The effect of oral vocabulary on reading visually novel words: a comparison of the dual-route-cascaded and triangle frameworks

Abstract

Dual-route-cascaded (DRC) (e.g. Coltheart & Rastle, Journal of Experimental Psychology: Human Perception and Performance 20 (1994) 1197) and triangle framework (e.g. Seidenberg & McClelland, Psychological Review 96 (1989) 523) predictions were tested regarding the effect of having a word in oral vocabulary prior to reading that same word. Over two sessions, at intervals of 2–3 days, 44 Grade 1 (6–7-year-old) children were aurally familiarized with the sound and meaning of ten novel words (semantic oral instantiation), and with just the sound of another ten novel words (non-semantic oral instantiation). Two to three days later non-word naming performance was significantly more accurate for aurally trained novel words compared to pseudohomophones, which were in turn advantaged over untrained non-words. The semantic manipulation had no effect. Experiment 2 manipulated articulation during (non-semantic) training. Forty Grade 1 children participated. Again, aurally trained items were named more accurately and quickly than equivalently trained pseudohomophones, which were in turn advantaged over untrained non-words. The articulation manipulation had no effect. The results suggest that word-specific phonological information is represented in the reading system independently of semantic or articulatory influences. The results are interpreted as being problematic for both the DRC and triangle frameworks, but more so for the latter.

Introduction

In the process of learning to read children are continually confronted with words that they have never seen in printed form before. In alphabetic languages such as English, the ability to decipher a novel letter string requires knowledge of the specific letter–sound (grapheme–phoneme) correspondences of the word. Once learned, knowledge of grapheme–phoneme correspondences enables the reader to sound out a candidate pronunciation for a novel letter string independently of outside assistance, an ability known as phonological recoding (Share, 1995).

Notwithstanding the importance of phonological recoding, it is possible that other sources of knowledge assist the beginning reader to decode visually unfamiliar words. One possible source of additional information is the child's extensive oral vocabulary. Indeed, most words that children encounter for the first time in age-appropriate reading material will already be familiar to them in spoken form (Chall, 1987). Given the substantial oral vocabulary that children bring to reading, we became interested in addressing the following question: is it reasonable to assume that all unfamiliar written words are processed alike or do those that are familiar to the reader in spoken form somehow benefit from their prior representation in the spoken language system?

Many theorists have speculated that previously encoded representations of the sound (phonology) and meaning (semantics) of spoken words might assist the beginning reader to achieve a correct match between a known pronunciation and an incomplete or inaccurate recoding attempt over and above pure phonological recoding ability (e.g. Ellis and Young, 1988, Henderson, 1985, Laxon et al., 1995, Share, 1995, Taft, 1982). However, attempts to demonstrate such an advantage have been indirect and inconclusive.

To date, investigations into the role of prior representations of phonology and semantics in reading novel letter strings have been largely confined to manipulations of pseudohomophones, a special class of non-words whose pronunciations match those of real words (e.g. brane). A number of studies have found pseudohomophones to be named both more accurately and more quickly than orthographically matched non-homophonic non-words (e.g. prane) on isolated non-word naming tasks. This naming advantage for pseudohomophones has been demonstrated with adults (e.g. Grainger et al., 2000, Marmurek and Kwantes, 1996, McCann and Besner, 1987, Taft and Russell, 1992), with normally-reading children (e.g. Laxon et al., 1995, Pring and Snowling, 1986), in acquired dyslexia (e.g. Derouesne & Beauvois, 1985), and in developmental phonological dyslexia (e.g. Temple & Marshall, 1983).

One explanation for the naming advantage observed for pseudohomophones is that it reflects the activation of the (phonological) lexical representation of the base-word from which the pseudohomophone was derived (e.g. Coltheart and Rastle, 1994, McCann and Besner, 1987). Such an explanation assumes the activation of word-specific information via either localist representations of whole words (e.g. Coltheart & Rastle, 1994) or through the interaction of semantic and phonological information in a distributed network (e.g. Harm, 1998).

Two alternative, non-lexical, explanations have also been proposed to explain the naming advantage for pseudohomophones. Firstly, it may be that the naming advantage reflects the possibility that pseudohomophones are more orthographically word-like than other non-words (i.e. that they share more features with the orthography–phonology mappings of previously encountered words). Indeed, the naming advantage for pseudohomophones over controls has been demonstrated to disappear when pseudohomophones and control non-words are matched exactly for onset frequency (Seidenberg, Peterson, MacDonald, & Plaut, 1996). A second non-lexical explanation is that the articulatory motor patterns for pronouncing pseudohomophones are more familiar, and therefore easier to produce (Seidenberg et al., 1996). These two alternative explanations are fully consistent with distributed connectionist models of the reading system which are based on frequency-sensitive (stochastic) learning mechanisms.

