The convergence of corpus linguistics, psycholinguistics and functionalist linguistics
As we have seen in Chapter 7, functionalist linguistics in the broad sense (including cognitive linguistics) is increasingly making use of corpusbased methods, and in turn informing the analyses of corpus linguists. In this chapter, we will show that this phenomenon extends as well to experimental psycholinguistics. We will also discuss the implications of the rapprochement of functionalist linguistics and psycholinguistics with corpus linguistics with regard to the neo-Firthian school of thought which we surveyed in Chapter 6; we will argue that in the neo-Firthian school, this rapprochement with functional linguistics has taken a very different form. As we saw in Chapter 6, one of the bases of the neo-Firthian or so-called ‘corpus-driven’ approach is a rejection of non-corpus-derived theoretical frameworks. To explicitly adopt a functionalist theory as the basis for a corpus-driven study would be distinctly peculiar from the
neo-Firthian perspective. Indeed, some of the stronger forms of the neo-Firthian position – such as that espoused by Teubert, for instance – explicitly reject the notion of a convergence of neo-Firthian corpus linguistics and functional or cognitive linguistics, with Teubert (2005: 2) claiming that corpus linguistics ‘offers a perspective on language that sets it apart from received views or the views of cognitive linguistics, both relying heavily on categories gained from introspection rather than from the data itself’. Nevertheless, we wish to argue that such a convergence is in fact taking place, stemming on the neo-Firthian side from work by Sinclair and others from the 1990s onwards. Our basis for making this case is that, when we closely examine the findings of the most extensively developed neo-Firthian theories – in particular, Pattern Grammar and Lexical Priming – we will find that many of these conclusions have also been arrived at by one or more branches of functional linguistics or psycholinguistics. These congruent conclusions stem from wildly different sets of evidence and are, of course, expressed using very different descriptive apparatus. But certain fundamental insights – namely, the inseparability of lexis and grammar, and the nature of grammar as secondary to, and emergent from, lexis – have been arrived at by both functional linguists and neo-Firthian corpus linguists, largely independently of one another.
8.2 Corpus methods and psycholinguistics
In this chapter, then, we have two main topics. Firstly, in section 8.2 we will consider the role of corpora in experimental psycholinguistics, as we considered their role in functionalism in Chapter 7. Psycholinguistics as a discipline is methodologically rather different to functionalist theoretical linguistics, but it shows signs of a similar trend with regard to corpus methods – that is, that over recent years there has been more and more use of corpus data within psycholinguistic research, and a convergence or rapprochement between the findings of psycholinguistic experiments and of corpus investigations. Secondly, section 8.3 discuss the convergence of findings, regarding in particular the ontological status of grammar, lexis and language itself, between neo-Firthian corpus linguistics, functional linguistics and psycholinguistics.
Corpus methods and psycholinguistics
Overlapping cognitive linguistics (which we discussed in the previous chapter), but in many ways distinct from it, is the field of psycholinguistics –
and in particular that branch of psycholinguistics whose methodology is mainly experimental. In the latter approach, the primary source of data is various types of laboratory tests on human subjects (or, as we will see later, computer models). While experimental psycholinguistics is not usually considered a branch of functional-cognitive linguistics, its fundamental methodological assumption – that the nature of language in the brain or mind can be investigated in much the same way that experimental psychology in general looks at other aspects of the nature of thought – is in accordance with the general tenet of functionalism that there is no absolute divide between form and function, between language and non-linguistic cognition. Psycholinguistics is a very broad field, and there is absolutely no room here for a full review of it – nor even to treat comprehensively all research which has linked psycholinguistics with corpus data and methods. We must therefore confine ourselves to an extremely brief and purely indicative survey. To characterise psycholinguistics in very broad terms, we might say that it is focused on two primary issues (which are closely interrelated, as Ellis 2002 illustrates): language learning and language processing. There are other topics of interest of course, such as the evolution of the language faculty. However, we will limit ourselves here to looking at how corpora have been used in some psycholinguistic investigations into first language acquisition, second language acquisition and language processing. 8.2.1
Corpus data in experiments on language processing ⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢⅢ
Language processing has been investigated experimentally in a number of ways. Two that are reasonably common are self-paced reading experiments
corpus, psycholinguistics and functionalism
