Do current connectionist learning models account for reading development in different languages?

Learning to read a relatively irregular orthography, such as English, is harder and takes longer than learning to read a relatively regular orthography, such as German. At the end of grade 1, the difference in reading performance on a simple set of words and nonwords is quite dramatic. Whereas children using regular orthographies are already close to ceiling, English children read only about 40% of the words and nonwords correctly. It takes almost 4 years for English children to come close to the reading level of their German peers. In the present study, we investigated to what extent recent connectionist learning models are capable of simulating this cross-language learning rate effect as measured by nonword decoding accuracy. We implemented German and English versions of two major connectionist reading models, Plaut et al.'s (Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: computational principles in quasi-regular domains. Psychological Review, 103, 56-115) parallel distributed model and Zorzi et al.'s (Zorzi, M., Houghton, G., & Butterworth, B. (1998a). Two routes or one in reading aloud? A connectionist dual-process model. Journal of Experimental Psychology: Human Perception and Performance, 24, 1131-1161); two-layer associative network. While both models predicted an overall advantage for the more regular orthography (i.e. German over English), they failed to predict that the difference between children learning to read regular versus irregular orthographies is larger earlier on. Further investigations showed that the two-layer network could be brought to simulate the cross-language learning rate effect when cross-language differences in teaching methods (phonics versus whole-word approach) were taken into account. The present work thus shows that in order to adequately capture the pattern of reading acquisition displayed by children, current connectionist models must not only be sensitive to the statistical structure of spelling-to-sound relations but also to the way reading is taught in different countries.     

A Computational and Empirical Investigation of Graphemes in Reading

It is often assumed that graphemes are a crucial level of orthographic representation above letters. Current connectionist models of reading, however, do not address how the mapping from letters to graphemes is learned. One major challenge for computational modeling is therefore developing a model that learns this mapping and can assign the graphemes to linguistically meaningful categories such as the onset, vowel, and coda of a syllable. Here, we present a model that learns to do this in English for strings of any letter length and any number of syllables. The model is evaluated on error rates and further validated on the results of a behavioral experiment designed to examine ambiguities in the processing of graphemes. The results show that the model (a) chooses graphemes from letter strings with a high level of accuracy, even when trained on only a small portion of the English lexicon; (b) chooses a similar set of graphemes as people do in situations where different graphemes can potentially be selected; (c) predicts orthographic effects on segmentation which are found in human data; and (d) can be readily integrated into a full‐blown model of multi‐syllabic reading aloud such as CDP++ (Perry, Ziegler, & Zorzi, 2010). Altogether, these results suggest that the model provides a plausible hypothesis for the kind of computations that underlie the use of graphemes in skilled reading.     

Beyond single syllables: Large-scale modeling of reading aloud with the Connectionist Dual Process (CDP++) model

Most words in English have more than one syllable, yet the most influential computational models of reading aloud are restricted to processing monosyllabic words. Here, we present CDP++, a new version of the Connectionist Dual Process model (Perry, Ziegler, & Zorzi, 2007). CDP++ is able to simulate the reading aloud of mono- and disyllabic words and nonwords, and learns to assign stress in exactly the same way as it learns to associate graphemes with phonemes. CDP++ is able to simulate the monosyllabic benchmark effects its predecessor could, and therefore shows full backwards compatibility. CDP++ also accounts for a number of novel effects specific to disyllabic words, including the effects of stress regularity and syllable number. In terms of database performance, CDP++ accounts for over 49% of the reaction time variance on items selected from the English Lexicon Project, a very large database of several thousand of words. With its lexicon of over 32,000 words, CDP++ is therefore a notable example of the successful scaling-up of a connectionist model to a size that more realistically approximates the human lexical system.     

Nested incremental modeling in the development of computational theories: the CDP+ model of reading aloud

At least 3 different types of computational model have been shown to account for various facets of both normal and impaired single word reading: (a) the connectionist triangle model, (b) the dual-route cascaded model, and (c) the connectionist dual process model. Major strengths and weaknesses of these models are identified. In the spirit of nested incremental modeling, a new connectionist dual process model (the CDP+ model) is presented. This model builds on the strengths of 2 of the previous models while eliminating their weaknesses. Contrary to the dual-route cascaded model, CDP+ is able to learn and produce graded consistency effects. Contrary to the triangle and the connectionist dual process models, CDP+ accounts for serial effects and has more accurate nonword reading performance. CDP+ also beats all previous models by an order of magnitude when predicting individual item-level variance on large databases. Thus, the authors show that building on existing theories by combining the best features of previous models--a nested modeling strategy that is commonly used in other areas of science but often neglected in psychology--results in better and more powerful computational models.     

