Kathleen Rastle

DRC: a dual route cascaded model of visual word recognition and reading aloud

This article describes the Dual Route Cascaded (DRC) model, a computational model of visual word recognition and reading aloud. The DRC is a computational realization of the dual-route theory of reading, and is the only computational model of reading that can perform the 2 tasks most commonly used to study reading: lexical decision and reading aloud. For both tasks, the authors show that a wide variety of variables that influence human latencies influence the DRC models latencies in exactly the same way. The DRC model simulates a number of such effects that other computational models of reading do not, but there appear to be no effects that any other current computational model of reading can simulate but that the DRC model cannot. The authors conclude that the DRC model is the most successful of the existing computational models of reading.

The broth in my brother’s brothel: Morpho-orthographic segmentation in visual word recognition

Much research suggests that words comprising more than one morpheme are represented in a “decomposed” manner in the visual word recognition system. In the research presented here, we investigate what information is used to segment a word into its morphemic constituents and, in particular, whether semantic information plays a role in that segmentation. Participants made visual lexical decisions to stem targets preceded by masked primes sharing (1) a semantically transparent morphological relationship with the target (e.g.,cleaner-CLEAN), (2) an apparent morphological relationship but no semantic relationship with the target (e.g.,corner-CORN), and (3) a nonmorphological form relationship with the target (e.g.,brothel-BROTH). Results showed significant and equivalent masked priming effects in cases in which primes and targets appeared to be morphologically related, and priming in these conditions could be distinguished from nonmorphological form priming. We argue that these findings suggest a level of representation at which apparently complex words are decomposed on the basis of their morpho-orthographic properties. Implications of these findings for computational models of reading are discussed.

Morphological and semantic effects in visual word recognition : A time-course study

Some theories of visual word recognition postulate that there is a level of processing or representation at which morphemes are treated differently fromwhole words. Support for these theories has been derived frompriming experiments in which the recognition of a target word is facilitated by the prior presentation of a morphologically related prime (departure-DEPART). In English, such facilitation could be due to morphological relatedness, or to some combination of the orthographic and semantic relatedness characteristic of derivationally related words. We report two sets of visual priming experiments in which the morphological, semantic, and orthographic relationships between primes and targets are varied in three SOA conditions (43 ms, 72 ms, and 230 ms). Results showed that morphological structure plays a significant role in the early visual recognitionof English words that is independent of both semantic and orthographic relatedness. Findings are discussed in terms of current approaches to morphological processing.

358,534 nonwords: The ARC Nonword Database

The authors present a model of the phonotactic and orthographic constraints of Australian and Standard Southern British English monosyllables. This model is used as the basis for a web-based psycholinguistic resource, the ARC Nonword Database, which contains 358,534 monosyllabic nonwords—48,534 pseudohomophones and 310,000 non-pseudohomophonic nonwords. Items can be selected from the ARC Nonword Database on the basis of a wide variety of properties known or suspected to be of theoretical importance for the investigation of reading.

Morphological decomposition based on the analysis of orthography

Recent theories of morphological processing have been dominated by the notion that morphologically complex words are decomposed into their constituents on the basis of their semantic properties. In this article we argue that the weight of evidence now suggests that the recognition of morphologically complex words begins with a rapid morphemic segmentation based solely on the analysis of orthography. Following a review of this evidence, we discuss the characteristics of this form of decomposition, speculate on what its purpose might be, consider how it might be learned in the developing reader, and describe what is known of its neural bases. Our discussion ends by reflecting on how evidence for semantically based decomposition might be (re)interpreted in the context of the orthographically based form of decomposition that we have described.

Serial and strategic effects in reading aloud.

