General Discussion

The present study addressed inhibitory effects in speech recognition. Furthermore, it explored the potentially different patterns of results that are obtained by electroencephalographic as compared to behavioural studies. In a behavioural and in an ERP experiment disyllabic prime fragments preceded trisyllabic target words.

The fragments were either identical to the target onset (e.g. ano – ANORAK), deviated from the target in the vowel of the second syllable (e.g. ana – ANORAK) or were completely unrelated to the target (e.g. paste – ANORAK). While the behavioural results showed blocking of activation in the related condition (Experiment 1) or weak inhibition (Experiment 2), the ERP amplitudes differentiated between the three conditions in terms of graded activation in deflections that are related to target processing (P350 and N400), and revealed blocking of the related condition in a response related deflection.

In the introduction to this Chapter we hypothesized that the inhibition effect observed by Soto-Faraco et al. (2001) could be due to strategic effects rather than lateral inhibition among competitors. This conclusion was based on the fact that they used systematically varying fragments in the control compared to the test condition. The results of the present experiments, where these aspects were well controlled, did not reveal clear evidence for inhibition (no effect in Experiment 1, only weak inhibition in the behavioural data of Experiment 2). These findings again indicate that behavioural inhibition is not an obligatory finding in fragment priming across studies. They further strengthen our assumptions that reaction times may be biased by strategic processes, which may have a differential impact across

experimental paradigms. Maybe inhibition occurred in the reaction time data of the EEG-Experiment because participants read both target words (e.g. ANANAS and ANORAK) in each of the four blocks. Hence, hearing for instance ano- activates the recently-read word ANORAK more than it usually would and the word ANANAS less. If this is true, EEG-results suggest that this is a later process, otherwise no graded activation would have been observed in the P350 component. Considering the models of speech processing discussed in Chapter 1, this speaks against the notion of lateral inhibition in the Shortlist model (Norris, 1994). The Trace model (McClelland & Elman, 1986) also incorporates lateral inhibition among competing word candidates, however, the level of activation cannot drop below zero, and hence an inhibited item would not be worse off than an item that has received neither activation nor inhibition. Shortlist B (Norris & McQueen), DCM (Gaskell

& Marslen-Wilson, 1997) and FUL (Lahiri & Reetz, 2002) do not incorporate lateral inhibition and hence are in accordance with the reported results.

Interestingly enough, the behavioural results in the purely behavioural Experiment 1 reveal a different pattern than the electrophysiological results.

Lexical decision responses in the identical condition were facilitated. However, a deviation in just one vowel between prime and target eliminated any reaction time advantage despite the considerable segmental overlap. On the other hand, amplitudes of event related potentials suggest graded activation in the mental lexicon, in terms of the P350 as well as the N400 effect. From 300-400ms the P350 effect as described by Friedrich and co-workers (2004; 2005; 2008; 2009) was observed. In the left anterior ROI amplitudes were most negative for the identical condition, less negative in the related condition and least negative in the control condition. The P350 effect has been interpreted as reflecting amount of lexical preactivation in fragment priming experiments. If a target word has been preactivated by the preceding prime fragment, the amplitude is more negative than if it was not preactivated. The observed graded levels in amplitude in the present study are consequently interpreted as indicators for graded levels of lexical activation. In the later 350-500ms time window, the N400 effect showed a similar sensitivity to phonological overlap, as did the P350 effect. As unexpected events usually elicit most negative N400 amplitudes (Kutas & Federmeier, 2000), in our study the control condition was most negative over posterior sites, the identical condition most positive and the related condition intermediate. Thus, in contrast to Friedrich (2005), both, the P350 and the N400 component provide evidence for graded activation.

A corresponding pattern of results between the behavioural experiment and the EEG experiment was only observed when analyzing response-locked ERPs.

Here the three conditions deviated approximately 200ms before the button press and showed the most negative amplitude for the identity condition and less negative amplitudes for the related and the control condition without a difference between the latter. This was interpreted as reflecting the status of a decision level rather than the actual pattern of lexical activation. It is reasonable that the lexical entry of a target word in the related condition is still intermediately active while at the decision level it has already been rejected as a possible candidate. Visual inspection of separate averages for right vs. left handed button presses confirmed that this effect does not simply reflect a response potential of the button press, because it was not lateralized. Furthermore, the response was the same in all three conditions, namely “word” (as opposed to “pseudoword”).

A lot of studies in the past decades have investigated the variability that is tolerated in lexical access. Following an initial study by Marslen-Wilson &

Zwitserlood (1989), it has been assumed that mismatch on a single phoneme (e.g., honing changed to woning) prevents lexical activation. However, results of other studies suggested that items are not activated in an all or none fashion but that there is graded activation, depending on the goodness of fit between information in the signal and information in the lexical representation (Connine, Blasko & Wang, 1994; Connine, Titone, Deelman & Blasko, 1997; Frauenfelder, Scholten &

Content, 2001; Bölte & Coenen, 2002; Friedrich, 2005). With that respect, several studies further suggested that it is not the goodness of fit on the segmental level that determines the amount of activation, but on a subsegmental, featural level (Connine, Titone, Deelman & Blasko, 1997; Frauenfelder, Scholten & Content, 2001; Bölte & Coenen, 2002; Friedrich, Lahiri & Eulitz, 2008).

The present ERP results show that gradual activation and blocking are temporally distinguishable. Given the P350 and the N400 results, we can conclude that in a first step word candidates are activated as a function of their goodness of fit between the input and the lexical representation. Matching words are more strongly activated than partially mismatching ones. However, it does not appear that partially mismatching candidates are inhibited if better matching entries exist.

Blocking of partially mismatching candidates, which is reflected in the reaction time data, occurs quite late during recognition. It is only reflected in ERPs that immediately precede the behavioural responses. Therewith, our data suggest that competition occurs relatively late during word recognition.