Rethinking the FUL Model’s Assumption of Ternary Matching Logic

lexical activation, whereas a feature that mismatches with the lexical entry removes the word from the cohort of possible items. If a feature is in the signal, but the lexical entry is underspecified for it, or if there is nothing extracted from the signal but something specified in the mental lexicon, this leads to a nomismatch. A nomismatch does not remove an item from the cohort of activated items, neither does it enhance the word’s activation level.

In the introduction several studies have been described which provide results that are in line with the model’s assumption of underspecification (Lahiri & Marslen-Wilson, 1991;

Lahiri & van Coillie, 1999; Eulitz & Lahiri, 2003; Eulitz & Lahiri, 2004; Friedrich, Lahiri &

Eulitz, submitted). At the same time these results support the notion of a ternary matching logic. However the exact nature of this ternary matching logic has not been examined yet.

What does it mean for an item to be removed from the cohort as soon as a mismatch occurs?

Is it actively inhibited or does it simply fall back into a status of no activation? Alternatively a mismatch could be punished with a reduction in activation, leaving the item still activated a little bit. Similarly, a match has to be treated differently from a nomismatch in terms of activation. This can happen in two ways. Either, as the FUL model claims, in that the match leads to more activation whereas the nomismatch does not change the activation level; or both lead to more activation, differing in the amount of power added to an item. Table 4.8 depicts four possible scenarios of a ternary matching logic.

Table 4.8: Four possible scenarios according to which match, nomismatch and mismatch could affect the activation level of a word in the mental lexicon. A↑ = increase activation level; 0 = do not change activation; A↓ = decrease activation level but do not remove; X = remove item from cohort; X- = remove and inhibit item;

Changes in Lexical Activation

FUL model Alternative 1 Alternative 2 Alternative 3

Match A ↑ A ↑ A ↑ A ↑↑

Nomismatch 0 0 0 A ↑

Mismatch X X- A ↓ 0

Out of these four scenarios of signal-to-lexicon-mapping different predictions concerning the outcome of priming experiments arise. The FUL model predicts positive priming effects for the match and the nomismatch conditions. Differences in priming between these two conditions will probably depend on factors like the amount of matching versus nomismatching features and the size of the cohort activated by a prime, determining the number of potential competitors that match the input equally well or better than the target

word. A mismatch condition should not prime the target word since this should have been removed from the active cohort. At this point the question arises of what happens to a word that is removed from the cohort. How quickly does it lose its activation? Does it fall back into a status of no activation, comparable to the one before the priming situation, or is it actively inhibited? Several possible scenarios emerge:

• If activation is simply removed from the item without inhibiting it, then the mismatching condition should reveal results that are comparable to those of the control condition.

• If, however, inhibition occurs after a mismatch (Alternative 1), then the mismatch condition ought to reveal even worse results than the control condition in terms of enhanced reaction time or difference in ERP amplitude.

• Alternative 2 predicts that a mismatch will elicit the weakest priming effects of the three conditions, but will still differ in a positive direction from the control condition.

This scenario might be difficult to distinguish from the FUL model, depending on how quickly an item loses its activation when thrown out of the cohort of active competitors. Paradigms employing delayed priming will be more useful for this purpose of investigation than direct fragment priming designs.

• Also Alternative 3, the last one mentioned here, would yield patterns of results similar to those of Alternative 2. It is maybe the least intuitive one since it is not quite logic that a nomismatch contributes additional activation to an item. Even after a mismatch between prime and target considerable priming would be observed.

Several behavioural studies have confirmed assumptions of abstract representations of words in the mental lexicon. However, they are inconclusive on the behaviour of the mismatch condition as compared to the unrelated control condition. The relationship between these two conditions is important for interpreting the meaning of being removed from a cohort. In the experiment by Lahiri and van Coillie (Lahiri & van Coillie, 1999) using a crossmodal lexical decision task with immediate semantic priming, the mismatching condition did not differ significantly from the control condition. Real words ending in a coronal segment (e.g. Bahn) primed semantically related words (e.g. Zug) as well as pseudowords did which differed from the real words in that the last coronal segment was replaced by a noncoronal one (e.g.

not differ from those in the control condition. These results are in line with the assumption of the FUL model, that a mismatching item is simply removed completely from the cohort.

On the contrary, using a crossmodal repetition priming experiment, Wheeldon and Waksler (2004) found significant priming for their mismatch condition. Although reaction time was significantly increased for the mismatch condition in comparison to the match and nomismatch condition, it was still facilitated with regard to the control condition. The experimental design differs in several important aspects from the design described by Lahiri and van Coillie (1999). It did not employ semantic priming but repetition priming. The critical words were embedded in auditorily presented sentences. At the offset of the critical prime word, a visual lexical decision target appeared on the screen. The prime word was either unchanged or changed in the place of articulation of its last segment. The target word was the unchanged form of the word. For the control condition completely unrelated prime words were used. Predictions were again that pseudowords with noncoronal endings activated the corresponding words with coronal endings in the mental lexicon, while coronal ending pseudowords do not activate target words ending in noncoronals. These asymmetric predictions were supported. In contrast to the experiment by Lahiri and van Coillie, there was a lot of segmental overlap between prime and target this time because repetition priming was used. This could be taken to explain the priming effects even for conditions with segmental mismatch in the last segment of a word. Still, if an item in the cohort was removed or actively inhibited after the occurrence of a mismatch, no priming should show up, despite considerable segmental overlap. In addition, it would be important to know whether, and if how many competitors were still active at the occurrence of the mismatch. Maybe the brain is reluctant to throw away the last surviving item in the recognition process. The FUL model could account for these results if one assumes that lexical deactivation takes some time.