However, another consistent experimental observation is that pseudohomophones used as foils in speeded lexical decision tasks are classified as non-words less rapidly than orthographically matched non-homophonic non-words (e.g. Coltheart et al., 1977, Rubenstein et al., 1971, Seidenberg et al., 1996). Clearly, the interference observed for pseudohomophones in lexical decisions cannot be explained in terms of articulatory familiarity because lexical decisions do not involve naming. Further, the magnitude of the pseudohomophone effect in lexical decisions has been shown to remain constant even when pseudohomophones and control non-words are matched exactly for onset and rime frequency (Seidenberg et al., 1996), and so cannot be explained in terms of pseudohomophones sharing more orthographic features with known words. Instead, the interference observed for pseudohomophones in lexical decisions requires an explanation in terms of the activation of word-specific information.

Whether it is possible for word-specific information to be activated in phonology without semantic mediation is an issue that separates the (local representationalist) dual-route-cascaded (DRC) account of the reading system put forward by Coltheart and colleagues (e.g. Coltheart et al., 1993, Coltheart and Rastle, 1994) from the parallel-distributed processing (PDP) account of the reading system proposed by Seidenberg and McClelland (1989) (see also Plaut, McClelland, Seidenberg, & Patterson, 1996). This representational distinction between the two models is of particular interest in the present context because it determines the predictions that each model makes regarding the likely effect of knowing the spoken form of a word prior to encountering it in printed form.

Before discussing the specific DRC and PDP predictions for the present study, it is important to understand how they were derived. Both the DRC and PDP accounts of reading have been partially implemented as computational models. Clearly, these incomplete implementations are more constrained than the broader theories from which they are derived, a situation that makes accurate comparisons of the theories difficult. So, with the goal of facilitating theory adjudication, we sought to explicate those theoretical commitments specific to each approach that constrain them independently of the details of any particular partial implementation. A detailed account of these constraints, and of how they determine the respective predictions of the two approaches in relation to the current investigation, is given below.

The DRC model is a partial computational implementation of Coltheart's dual-route theory of visual word recognition and reading aloud (Coltheart, 1978, Coltheart et al., 1993, Coltheart and Rastle, 1994). The dual-route theory proposes two processes (routes) for converting print to sound: a lexical route for the fast recognition of familiar orthographic forms, and a non-lexical route for ‘sounding out’ novel orthographic inputs via the sequential application of grapheme–phoneme conversion (GPC) rules. The lexical route enables the reader to pronounce irregularly spelled words correctly. The non-lexical route assembles a correct pronunciation for regularly spelled words, and allows the reader to produce a regular pronunciation for non-words (or for real words not previously encountered in print), but produces mispronunciations, or regularization errors, for irregularly spelled words (e.g. pint pronounced to rhyme with mint).

It is the putative existence of these two routes for converting print to sound that gives the dual-route theory its name. However, the term dual-route is somewhat misleading in so far as the lexical route is further subdivided into the semantic and non-semantic pathways. The semantic pathway allows direct access to word meanings from orthographic representations, prior to phonological retrieval, while the non-semantic pathway enables phonological representations to be addressed directly from orthography without semantic mediation. The distinction between the semantic and non-semantic pathways is made in order to account for the reading performance of some patients with acquired phonological dyslexia (substantially better word reading than non-word reading) who, in addition to displaying a near complete inability to read non-words, have a semantic impairment that renders the semantic pathway insufficient to explain their preserved ability to read real words (Coltheart et al., 1993).

The DRC model only implements the non-lexical route and the non-semantic pathway of the lexical route (henceforth, the lexical/non-semantic route; see Coltheart, Rastle, Perry, Langdon, & Ziegler, 1999 for a detailed account). Early verbal accounts of dual-route theory (e.g. Coltheart, 1978) precluded the possibility that a pronunciation generated for a novel letter string via the non-lexical route could access and be influenced by information stored in the lexicons. In contrast, the DRC model allows information from each route to influence processing in the other (e.g. Coltheart & Rastle, 1994) via bidirectional connections between the assembled output of the non-lexical route in the phoneme units and the representations of word entries in the phonological output lexicon (Coltheart & Rastle, 1994). Indeed, it is these bidirectional connections between the two routes that enables the DRC model to explain the apparently lexical basis of pseudohomophone effects (Coltheart & Rastle, 1994).