Reading and related skills in Grades 6, 7, 8 and 9: French normative data from EVALEC

IntroductionTo appropriately assess reading difficulties, tests designed according to an appropriate theoretical framework and based on normative data are required.ObjectiveWe used EVALEC (Sprenger-Charolles, Colé, Béchennec, & Kipffer-Piquard, 2005) to collect data on the word-level reading skills and reading-related skills (phonemic awareness, phonological short-term memory, and rapid naming) of middle school children (Grades 6 to 9, about 80 in each grade).MethodIn the tests focused on word-level reading skills, the effects of regularity (regular vs. irregular words), lexicality, and length (short vs. long irregular words and pseudowords) were examined. Accuracy and processing times were recorded for all tests.ResultsThe effects of both regularity and lexicality were significant, whatever the measure and independently of grade. Both accuracy and speed were lower for longer pseudowords, whereas length did not have a significant effect on irregular word latencies and, surprisingly, long irregular words were read more accurately than short ones. Reading level as assessed by a standardized test (Lefavrais, 2005) was not predicted by phonological short-term memory; rapid naming (color names) and phonemic awareness were both predictors but, in both cases, only response times predicted reading level.ConclusionThese results, and particularly those from the reading tasks, are discussed in relation to models of written-word recognition developed to account for the reading of multisyllabic items (Perry, Ziegler, & Zorzi, 2010) in orthographies shallower than English (Perry, Ziegler, & Zorzi, 2014).     

Zeading and reazing: which is faster? The position of the diverging letter in a pseudoword determines reading time

We present evidence that (a) at least some components of the reading process are serial and (b) pseudoword reading is affected by lexical knowledge, even in a transparent orthographic system like Italian. Pseudowords deriving from five-letter words by changing either the first or the fourth letter were presented for reading aloud. Results showed an effect of the position of the diverging letter: Early diverging pseudowords were read more slowly than late diverging pseudowords. The dual-route cascaded (DRC) model (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) successfully simulated the behavioural data.     

Modulation of regularity and lexicality effects in reading aloud

We examined the question of whether the sizes of the regularity and lexicality effects in naming can be modulated as a function of filler type (nonwords or low-frequency exception words). The lexicality effect was larger in the exception word filler condition than in the nonword filler condition, but the size of the regularity effect was essentially unaffected by filler type. This pattern is at odds with what is generally assumed to be the predictions from dual-route theories of reading aloud. An attempt was next made to determine whether the dual-route cascaded model of Coltheart, Rastle, Perry, Langdon, and Ziegler (2001) could possibly simulate this pattern when changes were introduced to each of the three parameters that affect the contribution of the nonlexical route. We discuss the implications of these results for the idea that reliance on the lexical and nonlexical routes is under strategic control.     

"Serial" effects in parallel models of reading

There is now considerable evidence showing that the time to read a word out loud is influenced by an interaction between orthographic length and lexicality. Given that length effects are interpreted by advocates of dual-route models as evidence of serial processing this would seem to pose a serious challenge to models of single word reading which postulate a common parallel processing mechanism for reading both words and nonwords (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001; Rastle, Havelka, Wydell, Coltheart, & Besner, 2009). However, an alternative explanation of these data is that visual processes outside the scope of existing parallel models are responsible for generating the word-length related phenomena (Seidenberg & Plaut, 1998). Here we demonstrate that a parallel model of single word reading can account for the differential word-length effects found in the naming latencies of words and nonwords, provided that it includes a mapping from visual to orthographic representations, and that the nature of those orthographic representations are not preconstrained. The model can also simulate other supposedly "serial" effects. The overall findings were consistent with the view that visual processing contributes substantially to the word-length effects in normal reading and provided evidence to support the single-route theory which assumes words and nonwords are processed in parallel by a common mechanism.     