Coltheart and Rastle (1994) reported that the size of th e regularity effect on word naming latency decreases across position of irregul arity, implicating a serial process in reading aloud. In response to criticism by Pl aut, McClelland, Seidenberg, and Patterson (1996), we replicate these results here using mo osyllabic words which have been controlled for consistency at each of five orthographic segments. A successful simulation of these data by the DRC model (C oltheart, Curtis, Atkins, & Haller, 1993) is presented. These findings were used in a second experiment to produce a strategy effect in reading aloud. Subjects named nonword or regular word targe ts mixed with either first position irregular fillers or third position irre gular fillers. Target naming was slowed when first position irregular fillers were pres ent compared with target naming when third position irregular fillers were present. The se data suggest that the use of the nonlexical route is not fixed; subjects can slow it s use if faced with very costly exception words. A simulation of these data by the DRC model is presented. Serial and Strategic Effec ts 3 Serial and Strategic Effects in Reading Aloud The term ‘dual-route theory’ refers to a particular cla ss of theories of visual word recognition and reading aloud. The defining feature of such theories is the postulate that there are two different procedures for con verti g print to speech, a dictionary-lookup or lexical procedure and a rule-based or nonl exical procedure (expositions of dual-route theories can be found in e.g., Baron & Strawson, 1976; Coltheart, 1978; Coltheart, 1985; Ellis & Young, 1988; Gough & Cosk y, 1977; Forster & Chambers 1973; Morton & Patterson, 1980; Ogden, 1996; P aap & Noel, 1991; Patterson & Morton, 1985; Patterson & Shewell, 1987). Coltheart, Curtis, Atkins, and Haller (1993; see also Colt heart, Langdon, & Haller, 1996; Coltheart & Rastle, 1994; Rastle & Coltheart, in press-a; Rastle & Coltheart, in press-b) described a computational realiza tion of dual-route theory, the Dual Route Cascaded (DRC) model. The model they described is computational in the sense that it exists as a computer program which can pe rform the tasks typically used in research on reading such as lexical decision and r e ding aloud. The number of processing cycles needed to perform any such task with a particular stimulus is an analogue of the number of milliseconds needed by human bein gs, so direct simulations of experiments that yield human reaction times is straightforward. Four other models of reading which are computational in this sense currently exist: the parallel-distributed-processing (PDP) implemen tations described by Plaut, McClelland, Seidenberg, and Patterson (1996), the MultipleLevels model of Norris (1994), the Connectionist Dual-Process model of Zorzi, Ho ughton, and Butterworth (in press), and the Multiple Readout Model (MROM) of Gra inger and Jacobs (1996). The DRC model differs from these models in that it is applicable to both the Serial and Strategic Effec ts 4 simulation of lexical decision and the simulation of reading aloud (that is, it is a model of both visual word recognition and reading aloud). The simulations contained in the Norris (1994), Zorzi et al. (in press), and Plaut e t al. (1996) implementations are of reading aloud only; those models, in their present for m, have no procedure for making lexical decisions. Similarly, the simulations in Grainger and Jacobs (1996) are of lexical decision only; the MROM model, in its c urrent form, does not read aloud. Since we are concerned in this paper with reading aloud rat her than with lexical decision, we focus on comparing the DRC model wi th the Multiple-Levels model, the Connectionist Dual-Process model, and the PDP implementations, since these are the only current computational models of readi ng aloud. The DRC Model The overall architecture of the DRC model is shown in Figure 1. As is evident, the model has both a dictionary lookup (lexical) procedure for converting print to speech and a rule-based (nonlexical) procedure for s uch conversion. _____________________ Insert Figure 1 about here _____________________ DRC’s Lexical Route The lexical route consists of a sequence of five process ing components or levels: feature detection, letter identification, orth graphic lexicon, phonological lexicon, and phoneme activation. The first three of these levels are simply a generalization of the Interactive Activation (IA) mo del of visual word recognition (McClelland & Rumelhart, 1981; Rumelhart & McClelland, 1982). Ho wever, instead of being restricted just to four-letter monosyllabic monom orphemic words (as was the Serial and Strategic Effec ts 5 case with the IA model), the DRC model operates with w ords of any length up to seven letters 1 and is not restricted to monomorphemic words (though is r estricted to monosyllabic words). Apart from these differences, th e architecture, connectivity, and mode