The implemented non-lexical route works by the serial application of GPC rules. In contrast, processing in the lexical/non-semantic route occurs in parallel, being modelled on the interactive activation architecture developed by McClelland and Rumelhart (1981) whereby the layers in the model interact with each other either via direct connections between layers or via activation cascading through one or more layers. The orthographic input and phonological output lexicons of the lexical/non-semantic route consist of nodes representing word entries, each with a frequency parameter that causes the rate at which activation accumulates for a given input to be proportional to its frequency of occurrence in written/spoken language. The two lexicons are linked to each other via bidirectional excitatory connections. Inhibitory connections within each lexicon cause competing words to inhibit each other (Coltheart, Rastle et al., 1999).

While the lexical/semantic pathway is yet to be computationally implemented (but see Coltheart, Woollams, Kinoshita, & Perry, 1999 for a partial implementation of the semantic route), it is possible to make general predictions concerning the likely effects of the lexical/semantic route given the box-and-arrow diagram for the full DRC model (see Fig. 1). Accordingly, the activation of semantic information for a familiar word would occur rapidly via recursive excitatory connections with the orthographic input lexicon. Semantic information would then cascade to the word's phonological representation in the phonological output lexicon, and through to the phoneme system. Given the bidirectional connections between the non-lexical and lexical routes, it is also possible for semantic information to be activated via phonological information assembled by the non-lexical route. Thus, the activation of semantic information from both lexical and non-lexical sources would serve to strengthen the activation for any orthographic input that corresponds to a known word meaning. It is the full framework for the DRC model shown in Fig. 1, which includes the un-implemented semantic pathway, from which the predictions for the current experiments were derived. It will be referred to as the DRC framework to differentiate it from the partially implemented DRC model.

Seidenberg and McClelland (1989) introduced a PDP, or connectionist, model of the reading system as an alternative to the dual-route theory. Seidenberg and McClelland's PDP framework defines the reading system as consisting of three representational domains – orthography, phonology and semantics, each linked to the other by a system of parallel feedforward and feedback connections. Located between any two domains of processing units are a set of hidden units that mediate between them. The entire framework is characterized by the triangular configuration depicted in Fig. 2, and is generally referred to as the triangle framework.

According to the triangle framework, the task of reading aloud can be accounted for without recourse to localist word-specific representations or distinct processing mechanisms for words and non-words. Specifically, the triangle framework does not represent knowledge of words explicitly, but the weights on connections between units. The weights encode the statistical properties of all the words that the system has encountered, with those items encountered most frequently having the largest influence on the weights. Further, the triangle framework proposes a single mechanism for the pronunciation of all types of letter strings, whether they are regular words, irregular words or non-words. When a letter string is presented to the system, orthographic, phonological and semantic information interact until the network as a whole settles into a stable pattern of activity that corresponds to the interpretation of that input. The pronunciation for a novel letter string is therefore generated on the basis of the combined influence of all known orthography-to-phonology-to-semantic mappings, with those most similar to the input string having the strongest effect.

A number of partial computational implementations of the triangle framework have been reported in the literature since the original Seidenberg and McClelland (1989) implementation of the orthography-to-phonology pathway (for improved implementations of the orthography-to-phonology pathway see Harm and Seidenberg, 1999, Plaut et al., 1996; for implementations of the orthography-to-semantics pathway see Hinton and Shallice, 1991, McLeod et al., 2000, Plaut and Shallice, 1993; for a feedforward only implementation of all three processing domains see Plaut, 1997). The absence of a complete implementation of the triangle framework in the published literature makes it difficult to derive precise predictions (but see Harm, 1998, for a full implementation of the triangle framework). However, by tracing the evolution of the triangle framework, it is possible to explicate a set of theoretical and implementational commitments that constrain the approach, and from which predictions can be derived, independently of any particular implementation. The two constraints most relevant to the current investigation are outlined below.

The first constraint applies to the nature of the representations that develop over the orthography-to-semantics mapping. The mapping between orthography and semantics for morphologically simple words is largely arbitrary in so far as visually similar words, such as cave and save, are usually unrelated in meaning. In order to learn such an arbitrary mapping, implementations of the orthography-to-semantics pathway of the triangle framework utilize an attractor architecture capable of representing higher order relationships between domains (e.g. Hinton and Shallice, 1991, McLeod et al., 2000, Plaut and Shallice, 1993). The architecture of the semantic attractor causes the network to develop an extreme sensitivity to small differences in the orthographic input so that, over the course of training, patterns of activity corresponding to the exact meanings of individual words come to be represented as stable attractors in the space of semantic features. This extreme sensitivity to small orthographic differences not only prevents the semantic features for visually similar words from becoming confused, but also prevents semantic features from being strongly activated by non-words that are visually similar to known words. Thus, the semantic domain provides a natural source of information for distinguishing words from non-words (Plaut, 1997, Plaut et al., 1996).