When parallel processing in visual word recognition is not enough: new evidence from naming

Low-frequency irregular words are named more slowly and are more error prone than low-frequency regular words (the regularity effect). Rastle and Coltheart (1999) reported that this irregularity cost is modulated by the serial position of the irregular grapheme-phoneme correspondence, such that words with early irregularities exhibit a larger cost than words with late ones. They argued that these data implicate rule-based serial processing, and they also reported a successful simulation with a model that has a rule-based serial component—the DRC model of reading aloud (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001). However, Zorzi (2000) also simulated these data with a model that operates solely in parallel. Furthermore, Kwantes and Mewhort (1999) simulated these data with a serial processing model that has no rules for converting orthography to phonology. The human data reported by Rastle and Coltheart therefore neither require a serial processing account, nor successfully discriminate among a number of computational models of reading aloud. New data are presented wherein an interaction between the effects of regularity and serial position of irregularity is again reported for human readers. The DRC model simulated this interaction; no other implemented computational model does so. The present results are thus consistent with rule-based serial processing in reading aloud.     

Serial processing in reading aloud: no challenge for a parallel model

K. Rastle and M. Coltheart (1999) challenged parallel models of reading by showing that the cost of irregularity in low-frequency exception words was modulated by the position of the irregularity in the word. This position-of-irregularity effect was taken as strong evidence of serial processing in reading. This article refutes Rastle and Coltheart's theoretical conclusions in 3 ways: First, a parallel model, the connectionist dual process model (M. Zorzi, G. Houghton, & B. Butterworth, 1998b), produces a position-of-irregularity effect. Second, the supposed serial effect can be reduced to a position-specific grapheme-phoneme consistency effect. Third, the position-of-irregularity effect vanishes when the experimental data are reanalyzed using grapheme-phoneme consistency as the covariate. This demonstration has broader implications for studies aiming at adjudicating between models: Strong inferences should be avoided until the computational models are actually tested.     

The DRC model of visual word recognition and reading aloud: An extension to German

The dual-route cascaded (DRC) model of word recognition and reading aloud was implemented in German. In this paper, we describe crosslinguistic differences and similarities between the German and the English DRC. The German DRC was evaluated with respect to its ability to correctly pronounce all German monosyllabic words and to simulate the loan word (regularity) effect. Furthermore, we obtained DRC predictions concerning a number of benchmark effects previously investigated in English, namely effects of word frequency, word length, and neighbourhood size.     

Can the dual-route cascaded computational model of reading offer a valid account of the masked onset priming effect?

The masked onset priming effect (MOPE) refers to the empirical finding that target naming is faster when the target (SIB) is preceded by a briefly presented masked prime that starts with the same letter/phoneme (suf) than when it does not (mof; Kinoshita, 2000, Experiment 1). The dual-route cascaded (DRC) computational model of reading (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) has offered an explanation for how the MOPE might occur in humans. However, there has been some empirical discrepancy regarding whether for nonword items the effect is limited to the first-letter/phoneme overlap between primes and targets or whether orthographic/phonological priming effects occur beyond the first letter/phoneme. Experiment 1 tested these two possibilities. The human results, which were successfully simulated by the DRC model, showed priming beyond the first letter/phoneme. Nevertheless, two recent versions of the DRC model made different predictions regarding the nature of these priming effects. Experiment 2 examined whether it is facilitatory, inhibitory, or both, in order to adjudicate between the two versions of the model. The human results showed that primes exert both facilitatory and inhibitory effects.     

Nonword reading: Comparing dual-route cascaded and connectionist dual-process models with human data

Two prominent dual-route computational models of reading aloud are the dual-route cascaded (DRC) model, and the connectionist dual-process plus (CDP+) model. While sharing similarly designed lexical routes, the two models differ greatly in their respective nonlexical route architecture, such that they often differ on nonword pronunciation. Neither model has been appropriately tested for nonword reading pronunciation accuracy to date. We argue that empirical data on the nonword reading pronunciation of people is the ideal benchmark for testing. Data were gathered from 45 Australian-English-speaking psychology undergraduates reading aloud 412 nonwords. To provide contrast between the models, the nonwords were chosen specifically because DRC and CDP+ disagree on their pronunciation. Both models failed to accurately match the experiment data, and both have deficiencies in nonword reading performance. However, the CDP+ model performed significantly worse than the DRC model. CDP++, the recent successor to CDP+, had improved performance over CDP+, but was also significantly worse than DRC. In addition to highlighting performance shortcomings in each model, the variety of nonword responses given by participants points to a need for models that can account for this variety. (PsycINFO Database Record (c) 2012 APA, all rights reserved).     