of operation of these three levels is identica l to what is found in the IA model. The feature level consists of eight sets of feature unit s, one for each of the eight possible letter positions of an input string. Each of t ese eight sets of feature units contains 16 feature-present units and 16 feature-absent uni ts (the number is 16 because the system operates with the 16-stroke font of R umelhart and Siple, 1974, exactly as in the IA model). The letter level also consists of eight sets of units. Each of these sets of letter units contains 27 units, one for each letter of the alpha bet nd one coding the absence of any letter in that position in the input string. Each unit at the feature level for a particular input pos ition has an excitatory connection to all of the letters for that input positio n which possess that feature and an inhibitory connection to all of the letters that do n t. There are no connections from the letter level back to the feature level, nor are th ere inhibitory connections within each set of feature units. At the letter level, howev er, there is within-level inhibition: for each letter position, all letter units have inhibit ory connections to all other letter units. The orthographic lexicon contains 7,980 units, one unit for ea ch monosyllabic word in the CELEX English database (Baayen, Piepenbroc k, & van Rijn, 1993), except that infrequent words of foreign origin such as KV ASS or LAKH have been culled. Each word unit in this lexicon has inhibitory conne ctions to all other word units in the orthographic lexicon. Each word unit also has a par meter representing Serial and Strategic Effec ts 6 its frequency; as in the IA model, the value of this param eter ranges from -0.05 (for the least frequent word) to 0.00 (for the most frequent word ). These frequency parameters cause the rate at which activation rises to be positively related to word frequency. For each position at the letter level, every letter unit has excitatory connections to every entry in the orthographic lexicon representing a word which possesses that letter in that position, and inhibitory connections to a ll other word units. Furthermore, each word unit has excitatory connections back to all of the letter units which represent its spelling and inhibitory connections to all ot her letter units. The phonological lexicon contains 7,117 units, with each unit representing the pronunciation (in Australian English) of one of the e ntries in the orthographic lexicon. Each word unit in this lexicon has inhibitory connections t o all other word units in the phonological lexicon. These units are frequency coded in t he same way as are the entries in the orthographic lexicon. Heterographic homoph ones such as SO and SEW have separate entries in the orthographic lexicon but act ivate the same entry in the phonological lexicon. Homographic heterophones such as LE AD have a single entry in the orthographic lexicon which activates two different e ntri s in the phonological lexicon. Apart from these cases, connections from the orthographic lexicon to the phonological lexicon are one to one. The phoneme level consists of eight sets of phoneme u nits. Each of these sets contains 44 phoneme units, one for each of the 43 phonemes i n DRC’s phonemic vocabulary (see the list of these phonemes in Appendix A) and one coding the absence of a phoneme in that position in the output st ring. Each entry in the phonological lexicon has excitatory connections to all o f its constituent phonemes and inhibitory connections to all other phonemes. In turn, each phoneme unit has Serial and Strategic Effec ts 7 excitatory connections back to all word units in the phon ol gical lexicon that contain that phoneme in that position and inhibitory connections back to all other units in the phonological lexicon. DRC’s Nonlexical Route The nonlexical route in the DRC model is a sequence of f our processing components or levels: feature detection, letter identif ica on, grapheme-phoneme conversion, and phoneme activation. The feature, lett r, and phoneme levels are shared with the lexical route and have been described. The rules used for graphemephoneme conversion are described in the next section. The nonlexical route operates as follows. For the fir st N processing cycles (N is set to 10 cycles in the standard set of parameters), t he route is inoperative. At cycle 11, GPC rules are applied to the first letter of the input string. After a further M processing cycles have elapsed (M is set to 17 cycles in the standard set of parameters), the next letter of the string becomes avai lable for nonlexical translation: the input to the GPC rules is now a two-letter string. Thus, the string becomes available for translation serially, letter by letter, from left to right; the translation process then operates on the entire available string in parallel. Nonwords are not processed entirely by the nonlexical ro ute, however. They activate word neighbors in the ortho

