Lexical Access in the Light of Behavioural versus EEG Data

In Chapter 3 we directly confronted behavioural and electrophysiological data, gained in the same experimental paradigm and with the same stimuli. This was motivated by contradictory findings concerning the mechanisms of lexical access and selection (Soto-Faraco, Sebastian-Gallés & Cutler, 2001; Friedrich, 2005). One finding of our experiments was, that the behavioural data gathered in a purely behavioural study differed from behavioural data gathered during EEG recording.

This was probably due to the repetition of prime fragments and target words in the

latter. As a consequence we suggested to rely only on behavioural data that has been recorded in a proper and purely behavioural paradigm. A second finding was, that the behavioural reaction time data showed a different pattern of results than the ERP data. Whereas reaction time data suggested that only the prime fragment in the identity condition (ano- ANORAK) preactivated the target word, ERP amplitudes suggested that most preactivation happened in the identity condition, and reduced preactivation in the related condition (ana- ANORAK). As a consequence we concluded, that the P350 and N400 components reflect different stages of processing than behavioural response data. While P350 and N400 may still tap into lexical activation processes, or at least mirror the pattern of activation in the lexicon, the response data depict a later stage of decision-making, where a complete match between prime and target wins over all other prime-target combinations. Further, we observed that response-locked rather than stimulus-locked ERPs reflected the pattern of the response time in the behavioural experiment, with a two way distinction between the identical condition on the one hand, and the related and the control condition on the other hand. This lent additional support to our interpretation that stimulus-locked ERPs reflect earlier stages of processing than behavioural responses.

However, contradictory findings are not only observed between behavioural and EEG data, but also within each of these methods. We will confine here to behavioural data obtained in the experiments reviewed in Chapter 3 that examined segmental processing and those in Chapter 4 that examined stress processing. These studies reported quite different findings concerning the consequences of a difference in one segment between prime and target (i.e. in what we called the related condition). Two characteristics of these studies struck us as possible confounds or explanations for the observed differences: the employment of semantic versus phonological priming or monitoring designs, and the use of words versus pseudowords as primes. Those studies that employed semantic priming (Marslen-Wilson & Zwitserlood, 1989; Connine, Blasko & Wang, 1994;

Lahiri & van Coillie, 1999; Bölte & Coenen, 2002) reported that response times for the related condition fell in between those for the identical and those for the control conditions. In other words, they observed graded activation for the related condition. Studies that employed phonological rather than semantic priming report more heterogeneous results, with all three possible patterns: graded activation of the related condition (Radeau, Morais & Dewier, 1989; Gaskell & Marslen-Wilson, 1996; Connine, Titone, Deelman & Blasko, 1997; Cutler & van Donselaar, 2001;

Frauenfelder, Scholten & Content, 2001; Gow, 2001; Cooper, Cutler & Wales,

2002; Wheeldon & Waksler, 2004), no activation, i.e. no difference between the related condition and the control condition (Gaskell & Marslen-Wilson, 2001;

Gow, 2002; Felder, Friedrich, Lahiri & Eulitz, 2008; Felder, Chapter 3), and inhibition, i.e. slower responses to the related condition than to the control condition (Radeau, Morais & Dewier, 1989; Soto-Faraco, Sebastian-Gallés &

Cutler, 2001; Cooper, Cutler & Wales, 2002; Donselaar, Koster & Cutler, 2005). If one considers the lexical status of the prime, it strikes that any time the prime was a pseudoword (Radeau, Morais & Dewier, 1989; Gaskell & Marslen-Wilson, 1996;

Connine, Titone, Deelman & Blasko, 1997; Marslen-Wilson & Zwitserlood, 1989;

Frauenfelder, Scholten & Content, 2001; Gow, 2001; Bölte & Coenen, 2002;

Lahiri & van Coillie, 1999; Wheeldon & Waksler, 2004) or a pseudoword fragment (Cutler & van Donselaar, 2001), priming or phoneme monitoring resulted in a pattern of graded activation. This held true for both, semantic and phonological priming. If existing words or word fragments were employed as primes, the picture is again more complex, with graded activation in case of semantic priming (Marslen-Wilson & Zwitserlood, 1989; Connine, Blasko & Wang, 1994) and in case of phonological priming with stress differences (Cutler & van Donselaar, 2001; Cooper, Cutler & Wales, 2002), no preactivation of the related condition in cases of phonological segmental priming (Gaskell & Marslen-Wilson, 2001; Gow, 2002; Felder, Friedrich, Lahiri & Eulitz, 2008; Felder, Chapter 3), and inhibition with segments and stress under investigation in phonological priming (Radeau, Morais & Dewier, 1989; Soto-Faraco, Sebastian-Gallés & Cutler, 2001; Cooper, Cutler & Wales, 2002; Donselaar, Koster & Cutler, 2005). This shows the rather immense impact of the experimental design on the outcome and interpretation of the results. The differences between semantic and phonological priming may suggest that phonological priming taps into different or more diverse processes, while semantic priming consistently yields graded activation, suggesting that a prime that slightly differs from a target word nevertheless activates the semantic content of this target. Regarding the lexical status of the prime, pseudowords consistently lead to graded activation, while words and word fragments often cause inhibition or no activation at all. Probably competitors or better matching alternatives play an important role here. A pseudoword has no better matching competitor to the later target word and hence might activate it more than a word (fragment) that finds a better match than the respective target in the mental lexicon.

In order to explain this, one either has to assume that items inhibit each other in case of competition, or that processes of lexical access and selection are aware of the number of alternatives and are willing to accept an imperfect goodness of fit

between signal and representation in case of no better choice. The latter is appealing from a psychological point of view and asks for more flexibility in lexical selection than models usually incorporate. In any event, understanding the reasons for these differing methodological impacts on behavioural results will add a lot to our understanding of the brain processes during language comprehension.

In this light it may be wiser to explore the assumptions of asymmetric mental representations made by the FUL model with pseudowords rather than with word fragments as we did it. The use of pseudowords as primes will allow for testing effects of match, nomismatch and mismatch ‘in isolation‘ so to speak, i.e.

independent of frequency and competitor effects, that show up as soon as existing words are used as primes. However, in a next step word primes and the relative contributions of factors such as neighbourhood, word frequency, and competitors should be assessed, always with respect to the ternary matching logic of the FUL model. The challenge will be to evaluate the weight of each contributing phenomenon in lexical access and selection. Ultimately, the model will have to incorporate and account for both result patterns, those obtained by pseudowords and those by words.