The many places of frequency: Evidence for a novel locus of the lexical frequency effect in word production

Abstract

The effect of lexical frequency on language-processing tasks is exceptionally reliable. For example, pictures with higher frequency names are named faster and more accurately than those with lower frequency names. Experiments with normal participants and patients strongly suggest that this production effect arises at the level of lexical access. Further work has suggested that within lexical access this effect arises at the level of lexical representations. Here we present patient E.C. who shows an effect of lexical frequency on his nonword error rate. The best explanation of his performance is that there is an additional locus of frequency at the interface of lexical and segmental representational levels. We confirm this hypothesis by showing that only computational models with frequency at this new locus can produce a similar error pattern to that of patient E.C. Finally, in an analysis of a large group of Italian patients, we show that there exist patients who replicate E.C.'s pattern of results and others who show the complementary pattern of frequency effects on semantic error rates. Our results combined with previous findings suggest that frequency plays a role throughout the process of lexical access.

Acknowledgments

We are especially grateful to E.C. for his participation in this study. We would like to thank Wheeler Ruml, Gary Dell, Matthew Goldrick, and an anonymous reviewer for their helpful comments and suggestions. We would also like to thank Jay Fournier and Janna Fucillo for help in testing E.C., Laruen Moo for interpreting E.C.'s MRI scans, and Gabriele Miceli and Rita Capasso for collecting the data on the Italian patients. In addition, we are grateful to Gary Dell for providing us the source code for his implementation of his model, and to Brenda Rapp and Matthew Goldrick for sharing a detailed specification of their modelling neighbourhood. The research reported here was supported in part by NIH Grant DC04542 to Alfonso Caramazza. Mark Knobel was supported by the Sackler Scholars Programme in Psychobiology.

Notes

¹ At a finer grain, Dell (1990) and Jescheniak and Levelt (1994) came to different conclusions about the locus of frequency effects within word production. Dell put the frequency effect at the lemma, or syntactic form, level, while Jescheniak and Levelt put it at the lexeme, or phonological form, level. These distinctions within the lexical level are not universally accepted (see Caramazza, 1997; Caramazza & Miozzo, 1997). We discuss the issue of multiple lexical levels in our General Discussion.

² A logistic regression was conducted with log frequency, length in letters, imageability, and noun–verb status. The overall model was significant, χ²(4) = 37.55, p < .001, Nagelkerke R ² = .199. Log frequency (Wald = 7.27, p = .007), length in letters (Wald = 13.84, p < .001), and imageability (Wald = 6.85, p = .009) were the significant individual factors. His repetition data were analysed using a logistic regression with the same four factors. This time the overall model was also significant, χ²(4) = 9.54, p = .049, Nagelkerke R ² = .056. The only significant factor was length in letters (Wald = 6.28, p = .012).

³ Regularization errors (e.g., [hϵrθ] for “hearth”) and morphological errors were also common in word reading, while lexicalization errors were common in nonword reading.

⁴ Frequency was also found to affect mixed errors, but not phonological (word) errors. Since there were few errors in each of these categories (2.4% for phonological; 2.5% for mixed) we do not discuss these findings further.

⁵ Our implementation of Dell et al.'s (1997) model uses George Marsaglia and Wai Wan Tsang's (2000) ziggurat algorithm for generating normally distributed random numbers. We thank them for making their code publicly available.

⁶ We would like to thank Wheeler Ruml for providing us with this neighbourhood and for creating the algorithm by which it was constructed.

⁷ We wish to stress the difference between theory and implementation, since it is possible to implement the same theory in different ways. For example, see Rapp and Goldrick (2000) for other implementations of the global damage and levels of damage hypotheses.

⁸ Both our mapping procedure and that of Foygel and Dell (2000) used a smart procedure to only add points to the parameter map if they were significantly different from their surrounding neighbours (based on a χ² criterion). The slight difference between our procedure and theirs is that in ours the decision to add each new point is done on a point-by-point basis and not for groups of five points. Our maps allowed connection strength to vary from 0.001 to 0.1 and decay to vary from 0.5 to 1.0. The maximum level of expansion (i.e., grain) that we allowed was 1/128. We created a map for each of the nine models (i.e., three levels of interactivity crossed with three neighbourhoods) with each of the two implementations of damage, for a total of 18 maps. In testing our procedure, we obtained fits for their patients that were comparable to published results.

⁹ Another possible locus of frequency within the model is in the jolt strength. Higher frequency lexical nodes can receive a higher jolt of activation after they are selected. The results of such an implementation should be quite similar to the lexical-segmental connection locus that we do implement, since both have frequency effects occurring mostly after lexical selection. Please see our General Discussion for our reasons for not manipulating jolt strength in the model.

¹⁰ The Huynh–Feldt correction was applied because the sphericity assumption was not met.

¹¹ The Huynh–Feldt correction was applied because the sphericity assumption was not met.

¹² A potential challenge to this conclusion comes from a proposal by Dell (1990). He proposed, but never implemented, the hypothesis that lexical frequency effects in production could be explained as semantic and/or phonological neighbourhood effects. This hypothesis challenges an underlying assumption of our work, which is that the semantic and phonological neighbourhoods of a higher frequency word are the same as those neighbourhoods of a lower frequency word. This assumption is probably incorrect, at least for phonological neighbourhoods, given the significant correlation (r = −.49, p < .01) between frequency and phonological neighbourhood size in our own data (see also Gordon, 2002). Nevertheless, while phonological neighbourhood size affects E.C.'s rates of phonological and mixed errors, significant effects of lexical frequency persist in regression analyses of E.C.'s performance, despite the inclusion of phonological neighbourhood size and other semantic and phonological measures as cofactors. Putting these analyses aside, we explored the possibility further by trying to simulate E.C.'s lexical frequency effects through modifying semantic and phonological neighbourhoods in the largest neighbourhood (that of Rapp & Goldrick, 2000). Adding more semantic neighbours led to the models (H, R, and C) producing more semantic and nonword errors, with a much larger increase in semantic errors. Overall accuracy also decreased. This pattern was quite unlike E.C.'s. In the case of adding more phonological neighbours, models produced slightly fewer nonword and semantic errors, but instead of leading to fewer errors as a function of frequency (as in E.C.'s case), this decrease led to a corresponding increase in phonological errors (see also Dell & Gordon, 2003). It is clear from these investigations that semantic and phonological neighbourhoods cannot explain the lexical frequency effects that we observe in E.C.'s performance.

¹³ We were unable to obtain enough items with frequency ratings for patients A.S. and C.L.B.

¹⁴ Given Italian's shallow orthography, letter length is very close to the phoneme length.

¹⁵ Unfortunately we were unable to get norms for enough of our items on semantic or other phonological factors to merit an analysis including them. Since we did not include imageability and phonological neighbourhood size into our analyses of the Italian patients, any frequency effects that we describe here could potentially be effects of imageability and/or phonological neighbourhood size, even though those factors were not responsible for the frequency effects observed in patient E.C.