Masked Phonological Priming Effects in English: Are They Real? Do They Matter?.

For over 15 years, masked phonological priming effects have been offered as evidence that phonology plays a leading role in visual word recognition. The existence of these effects-along with their theoretical implications-has, however, been disputed. The authors present three sources of evidence relevant to an assessment of the existence and implications of these effects. First, they present an exhaustive meta-analytic literature review, in which they evaluate the strength of the evidence for masked phonological priming effects on English visual word processing. Second, they present two original experiments that demonstrate a small but significant masked priming effect on English visual lexical decision, which persists in conditions that may discourage phonological recoding. Finally, they assess the theory of visual word recognition offered by the DRC model (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) in the context of their empirical data. Through numerous simulations with this model, they argue that masked phonological priming effects might best be captured by a weak phonological (i.e., dual-access) theory in which lexical decisions are made on the basis of phonological information.

Can cognitive models explain brain activation during word and pseudoword reading? A meta-analysis of 36 neuroimaging studies.

Reading in many alphabetic writing systems depends on both item-specific knowledge used to read irregular words (sew, yacht) and generative spelling-sound knowledge used to read pseudowords (tew, yash). Research into the neural basis of these abilities has been directed largely by cognitive accounts proposed by the dual-route cascaded and triangle models of reading. We develop a framework that enables predictions for neural activity to be derived from cognitive models of reading using 2 principles: (a) the extent to which a model component or brain region is engaged by a stimulus and (b) how much effort is exerted in processing that stimulus. To evaluate the derived predictions, we conducted a meta-analysis of 36 neuroimaging studies of reading using the quantitative activation likelihood estimation technique. Reliable clusters of activity are localized during word versus pseudoword and irregular versus regular word reading and demonstrate a great deal of convergence between the functional organization of the reading system put forward by cognitive models and the neural systems activated during reading tasks. Specifically, left-hemisphere activation clusters are revealed reflecting orthographic analysis (occipitotemporal cortex), lexical and/or semantic processing (anterior fusiform, middle temporal gyrus), spelling-sound conversion (inferior parietal cortex), and phonological output resolution (inferior frontal gyrus). Our framework and results establish that cognitive models of reading are relevant for interpreting neuroimaging studies and that neuroscientific studies can provide data relevant for advancing cognitive models. This article thus provides a firm empirical foundation from which to improve integration between cognitive and neural accounts of the reading process.

Whammies and double whammies: The effect of length on nonword reading

In the work presented here, the length effect in nonword reading aloud was investigated in order to assess whether that effect is driven by the number of letters in a string or by the number of graphemes in a string. Simulation work with the Dual-Route Cascaded (DRC) model (e.g., Coltheart, Curtis, Atkins, & Haller, 1993; Coltheart & Rastle, 1994) uncovered a surprising finding regarding the length effect; the same result was obtained in an experiment with human subjects. The results are discussed in terms of the DRC model, with particular reference to serial processing and interphoneme inhibition, two properties critical to understanding the effect reported here.

ERP Evidence of Morphological Analysis from Orthography: A Masked Priming Study

There is broad consensus that the visual word recognition system is sensitive to morphological structure (e.g., hunter = hunt + er). Moreover, it has been assumed that the analysis of morphologically complex words (e.g., hunter) occurs only if the meaning of the complex form can be derived from the meanings of its constituents (e.g., hunt and er). However, recent behavioral work using masked priming has suggested that morphological analysis can occur at an early, orthographic level, with little influence from semantics. The present investigation examined the neurophysiological correlates of masked priming in conditions of a genuine morphological relationship (e.g., hunter-HUNT), an apparent morphological relationship (corner-CORN), and no morphological relationship (brothel-BROTH). Neural priming was indexed by the reduction of the N400 ERP component associated with targets preceded by related primes, as compared to targets preceded by unrelated primes. The mere appearance of morphological structure (corner-CORN) resulted in robust behavioral and neural priming, whose magnitude was similar to that observed in pairs with genuine morphological relationship and greater than that in the nonmorphological pairs. The results support a purely structural morphemic segmentation procedure operating in the early stages of visual word perception.