The second constraint on the triangle framework relates to the representations formed in the phonological domain. The triangle framework's commitment to processing both words and non-words via the orthography–phonology connections makes the highly specific attractors formed in the semantic domain unsuitable for the phonological domain. That is, because the mappings between orthography and phonology are much more systematic than those between orthography and semantics, highly specific attractors would have the effect of drawing non-words that are visually similar to known words (e.g. sare) into the attractors of known words (e.g. care, save), the result being an unacceptably high level of lexicalization errors for unfamiliar stimuli, and poor non-word generalization (Plaut et al., 1996). Instead, much broader, so-called componential attractors, are required if both word and non-word reading are to be accomplished by the orthography-to-phonology pathway (e.g. Harm, 1998, Plaut et al., 1996). Componential attractors in the triangle framework develop a substructure representing the sublexical categories of onset, vowel and coda. The degree of componentiality varies according to the degree of systematicity in the mapping, ranging from highly componential for regular/consistent words to much more specific attractors for exception words (Plaut et al., 1996, see Fig. 15).

The key point for the present investigation is that the highly componential (sublexical) phonological attractors that develop for regular/consistent words are unlikely to show word-specific effects, analogous to the way that word-specific effects are not generated by an isolated GPC route in the DRC framework. Instead, the representational constraints on the triangle framework make it likely that word-specific effects for regular/consistent stimuli will only arise with support from the much more specific semantic attractors. Indeed, semantic mediation is invoked to explain the apparently lexical basis of the pseudohomophone effect in lexical decisions (e.g. Harm, 1998, Seidenberg et al., 1996). It is this point that distinguishes the predictions of the DRC and triangle frameworks. Specifically, where the input consists of regular/consistent, phonologically familiar but orthographically novel words, the DRC model predicts an advantage of phonological familiarity independently of semantics (due to the lexical/non-semantic pathway). In contrast, the triangle framework must invoke the notion of semantic mediation.

The results of a study conducted with beginning readers by Pratt, Johnston, and Morton (1999) suggest that familiarity with the spoken form of a word does assist the reader to produce a correct pronunciation. Pratt et al. (1999) piloted a new research technique for exploring the effect of prior exposure to the spoken form and meaning of visually novel words. The procedure involved teaching a group of children in Grades 2 and 3 (aged 7–9 years) to associate non-words presented orally in the context of a story with pictures of imaginary creatures. In this way, the non-words were instantiated into the children's spoken vocabularies. In order to maximize the likelihood that the non-words would be instantiated, the story was presented on at least two separate occasions at intervals of between 2 and 3 days. Two to three days after the final training session a non-word naming test was administered containing the ten orally instantiated items as well as ten matched non-words for which no training had been given (control). It was found that the children named the orally instantiated non-words nearly twice as accurately as the control non-words.

The method piloted by Pratt et al. (1999) provides an analogue for the experience of encountering an aurally familiar word in print for the first time, and avoids some of the confounds inherent in the pseudohomophone task (e.g. the existence of a competing orthographic representation, prior articulatory experience, etc.). However, demonstrating an advantage for naming orthographically unfamiliar letter strings for which both the phonology and meaning are known (recall that the items were presented in the context of a story) does not serve to adjudicate between DRC and triangle framework predictions. The DRC framework could account for the Pratt et al. results by proposing that training resulted in the establishment of discrete phonological and semantic lexical nodes for each of the trained items, which subsequently assisted processing at test. Equally plausibly, the triangle framework would propose that highly specific (non-componential) semantic attractors were formed for the novel words during training, and were subsequently activated via the computed phonological code at test, producing the advantage over untrained non-words.¹

One way to tease apart the two accounts would be to train items under both semantic and non-semantic conditions. Indeed, it is conceivable that a word might be phonologically familiar, while its meaning is unknown. While the triangle framework would have difficulty in accounting for an observed advantage for trained (regular/consistent) novel words over untrained non-words in the absence of support from semantics, the DRC framework explicitly allows for this possibility via the activation of whole word representations in the lexical/non-semantic pathway.