus, as in picture-naming experiments. In cases of the former, such as preparing and giving a speech, the four criteria are obviously met. In the latter case, it is still true that people are aware of responding, have to decide to respond to the stimulus (e.g., to follow instructions in an experiment), and can stop themselves midstream or change their description (e.g., if the picture changes: Van Wijk & Kempen, 1987). Picture naming presumably involves less effort than coming up with an idea from scratch, but almost certainly it is impossible to remove any effort from the process. There is also evidence from picture–word interference experiments that pictures do not automatically activate their names except when the task is to name the picture. In a word-to-word translation task, Bloem and La Heij (2003) found that semantically related word contexts interfered with translation, but semantically related picture contexts facilitated translation. They suggested that this is because the picture contexts automatically activate only semantic features (hence priming selection of semantically related words), whereas the word contexts automatically activate the word forms themselves (hence interfering with selection of semantically related words). In other words, there is no automatic lexical activation from pictures unless you intend to name them (see Roelofs, 2003, for a related argument about Stroop effects). In fact, Levelt (1989, p. 20) argued that conceptualization is the only truly controlled process in speech production; all subsequent stages he considered automatic.

One possibility is that the results of this conceptualization are fed into the formulator. For example, when deciding on the concept TIGER the speaker might simply activate the features that uniquely specify tigers, such as CAT-FAMILY, HAS-STRIPES, and so on. These features can be seen as corresponding to the (nonlinguistic) idea and are therefore part of the “vocabulary” of thinking. These would then be used in performing lexical access on the word tiger. However, Levelt (1989) assumed that the prelinguistic idea is converted into a linguistic concept (principally to avoid the hyperonym problem), and that this representation is the input to the formulation process (see also Levelt et al., 1999). So we can ask whether the process of accessing the concept TIGER from features like CAT-FAMILY and HAS-STRIPES is automatic or not. Although there is no research that directly addresses this question, there is indirect evidence from models of Stroop interference results that establishing the linguistic concept is a controlled process (Roelofs, 2003). This is used to explain why conflicting ink colour (e.g., green) does not interfere with naming the word written in that colour (e.g., RED), whereas the conflicting word does interfere with naming the ink colour. In other words, whereas reading words is automatic and not subject to interference, naming colours is not automatic and so is subject to interference.

In Levelt’s model, speakers then formulate their utterance, turning the message into a series of linguistic representations. Although specific models differ on exactly the levels proposed, they tend to agree that there is a stage of grammatical encoding, a stage of phonological encoding, and an interface with lexical items encoded as lemmas, which inform both grammatical and phonological encoding. Crudely, the speaker has to decide the grammatical form of what she is saying, decide what words (and morphemes) to use, and decide what phonological form to employ. Here, the question of automaticity becomes more complex.

Choice of lexical items appears to be the most controlled. For example, in describing a particular dog crossing a road the speaker needs to choose an appropriate level of description for the dog. Should she say Rover, a spaniel, a dog or even a four-legged mammal? Similarly, she has to decide between synonyms and near synonyms (chair vs. seat, cook vs. chef, etc.). To the extent that speakers are always presented with the problem of choosing an appropriate level of lexical specification, the process cannot be completely automatic. In terms of the four horsemen, speakers can become aware of lexical choice, before, during, and after uttering a word. It is less clear whether they have to be aware of all word choices. For example, they might never be aware of some short function words that they produce. But there is no doubt that people can be aware of lexical selection and, presumably, normally are. They can also be aware of making a choice between alternatives, without necessarily being aware of all the factors that bias them toward one choice or the other. Lexical access must surely be intentional: it is hard to see how it could take place without voluntary instigation of the process. Again, however, one possible exception involves function words without any inherent meaning (e.g., the complementizer that), for which Levelt (1989) and others propose “indirect election”: such words are selected on the basis of the compatibility with other words, and do not require prior activation of a concept.

Lexical access is clearly not entirely efficient, in that there are some conditions...