and eye-tracking experiments. Both are means of investigating the speed with which particular segments of language are processed. In a self-paced reading experiment, participants work at a computer running a specially designed program. The computer shows one word of a sentence at a time to the participant, who presses a button to get the next word once they have read the word currently on screen. The program records the time for each button-press, so that the relative speed of reading for each word is known. Typically, after each sentence participants have to answer a (very easy) question about the content of the sentence – this prevents participants from just clicking through sentences without actually reading for meaning. The results of such an experiment can be used to infer what elements (morphological, syntactic or semantic) are processed easily, and which are more difficult and thus require more processing time. This in turn can give indications about what is actually happening in the brain. Although useful, selfpaced reading experiments may potentially be misleading in that fluent readers do not typically read one word at a time, in sequence, without ever going back in the text. In fact, it is known that a reader’s journey through a sentence of printed text can be quite complex, with multiple movements back and forth. This type of evidence is gathered in eye-tracking experiments (see Rayner 1998 for a review). Again, participants are given the task of reading sentences presented on a computer screen, but this time an entire sentence is presented at one time, and specialised video equipment records the movements of one of the participant’s eyes as it looks at different positions in the sentence immediately after the sentence appears on screen. The resulting data is much richer, but correspondingly rather more difficult to interpret, than self-paced reading data. These kinds of experiments may seem remote from the concerns of corpus linguistics. However, there are at least two ways in which corpus data can play an important role in the design and interpretation of such experiments. Firstly, corpus data can be used as a check on the naturalness of the language task that the experiment sets its participants. For instance, Frisson and Pickering (2001) summarise the results of a series of eye-tracking experiments aimed at investigating the processing of words which are ambiguous between a literal and a metaphorical meaning, when the part of the sentence prior to the ambiguous word does not provide sufficient cues to indicate which meaning is intended. But Deignan (2005: 114–17), in a review of this study, points out that in fact, such cases almost never occur in corpora of real usage: in all the examples she looks at, some aspect of the preceding context – possibly in an earlier sentence – indicates which meaning is intended. So, for instance, the word campaign literally relates to warfare and metaphorically relates to politics. In any given real example of campaign from a corpus, the prior context is overwhelmingly likely to give some indication whether a military campaign or a political campaign is intended; so by the time the reader gets up to campaign, it is already effectively disambiguated. On this basis, Deignan argues that if an experiment presents participants with a word such as campaign without any indication in the foregoing text as to whether it is literal or metaphorical, as Frisson and Pickering’s experiment did, then that
8.2 Corpus methods and psycholinguistics
experiment is actually ‘forcing participants to tackle problems that are not faced in normal discourse’ (Deignan 2005: 117). If this is the case, then it may be argued that while such an experiment may indeed tell us something interesting about the processing of ambiguously metaphorical words, it cannot tell us about the normal processing of language in use. We can see, then, that a corpus-derived awareness of how words (and other linguistic items) are actually used can serve as a useful anchoring-point for psycholinguistic experimentation. This is not to say that unnatural language should never be used in an experiment – there are cases where non-idiomatic language may itself be the object of study, for instance Millar’s (2011) study of how errors in collocation, of the type made by non-native speakers of English, can affect processing speed in self-paced reading. What is undesirable is a situation where experimental tasks include highly unnatural language without the experimenter being aware that this is the case. Secondly, corpus data can be used as a source of frequency data in the construction of test sentences in self-paced reading or eye-tracking experiments. Often, the test sentences used will not be drawn directly from corpus data, because the analysis of the resulting data may require certain aspects of the sentences to be controlled across different examples. For instance, if we are primarily interested in the time taken to process (say) the verb in a sentence, then we might well wish to control the length and syntactic structure of the preverbal elements (as well as, potentially, that of the rest of the sentence). We are unlikely to find such controlled sentences in a corpus! But even when invented example sentences are used, it is entirely possible for the creation of the sentences to be informed by frequency data of various sorts extracted from a corpus. The study by Millar (2011) which we cited above uses this approach: Millar’s test sentences are all fabricated, but each is built around an observed non-idiomatic collocation extracted from a learner corpus. A perhaps more straightforward use of frequency data drawn from corpora is exemplified by the eye-tracking experiments of McDonald and Shillcock (2003a, 2003b). They investigate whether the co-occurrence frequency of a pair of words (as established in a large corpus, in this case the BNC) can predict the ease of processing of the second word in that pair. The co-occurrence frequencies are expressed, in this case, as transition probabilities; that is, given that the first word in the pair is X, what is the probability that the second word is Y? In this case, the probability is equal to the number of times the sequence X-then-Y occurs in the BNC, divided by the total number of instances of word X – this is fundamentally very similar to a collocation calculation. McDonald and Shillcock (2003a) look at the processing of verb–object pairs, contrasting pairs where the object is probable, given the verb – e.g. avoid confusion – and pairs where it is less probable – e.g. avoid discovery. The frequencies of these bigrams in the BNC are 50 and 2 respectively, relative to 7,823 instances of the wordform avoid in total. McDonald and Shillcock’s eye-tracking data showed that participants’ eyes fixed on the object noun for a shorter time when they were reading a high-probability transition than when reading a low-probability transition. This suggests that the