Why learning to read is easier in Welsh than in English: Orthographic transparency effects evinced with frequency-matched tests

This study compared the rate of literacy acquisition in orthographically transparent Welsh and orthographically opaque English using reading tests that were equated for frequency of written exposure. Year 2 English-educated monolingual children were compared with Welsh-educated bilingual children, matched for reading instruction, background, locale, and math ability. Welsh children were able to read aloud accurately significantly more of their language (61% of tokens, 1821 types) than were English children (52% tokens, 716 types), allowing them to read aloud beyond their comprehension levels (168 vs. 116%, respectively). Various observations suggested that Welsh readers were more reliant on an alphabetic decoding strategy: word length determined 70% of reading latency in Welsh but only 22% in English, and Welsh reading errors tended to be nonword mispronunciations, whereas English children made more real word substitutions and null attempts. These findings demonstrate that the orthographic transparency of a language can have a profound effect on the rate of acquisition and style of reading adopted by its speakers. Systems that are noisy and inconsistent are harder to sort out than systems that are reliable and categorical. There is now a large body of research demonstrating that greater ambiguity in the mappings between the forms and functions of a particular language causes less successful learning because of a larger degree of competition among the cues in the learning set (Bates & MacWhinney, 1987; MacWhinney, 1987). Constructivist, emergentist, and connectionist perspectives on language acquisition emphasize that human rulelike processing of the structural regularities of language emerges from learners’ lifetime analysis of the distributional characteristics of the language input (MacWhinney, 1999). Some cues are more reliable than others, and the language learner’s task is to work out the most valid predictors. The Competition Model shows how Bayesian cue use can resolve in the activation of a single interpretative hypothesis from an interaction of cues, which vary in their frequency, reliability, and validity (MacWhinney, 1997).  2001 Cambridge University Press 0142-7164/01 $9.50 Applied Psycholinguistics 22:4 572 Ellis & Hooper: Reading in Welsh and English These processes pervade all domains of language acquisition. Consider, for example, how the acquisition of grammatical gender is determined by the degree of transparency of its morphophonological marking. Brooks, Braine, Catalano, Brody, and Sudhalter (1993) demonstrated that children and adults showed better learning of the noun subclasses of artificial languages when ambiguity was reduced by there being a subset of nouns of each subclass that shared a phonological feature than in a condition in which the phonological features were less reliable cues in distinguishing the subclasses. Taraban and Roark (1996) manipulated the ambiguity in the mapping of noun forms onto genders in two sets of French nouns and showed that learning the same set of feminine nouns took longer if the nouns in the masculine class were, as a set, more ambiguous in the mappings of their noun endings onto gender. This demonstration is important because it illustrates how the presence of nontransparent marking not only affects the speed at which the nontransparent items themselves are acquired but also slows the learning of the whole system. Recent studies have simulated language-learning data using simple connectionist models that relate cues and their functional interpretations. For example, the simulations of Kempe and MacWhinney (1998) showed why the Russian case inflection system is acquired more rapidly than is that of German: even though case marking in Russian is more complex than in German, the Russian inflections are more reliable cues to sentence interpretation. One area in which the effect of consistency of mapping has been extensively researched is that of relating symbols and their sounds in reading aloud. To the extent that readers are able to construct the correct pronunciations of novel words or nonwords, they must be able to apply sublexical rules or mappings that relate graphemes and phonemes (Coltheart, Curtis, Atkins, & Haller, 1993; Patterson & Morton, 1985) or larger orthographic units and their corresponding rimes or syllables (Ehri, 1998; Glushko, 1979; Goswami, 1999; Treiman, Mullennix, Bijeljac-Babic, & Richmond-Welty, 1995), and it is likely that it is the operation of this system that explains why regular or consistent words are read better than irregular or inconsistent words. For the case of adult fluency in English, words with regular spelling–sound correspondences (such as mint) are read with shorter naming latencies and lower error rates than words with exceptional correspondences (such as pint; Coltheart, 1978); in development, the reading of exception words (blood, bouquet) is acquired later than that of regular words (bed, brandy; Coltheart & Leahy, 1996). Similarly, in fluent performance, words that are consistent in their pronunciation in terms of whether this agrees with those of their neighbors with similar orthographic body and phonological rime (best is regular and consistent in that all -est bodies are pronounced in the same way) are named faster than inconsistent items (mint is regular in terms of its GPC rule, but inconsistent in that it has pint as a neighbor) (Glushko, 1979). The magnitude of the consistency effect for any word depends on the summed frequency of its “friends” (similar spelling pattern and similar pronunciation) in relation to that of its “enemies” (similar spelling pattern but dissimilar pronunciation) (Jared, McRae, & Seidenberg, 1990). Adult naming latency decreases monotonically with increasing consistency on this measure Applied Psycholinguistics 22:4 573 Ellis & Hooper: Reading in Welsh and English (Taraban & McClelland, 1987). In development, Laxon, Masterson, and Coltheart (1991) showed that within regular words, consistent (pink, all -ink) and consensus (hint, mostly as in mint, but cf. pint) items are acquired earlier than ambiguous ones (cove versus love, move), and that within irregular words, those in deviant gangs (like look, cold, and calm) are acquired earlier than ambiguous ones (love). According to the power law of learning that relates reaction time to amount of exposure, performance converges asymptotically at high levels of practice, and thus these effects of regularity and consistency are more evident with low-frequency words than with high-frequency ones (Seidenberg, Waters, Barnes, & Tanenhaus, 1984). As with the learning of other quasiregular language domains, these effects of consistency and ambiguity of spelling–sound correspondence within language have been successfully simulated in connectionist (Harm & Seidenberg, 1999; Plaut, McClelland, Seidenberg, & Patterson, 1996; Seidenberg & McClelland 1989; Zorzi, Houghton, & Butterworth, 1998) and exemplar-based (Ellis & Hicks, 2000) computational models. These investigations have compared the learnability and processing of words of different degrees of spelling–sound ambiguity within a language. What about the large cross-linguistic issue: what are the effects of the overall ambiguity of a language’s symbol–sound mappings on its speakers’ rate of literacy acquisition? The orthographies of languages such as Finnish, Italian, Spanish, Dutch, Turkish, and German are, on the whole, much more transparent than those of opaque languages such as English and French. In transparent orthographies, the mappings from letters to sounds are consistent. In opaque orthographies, the same grapheme many represent different phonemes in different words, and, as just illustrated for English, there are many words that are irregular in terms of the default grapheme–phoneme rules. It seems likely that these language differences in overall orthographic transparency have a determining effect on rate of reading acquisition, segmental phonological awareness, reading strategy, and reading disorder. We will briefly consider each in turn. Rate of reading acquisition Theories of reading acquisition in alphabetic languages commonly hold that there is a prolonged alphabetic stage of reading in which words are decoded on the basis of learned symbol–sound associations, and that this provides the practice that allows for the eventual development of skilled orthographic reading abilities (e.g., Ehri, 1979, 1998; Frith, 1985; Goswami, 2000; Marsh, Friedman, Welch, & Desberg, 1981). An orthographic transparency hypothesis therefore predicts that children learning to read a transparent orthography, in which sound–symbol mappings are regular and consistent, should learn to read and spell faster than those learning an opaque orthography, in which the cues to pronunciation are more ambiguous. Empirical research supports this prediction; for example, children learning to read German were more able to read their transparent orthography instantiation of pairs of translation equivalents (e.g., Pflug–plough) than were matched learners of English (Landerl, Wimmer, & Frith, 1997); Spanish children were able to read more of a sample of eight Applied Psycholinguistics 22:4 574 Ellis & Hooper: Reading in Welsh and English monosyllabic and eight disyllabic words in their language than were matched French or English children (Goswami, Gombert, & De Barrera, 1998); and Turkish children are able to read and spell with high degrees of accuracy by the end of the first grade (Öney & Durgunoglu, 1997). Segmental phonological awareness Whereas young children are aware of the structure of spoken language at the syllable, onset, and rhyme levels, the representation of segmental information at the phoneme level seems partly dependent on learning to read an alphabetic orthography. The written graphemes provide explicit feedback that clarifies the phonemic code (Brown & Ellis, 1994; Bradley & Bryant, 1983; Ellis & Large, 1987; Goswami & Bryant, 1990; Morais, Alegria, & Content, 1987; Wimmer, Landerl, Linortner, & Hummer, 1991), and explicit focus on the representations of sounds as they are written during spelling instruction seems to push along the acquisition of explicit phoneme segmentation abilities (Ellis & Cataldo, 1990; Frith, 1985; Goswami & Bryant, 1990). This suggests that the representation of phonemic information as a consequence of the acquisition of reading should depend on the degree of consistency or ambiguity of the system of symbol–sound mappings in a language. Research has demonstrated this to be so: phonemic segmentation develops in a rapid spurt in languages such as Greek or German, in which such graphemic feedback is a reliable index of segmental phonology, and it tends to be more protracted in orthographies such as French or English, which are more ambiguous in this respect (Cossu, Shankweiler, Liberman, Katz, & Tola, 1988; Frith, Wimmer, & Landerl, 1997; Goswami et al., 1998; Goswami, Porpodas, & Wheelwright, 1997; Wimmer & Goswami, 1994). Onset-rime awareness is a predictor of reading development for English, with its relatively high degree of spelling–sound consistency at the level of rime compared to the level of the vowel phoneme, but not in German, in which phoneme awareness is a much stronger predictor (Wimmer et al., 1991; Wimmer, Landerl, & Schneider, 1994).