Zeading and reazing: Which is faster? The position of the diverging letter in a pseudoword determines reading time

We present evidence that (a) at least some components of the reading process are serial and (b) pseudoword reading is affected by lexical knowledge, even in a transparent orthographic system like Italian. Pseudowords deriving from five-letter words by changing either the first or the fourth letter were presented for reading aloud. Results showed an effect of the position of the diverging letter: Early diverging pseudowords were read more slowly than late diverging pseudowords. The dual-route cascaded (DRC) model (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001 Coltheart, M., Rastle, K., Perry, C., Langdon, R. and Ziegler, J. 2001. DRC: A dual route cascaded model of visual word recognition and reading aloud. Psychological Review, 108: 204–256. [Crossref], [PubMed], [Web of Science ®] , [Google Scholar]) successfully simulated the behavioural data.     

There goes the neighbourhood: contextual control over the breadth of lexical activation when reading aloud

There are currently two computational accounts of how the time to read pseudohomophones (like BRANE) and their nonword controls (like FRANE) varies with changes in context. In Reynolds and Besner's (2005) account, readers vary the breadth of lexical activation in response to changes in context. A competing account proposed by Kwantes and Marmurek (2007) and independently by Perry, Ziegler, and Zorzi (2007) has readers varying their response criterion in response to changes in context. The present work adjudicates between these two accounts by examining how the effect of neighbourhood density changes as a function of list context when reading pseudohomophones aloud. The results of an experiment and simulations from a leading computational model support the lexical breadth account, but are inconsistent with the response criterion account.     

On the Anglocentricities of current reading research and practice: the perils of overreliance on an "outlier" orthography

In this critique of current reading research and practice, the author contends that the extreme ambiguity of English spelling-sound correspondence has confined reading science to an insular, Anglocentric research agenda addressing theoretical and applied issues with limited relevance for a universal science of reading. The unique problems posed by this "outlier" orthography, the author argues, have focused disproportionate attention on oral reading accuracy at the expense of silent reading, meaning access, and fluency, and have significantly distorted theorizing with regard to many issues-including phonological awareness, early reading instruction, the architecture of stage models of reading development, the definition and remediation of reading disability, and the role of lexical-semantic and supralexical information in word recognition. The dominant theoretical paradigm in contemporary (word) reading research--the Coltheart/Baron dual-route model (see, e.g., J. Baron, 1977; M. Coltheart, 1978) and, in large measure, its connectionist rivals--arose largely in response to English spelling-sound obtuseness. The model accounts for a range of English-language findings, but it is ill-equipped to serve the interests of a universal science of reading chiefly because it overlooks a fundamental unfamiliar-to-familiar/novice-to-expert dualism applicable to all words and readers in all orthographies.     

Dual-Route and Connectionist Models: A Step Towards a Combined Model

Current models of word recognition are mainly constructed within the frameworks of either dual-route or connectionist theories. The most important test of a word recognition model is how it succeeds in accounting for various reading behaviors. In the present paper dual-route and connectionist word recognition models are briefly described and evaluated. As a further development of these models, a combined framework is proposed. An amalgamation of the two main types of models might give a more satisfactory account of various phenomena within word recognition.     

German inflection: single route or dual route?

The German plural system has recently become a focal point for conflicting theories of language, both linguistic and cognitive. Marcus et al. (1995) highlight the German plural as support for the dual-route account of inflectional morphology first proposed by Pinker and colleagues (Pinker & Prince, 1988). On the dual-route account, inflectional morphology is universally subserved by a symbolic rule route which deals with regular inflection and an associative memory component which deals with irregular inflection. This contrasts with single-route connectionist systems. We seek to counter supposed evidence for the dual-route account through large-scale simulations as well as through experimental data. We argue that, in its current form, the dual-route account is incapable of generating experimental data provided by Marcus et al. (1995) as support. Finally, we provide direct quantitative comparisons between single-route and dual-route models of German plural inflection and find single-route performance superior on these tests.