<?xml version="1.0" standalone="yes"?>
<Paper uid="P99-1056">
  <Title>The grapho-phonological system of written French: Statistical analysis and empirical validation</Title>
  <Section position="5" start_page="439" end_page="441" type="metho">
    <SectionTitle>
2.2. Results
</SectionTitle>
    <Paragraph position="0"> Data associated with inappropriate triggering of the microphone were discarded from the error analyses. In addition, for the response time analyses, pronunciation errors, hesitations, and anticipations in the delayed naming task were eliminated. Latencies outside an interval of two standard deviations above and below the mean by subject and condition were replaced by the corresponding mean. Average reaction times and error rates were then computed by subjects and by items in both the immediate naming and the delayed naming task. By-subjects and by-items (Ft and F2, respectively) analyses of variance were performed with grapheme frequency and grapheme entropy as within-subject factors.</Paragraph>
    <Paragraph position="1"> Grapheme frequency. For naming latencies, pseudowords of low grapheme frequency were read 24 msec more slowly than pseudowords of high grapheme frequency. This difference was highly significant both by subjects and by items; Fj(1, 19) = 24.4, p &lt; .001, Fe(1, 31) = 7.5, p &lt; .001. On delayed naming times, the same comparison gave a nonsignificant difference of-7 msec. For pronunciation errors, there was no significant difference in the immediate naming task. In the delayed naming task, pseudowords of low mean grapheme frequency caused 1.2% more errors than high ones. This difference was marginally significant by items, but not significant by subjects; F2(1, 31) = 3.1,p &lt; .1.</Paragraph>
    <Paragraph position="2"> Grapheme entropy. In the immediate naming task, high-entropy pseudowords were read 48 msec slower than low-entropy pseudowords; FI(1,</Paragraph>
    <Paragraph position="4"> the delayed naming task, the same comparison showed a significant difference of 27 msec; FI(1, 19) = 22.9 p &lt; .001, F2(1, 31) = 12.5, p &lt; .005.</Paragraph>
    <Paragraph position="5"> Because of this articulatory effect, delta scores were computed by subtracting delayed naming times from immediate naming times. A significant difference of 21 msec was found on delta scores; FI(1, 19) = 5.7,p &lt; .05, F2(1, 31) = 4.7,p &lt; .05.</Paragraph>
    <Paragraph position="6"> The pattern of results was similar for errors. In the immediate naming task, high-entropy pseudowords caused 5% more errors than low-entropy pseudowords. This effect was significant by subjects but not by items; Ft(1, 19) = 7.4, p &lt; .05, F2(1, 31) = 2.1,p &gt; .1. The effect was of 6.5% in the delayed naming task and was significant by subjects and items; FI(1, 19) = 17.2, p &lt; .001, F2(1, 31) = 8.3,p &lt; .01.</Paragraph>
    <Section position="1" start_page="439" end_page="441" type="sub_section">
      <SectionTitle>
2.3. Discussion
</SectionTitle>
      <Paragraph position="0"> A clear effect of the grapheme frequency and the grapheme entropy manipulations were obtained on immediate naming latencies. In both manipulations, the stimuli in the contrasted lists were selected pairwise to be as equivalent as possible in terms of potentially important variables.</Paragraph>
      <Paragraph position="1"> A difference between high and low-entropy pseudowords was also observed in the delayed naming condition. The latter effect is probably due to phonetic characteristics of the initial consonants in the stimuli. Some evidence confirming this interpretation is adduced from a further control experiment in which participants were required to repeat the same stimuli presented auditorily, after a variable response delay. The 27 msec difference in the visual delayed naming condition was tightly reproduced with auditory stimuli, indicating that the effect in the delayed naming condition is unrelated to print-to-sound conversion processes. Despite this unexpected bias, however, when the influence of phonetic factors was eliminated by computing the difference between immediate and delayed naming, a significant effect of 21 msec remained, demonstrating that entropy affects grapheme-phoneme conversion.</Paragraph>
      <Paragraph position="2"> These findings are incompatible with current implementations of the dual-route theory (Coltheart et aL, 1993). The &amp;quot;central dogma&amp;quot; of this theory is that the performance of human subjects on pseudowords is accounted for by an analytic process based on grapheme-phoneme conversion rules. Both findings are at odds with the additional core assumptions that (1) only  dominant mappings are retained as conversion rules; (2) there is no place for ambiguity or predictability in the conversion.</Paragraph>
      <Paragraph position="3"> In a recent paper, Rastle and Coltheart (1999) note that &amp;quot;One refinement of dual-route modeling that goes beyond DRC in its current form is the idea that different GPC rules might have different strengths, with the strength of the correspondence being a function'of, for example, the proportion of words in which the correspondence occurs.</Paragraph>
