A tale of two frequencies: Determining the speed of lexical access for Mandarin Chinese and English compounds

Abstract

Two picture naming experiments show that compound word production in Mandarin Chinese and in English is determined by the compound's whole-word frequency, and not by its constituent morpheme frequency. Four control experiments rule out that these results are caused by recognition or articulatory processes. These results are consistent with models of lexical access that assume compounds are stored in their full-form and that frequency affects the retrieval of whole words. The present results corroborate the results from previous studies that have investigated compound word production in Mandarin Chinese, but also differ from those previously reported on compound word production in Dutch. The possibility that this inconsistency arises due to cross-linguistic, or task differences is discussed.

Keywords:

• Compounds
• Frequency
• Language production
• Morphology
• Picture naming

Acknowledgements

The research reported here was supported by NIH Grant DC04542 to Alfonso Caramazza. Experiment 1a and 1b were presented at the International Conference of Cognitive Sciences in Beijing, China, August, 2001 by Yanchao Bi. Experiment 2a and 2b were presented at the Trends in Cognitive Psychology Symposium, in Naples, Italy, March, 2004 by Niels Janssen. Niels Janssen was supported by the Sackler Scholar Programme in Psychobiology and a post-doctoral fellowship from the Fyssen Foundation. We thank Jorge Almeida, Raymond Bertram, and one anonymous reviewer for helpful comments and discussion.

Notes

¹Note that the reverse is not true. That is, evidence against the decomposition hypothesis of compounds would not constitute evidence against the lemma/lexeme distinction. This is because the full-form representation hypothesis can be implemented in a lexical access system that distinguishes between lemma and lexeme lexical layers by assuming that compounds are represented as whole words at both levels of representation. That is, lexical models that assume a distinction between lemma and lexeme levels are compatible with both the decomposition and the full-form hypotheses of compounds.

²Note that with additional assumptions about incremental language production different predictions could be made about the size of frequency effects for the first and second morpheme. However, independent of these assumptions, the model predicts that RTs should decrease with increasing average morpheme frequency.

³It has been demonstrated that frequency and age of acquisition (AoA) are highly correlated (see Carroll & White, 1973b; Ellis & Morrison, 1998; Morrison, Ellis, & Quinlan, 1992). We did not control for the AoA factor in our experiments. Therefore, it is possible that any effects we attribute to the factor frequency might in fact be due to AoA. However, the discussion concerning the locus of the frequency effect in lexical access has been couched in terms of the factor frequency without controlling for AoA. It is not the purpose of this study to distinguish between these factors. Future experiments are required to disentangle which of these two factors is responsible for producing the effects discussed here.

⁴We acknowledge the potential importance of the role of semantic transparency in the representation of compounds (e.g., Libben et al., 2003 in word comprehension). However, the language production models under consideration here do not make assumptions about a role of semantic transparency in compound representation, and therefore, the predictions we derive from the models under consideration in this paper do not take into account the influence of this variable.

⁵Collinearity statistics on excluded variables revealed tolerance values of .95 for frequency of the first noun, .95 for frequency of the second noun, .92 for the mean frequency of the two constituents, .98 for the minimum frequency of the two constituents, and .90 for the maximum frequency of the two constituents. The largest condition index was 3.31 suggesting no problems of collinearity were present in the current data set.

⁶As acknowledged earlier, in Experiment 1b there are several items for which the English translation is also a compound word. To investigate whether the inclusion of these compound words influenced the pattern of results obtained in Experiment 1b, we reanalysed the results excluding those words that are also compounds in English (see Appendix A for a list). Exactly the same results were obtained in Experiment 1a and 1b when we excluded the 13 items that translated into compounds in Experiment 1b.

⁷A regression analysis on the compound pictures from this experiment (N = 42) with familiarity rating as a dependent variable revealed that the log transformed frequency of the first and second constituent were not significant predictors. Hence, the familiarity ratings collected for all compound pictures were not influenced by the compound's constituent frequency.

⁸Collinearity statistics on excluded variables revealed tolerance values of .93 for frequency of the first noun, .93 for the mean frequency of the two constituents, .93 for the minimum frequency of the two constituents, and .99 for the maximum frequency of the two constituents. The largest condition index was 4.88, suggesting no problems of collinearity were present in the current data set.

⁹These results do not rule out the class of models with a lemma/lexeme distinction as a whole. If such models were to assume that compounds are stored in their full form at the lexeme level, or if they were to assume that the frequency variable affects lemma retrieval (with decomposed representations at the lexeme level), they would predict the results obtained here.