Discussion of Experiment 3

This study revealed several unexpected results, mainly the more fine-grained differentiation in the N400 rather than in the P350 effect, the direction of the asymmetry in the global data analyses, and the selective influence of tongue height on lexical activation. All three points will be considered separately in the following discussion.

3.5.3.1 N400 versus P350 Effect

In the study on place specification of word initial consonants by Friedrich and colleagues (Friedrich, Lahiri & Eulitz, 2008) (reported in the introduction to this experiment), the P350 effect differed from the N400 effect and could be distinguished from it as a separate neurophysiological deflection. The two components differed from each other with respect to latency, scalp topography, polarity of the elicited differences, and sensitivity to the experimental manipulation. While the P350 effect indeed reflected the fine-grained, asymmetric differences predicted by the FUL model, the N400 was sensitive only to more global distinctions of related versus unrelated prime and target words and did not show any asymmetries. Similar results were expected for the present experiment that investigated word-medial vowels rather than word-initial consonants and used fragments of real words as primes rather than pseudowords. However, the present results reveal more fine-grained and asymmetric effects in the N400 component and a more global differentiation between the three conditions in the P350 component. Visual inspection of the waveforms revealed a similar, but less pronounced asymmetric pattern in the P350 component as in the N400 component, but this failed to reach significance. Furthermore, there is no latency difference of the main effect of Condition between the P350 and the N400 effect. According to the 50ms time window analyses, both ROIs show the effect around 300 ms after target onset. Naturally, the polarity of the elicited differences is opposite in the P350 and the N400 effect, with most negative amplitudes for the identical condition in the P350 component, and most positive amplitudes for the same condition in the N400 component. However, it is hard to tell these two effects apart in the present experiment.

However, the topography of the P350 effect has changed over past studies, depending on what was investigated. In the first study that reported the P350 effect (Friedrich, Kotz, Friederici & Gunter, 2004), it was described as a reduced positivity for matching as compared to mismatching prime-target pairs in left posterior temporal regions. In a study that investigated the effects of match and mismatch in pitch between prime and target (Friedrich, Kotz, Friederici & Alter, 2004), the P350 effect was distributed over left and right posterior regions.

Segmental match and mismatch among the same word items revealed a left-lateralised distribution including anterior and posterior electrodes. If the target words deviated from their preceding primes in only one segment, amplitudes differed from control words only in the left anterior Region of Interest (Friedrich, 2005). Finally, in the study by Friedrich and colleagues (Friedrich, Lahiri & Eulitz,

2008), where the most fine-grained differences were investigated, the P350 topography was restricted to the left fronto-temporal region. This shows that the P350 effect differs in scalp topography and polarity depending on what is investigated. Since we do not assume major differences in the processing of vowels and consonants, we had expected to see a similar pattern as in the study by Friedrich and colleagues (2008). Both, theirs and our study investigated featural, asymmetric representations of place of articulation, once in consonants, and once in vowels. Possibly vowels are processed by slightly different cell assemblies, yielding different topographies than consonants or pitch.

3.5.3.2 Direction of the Asymmetry in the Global Data Analysis and Selective Impact of Height Information

Based on the FUL model and the hypothesis of lexical underspecification of the feature [CORONAL], we had expected that a target word with an unspecified coronal vowel (e.g. Tennis) is activated by its own first syllable (e.g. ten-) as well as by a related syllable with a dorsal vowel (e.g. tan-), because both primes should lead to a so-called nomismatch in the process of signal to lexicon mapping. On the contrary, a target word with a specified dorsal vowel (e.g. Tanne) should be activated only by its own first syllable (e.g. tan-), but not by a syllable with a mismatching coronal vowel (e.g. ten-). Consequently we expected similar amplitudes in the identical and the related condition for targets with coronal vowels, and differences in amplitudes between the identical and the related condition for targets with dorsal vowels. The results of the centro-parietal ROI showed exactly the opposite effect. According to these amplitude values and our previous interpretations of this effect, we needed to conclude that a target with a dorsal vowel is activated by both, a dorsal as well as a coronal vowel in the prime fragment, while a target with a coronal vowel is activated only by a coronal vowel in the prime fragment, but not by a dorsal one. This effect is predicted neither by the FUL model, nor by any other theory on language processing.

If we additionally consider the effect of height features, the picture gets more complex. A target with a dorsal vowel was not affected by the height information in prime and target. That is, amplitudes of the identical and the related condition did not differ from each other, no matter whether the vowel in the prime fragment conveyed the same height information as the vowel in the target word, or not. However, for targets with a coronal vowel, a difference in height between prime and target considerably suppressed the activation of the target word in the related condition. In other words, a target with a coronal vowel was activated by a

prime with a dorsal vowel as long as height information was the same, but it was not activated if there was a difference in height between prime and target vowel. If we stick to the hypothesis that coronality is not specified in the mental lexicon, it makes sense that the language processor uses other cues for lexical access and selection, as for example tongue height information, and it might need to rely on these cues more heavily than a fully specified entry does. However, this would imply that we needed to assign a weighted importance to features in lexical access and that these weights changed in accordance with the details of specification. Also note, that the differences in height information in this experiment did not include a full mismatch in height, i.e. there was no high vowel in the prime and a low vowel in the target or vice versa. The only contrasts were between high and mid vowels and low and mid vowels. Since the FUL model assumes that the feature [MID] is neither stored in the mental lexicon nor extracted as a feature from the speech signal, no effect on lexical access can be expected.