Besides, the patterns described as Alternative 2 and 3 would also predict such a result.

In order to complete the picture, among others Soto-Faraco and colleagues (Soto-Faraco, Sebastián-Gallés & Cutler, 2001) report inhibition of priming effects in the case of a mismatch. In a crossmodal fragment priming experiment, they first presented disyllabic auditory prime fragments immediately followed by visual target words. A fragment was either identical (e.g. aban-) to the first two syllables of the target word (abandono,

“abandonment”), or differed in the nucleus of its second syllable from that of the target (abun- from abundancia, “abundance”), or was completely unrelated to it (elas- from elastico, “elastic”). Responses to matching targets were significantly faster than responses to control conditions, whereas responses to mismatching targets were significantly slower than

responses to control conditions. These results partially contradict those of the two former experiments. Again, the design differed, in that it was neither semantic nor repetition priming but fragment priming. Another important difference lies in the use of pseudowords as primes in the two former experiments, as opposed to fragments of real words in this study.

It is of major importance to examine the effects that different experimental designs, the size of the activated cohort, the use of pseudowords as compared to real words as primes, the amount of matching versus nomismatching features, auditory versus visual presentation of prime and target words and other factors not discussed here, have on the activation, selection and rejection of items in the mental lexicon. This is not only important for better understanding the processes underlying speech recognition but also for allowing precise hypotheses and for being able to adequately discuss seemingly contradictory findings to the FUL model.

The above mentioned studies are inconclusive on the nature of the ternary system of match, mismatch and nomismatch, but do not contradict the FUL model’s claim of underspecification. In an attempt to rule out the possibility of lexical underspecification, Gow (2001) reports, that in a crossmodal repetition priming experiment with a lexical decision task he obtained priming using phonologically implausible primes. For instance, mispronounced bloom in the sentence The first [blun] dies off with a late frost speeded reaction times for the visual target BLOOM as compared to the unrelated target CLAIM. The FUL model assumes labial to be specified in the mental lexicon and hence [blun] should not prime the underlying form of bloom. Moreover, reaction times did not differ between the phonologically implausible condition and the phonologically plausible condition in which a word-final coronal was replaced by a labial or dorsal. First of all it has to be noted that one usually varies the prime word rather than the target word between experimental and control conditions, in order to calculate reaction time for the same words across groups. Besides, the experimental setup resembles very closely the one of the study by Wheeldon and Waksler (2004). While they found significant differences in reaction time between an assumed nomismatch and a mismatch, Gow reports equal priming effects for both. As primes are embedded in sentences, care has to be taken to avoid context effects. The embedded prime word should not be anticipated by the preceding sentence context. Wheeldon and Waksler

regard to its nature (e.g.: “The coverage of the terrible crime drew the wrath of the critics.”

“To stay warm they had to burn poles (…)”. Gow, 2001, p.156)).

In all studies reported so far, the prime word always turned into a pseudoword when the last segment was changed for place of articulation. This implies that no lexical competition has taken place. Gow (2002) conducted a priming experiment very similar in nature to the one above, but this time the change in the last segment of the prime word turned this prime into another, already existing word. His intention was to test whether coronal place assimilation creates lexical ambiguity as the FUL model would predict. He embedded the prime items in sentence contexts that could induce labialisation of the final segment (e.g.

“This time she tried to get the right berries for her pie.”) After hearing the prime (right) a target word appeared on a screen. This was either the original form RIGHT, the equally possible assimilated form RIPE or unrelated items like HEM and HEN. Participants hearing the assimilated token of the word right that perceptually approximated ripe selectively accessed right but not ripe. That means, listeners accessed only the intended word. The FUL model predicted that ripe activates both forms, right and ripe.

Gow (2002) proposes that assimilation produces a range of modification, rather than discrete feature change. According to him, assimilated items show a pattern of formant movement that is intermediate between the patterns associated with unmodified coronal place and the non-coronal place. With sufficient information about coronality left in the signal no lexical ambiguity would arise. However, leaving the listener with sufficient information about the nonassimilated value of the place feature, would mean that coronality is extracted and hence will mismatch with the labial specification of ripe and thus the FUL model could again explain these results.

Clearly, more research has to be conducted on all the variables that could affect the mapping process from the speech signal to the mental lexicon and the competition between items in the cohort. The present study does not contribute to this discussion because no mismatch condition was included. Other P350 studies using a crossmodal fragment priming design and including a mismatch as well as a control condition revealed significant differences between all three conditions, match, mismatch and control (Friedrich, Lahiri &

Eulitz, submitted; Friedrich, 2005). The mismatching condition showed less negative amplitudes than the matching conditions but more negative amplitudes than the control conditions, arguing for some residual activation of items despite of a mismatch.