      <Paragraph position="4"> Although simple to implement, we have not explored the notion of rule strength in the DRC model because we are not aware of any work which demonstrates that any kind of rule-strength variable has effects on naming latencies when other variables known to affect such latencies such as neighborhood size (e.g., Andrews, 1992) and string length (e.g., Weekes, 1997) are controlled.&amp;quot; We believe that the present results provide the evidence that was called for and should incite dual-route modelers to abandon the idea of all-or-none rules which was a central theoretical assumption of these models compared to connectionist ones. As the DRC model is largely based on the interactive activation principles, the most natural way to account for graded effects of grapheme frequency and pronunciation predictability would be to introduce grapheme and phoneme units in the nonlexical system.</Paragraph>
      <Paragraph position="5"> Variations in the activation resting level of grapheme detectors as a function of frequency of occurrence and differences in the strength of the connections between graphemes and phonemes as a function of association probability would then explain grapheme frequency and grapheme entropy effects. However an implementation of rule-strength in the conversion system of the kind suggested considerably modifies its processing mechanism, notably by replacing the serial table look-up selection of graphemes by a parallel activation process. Such a change is highly likely to induce non-trivial consequences on predicted performance.</Paragraph>
      <Paragraph position="6"> Furthermore, and contrary to the suggestion that the introduction of rule-strength would amount to a mere implementational adaptation of no theoretical importance, we consider that it would impose a substantial restatement of the theory, because it violates the core assumption of the approach, namely, that language users induce all-or-none rules from the language to which they are exposed. Hence, the cost of such a (potential) improvement in descriptive adequacy is the loss of explanatory value from a psycholinguistic perspective. As Seidenberg stated, &amp;quot;\[we are\] not claiming that data of the sort presented \[here\] cannot in principle be accommodated within a dual route type of model. In the absence of any constraints on the introduction of new pathways or recognition processes, models in the dual route framework can always be adapted to fit the empirical data. Although specific proposals might be refuted on the basis of empirical data, the general approach cannot.&amp;quot; (Seidenberg, 1985, p. 244).</Paragraph>
      <Paragraph position="7"> The difficulty to account for the present findings within the dual-route approach contrasts with the straigthforward explanation they receive in the PDP framework. As has often been emphasized, rule-strength effects emerge as a natural consequence of learning and processing mechanisms in parallel distributed systems (see Van Orden, Pennington, &amp; Stone, 1990; Plaut et al., 1996). In this framework, the rule-governed behavior is explained by the gradual encoding of the statistical structure that governs the mapping between orthography and phonology.</Paragraph>
      <Paragraph position="8"> Conclusions In this paper, we presented a semi-automatic procedure to segment words into graphemes and tabulate grapheme-phoneme mappings characteristics for the French writing system. In current work, the same method has been applied on French and English materials, allowing to provide more detailed descriptions of the similarities and differences between the two languages. Most previous work in French (e.g.</Paragraph>
      <Paragraph position="9"> Vrronis, 1986) and English (Venezky, 1970) has focused mainly on the extraction of a rule set. One important feature of our endeavor is the extraction of several quantitative graded measures of grapheme-phoneme mappings (see also Bern&amp;, Reggia, &amp; Mitchum, 1987, for similar work in American English).</Paragraph>
      <Paragraph position="10"> In the empirical investigation, we have shown how the descriptive data could be used to probe human readers' written word processing. The results demonstrate that the descriptive statistics  capture some important features of the processing system and thus provide an empirical validation of the approach. Most interestingly, the sensitivity of human processing to the degree of regularity and frequency of grapheme-phoneme associations provides a new argument in favor of models in which knowledge of print-to-sound mapping is based on a large set of graded associations rather than on correspondence rules.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="441" end_page="441" type="metho">
    <SectionTitle>
Acknowledgements
</SectionTitle>
    <Paragraph position="0"> This research was supported by a research grant from the Direction Grn6rale de la Recherche Scientifique -- Communaut6 fran~aise de Belgique (ARC 96/01-203). Marielle Lange is a research assistant at the Belgian National Fund for</Paragraph>
  </Section>
class="xml-element"></Paper>