In order to make sure that the auditory stimulus material did not contain misleading acoustic information that could explain the unexpected outcome, we calculated and plotted the mean formant frequencies of F0 for pitch information, F1 for height information and F2 for place information for every single vowel in the auditory prime fragments. The formants of all vowels nicely fell into the expected categories, thus rendering an acoustic explanation unlikely. We also redid the statistical analyses for word frequency, cohort size and number of more frequent competitors separately for the subgroup with same height information in both pair words and for the subgroup with different height information. None of the analyses reached significance (p > .16), suggesting that the material was well controlled also throughout the subgroups.

Furthermore, we globally looked at ERP responses to word versus pseudoword targets to assure that the overall polarity of the N400 met the common expectations of a more negative N400 amplitude for pseudowords as compared to words. This was indeed the case, allowing for traditional interpretations of N400 amplitude effects.

Apart from using word rather than pseudoword fragments as primes and investigating vowels rather than consonants, another major difference between Friedrich et al. (2008) and our study is the timing in the experimental design.

Friedrich et al. presented the visual target word immediately at the offset of the auditory prime fragment, while we had inserted a 500ms pause between prime fragment and target word in order to have a silent baseline condition in the ERPs and to avoid prime repetitions. In word processing, 500ms is a long time span, and

consequently it could be the case that Friedrich et al. tapped into different processing stages than we did. With no pause between prime fragment and target word, the target taps into a stage of lexical activation where the cohort of possible word candidates has just been activated. In this stage words are activated that perfectly match the input signal, as well as words with coronal segments that do not mismatch with the pseudoword-fragments used. This leads to the asymmetric activation pattern observed by Friedrich et al. (2008). In our study there is a 500ms pause after the presentation of the prime fragments. This time may be sufficient for the mental lexicon to narrow down the cohort of activated words to the best matching word. In case of a fragment with a dorsal vowel (e.g. tan-), both target words, Tanne and Tennis will be initially activated, but as time passes, Tanne will be selected as the better match and Tennis will be removed from the cohort. If Tennis then appears as target word after tan-, it has to be reactivated again, which might have higher costs than activating a neutral target word that has not just been discarded. On the other hand, if ten- is perceived, it does not activate Tanne in the first place, and if Tanne then appears, it is treated in a neutral fashion. This could explain why Tennis is not facilitated after tan-, although it theoretically is a nomismatch. It does not, however, explain, why Tanne is as effectively activated by ten- as by tan-. In this light this explanation is not satisfactory but raises the importance to explore the effects of a pause between prime and target in cross-modal fragment priming.

3.5.3.3 Word- versus Pseudoword-Fragments as Primes

If it is true that the feature [MID] plays a role in lexical access of coronal vowels, we need to reconsider results obtained by Felder (2006). In a cross modal fragment priming study the impact of the height feature on coronal vowels was investigated.

The target word contained either the mid, coronal vowel [ɛ] (e.g. Becher), or the high, coronal vowel [ɪ] (e.g. Skizze). P350 results showed that both target-types were activated by their own first syllables (e.g. bech- Becher, skiz- Skizze) as well as by the syllables with changed height (e.g. *bich- Becher, *skez- Skizze). There was no difference in amplitude between these two conditions. The present study involved a place feature change in addition to the change in height, but still the differential effects of a height difference needs to be accounted for. One obvious difference between the two studies is the use of pseudoword-fragments in the study by Felder (2006) versus the use of fragments of existing words in the present study.

This might suggest that if no better matching alternative is available, a difference in height might be less severe than if competitors are activated. Similarly, we do

not know whether Friedrich et al. (2008) would have obtained their asymmetric priming results, had they used fragments of real words rather than pseudowords as deviating primes. This brings back to mind the studies by Gaskell and Marslen-Wilson (1996, 2001), who found facilitation for words with coronal final consonants (e.g. lean) that were primed by pseudowords with non-coronal consonants (e.g. *leam), but not for words with final coronal consonants (e.g. run) that were primed by other existing words with non-coronal consonants (e.g. rum).

This could suggest that the brain always goes for the best matching alternative (e.g.

rum rather than run after hearing rum) and rules out other less likely possibilities, except for cases where discarding that item would leave it with no possible word candidate (e.g. lean after hearing *leam). In other words, the brain sorts out more strictly in the face of choice (e.g. rum or run) than in the face of no choice (e.g.

lean or ?).