Grapheme-phoneme Mappings Research Articles

Evaluating Transformer Models and Human Behaviors on Chinese Character Naming

Abstract Neural network models have been proposed to explain the grapheme-phoneme mapping process in humans for many alphabet languages. These models not only successfully learned the correspondence of the letter strings and their pronunciation, but also captured human behavior in nonce word naming tasks. How would the neural models perform for a non-alphabet language (e.g., Chinese) unknown character task? How well would the model capture human behavior? In this study, we first collect human speakers’ answers on unknown Character naming tasks and then evaluate a set of transformer models by comparing their performance with human behaviors on an unknown Chinese character naming task. We found that the models and humans behaved very similarly, that they had similar accuracy distribution for each character, and had a substantial overlap in answers. In addition, the models’ answers are highly correlated with humans’ answers. These results suggested that the transformer models can capture humans’ character naming behavior well.1

Impact of Learning to Read in a Mixed Approach on Neural Tuning to Words in Beginning Readers.

The impact of learning to read in a mixed approach using both the global and phonics teaching methods on the emergence of left hemisphere neural specialization for word recognition is yet unknown in children. Taking advantage of a natural school context with such a mixed approach, we tested 42 first graders behaviorally and with Fast Periodic Visual Stimulation using electroencephalographic recordings (FPVS-EEG) to measure selective neural responses to letter strings. Letter strings were inserted periodically (1/5) in pseudofonts in 40 s sequences displayed at 6 Hz and were either words globally taught at school, that could therefore be processed by visual whole-word form recognition (global method), or control words/pseudowords eliciting grapheme-phoneme (GP) mappings (phonics method). Results show that selective responses (F/5, 1.2 Hz) were left lateralized for control stimuli that triggered GP mappings but bilateral for globally taught words. It implies that neural mechanisms recruited during visual word processing are influenced by the nature of the mapping between written and spoken word forms. GP mappings induce left hemisphere discrimination responses, and visual recognition of whole-word forms induce bilateral responses, probably because the right hemisphere is relatively more involved in holistic visual object recognition. Splitting the group as a function of the mastery of GP mappings into “good” and “poor” readers strongly suggests that good readers actually processed all stimuli (including global words) predominantly with their left hemisphere, while poor readers showed bilateral responses for global words. These results show that in a mixed approach of teaching to read, global method instruction may induce neural processes that differ from those specialized for reading in the left hemisphere. Furthermore, given their difficulties in automatizing GP mappings, poor readers are especially prone to rely on this alternative visual strategy. A preprint of this paper has been released on Biorxiv (van de Walle de Ghelcke et al., 2018).

Automaticity of word recognition is a unique predictor of reading fluency in middle-school students.

Automaticity in word recognition has been hypothesized to be important in reading development (LaBerge & Samuels, 1974; Perfetti, 1985). However, when predicting educational outcomes, it is difficult to isolate the influence of automatic word recognition from factors such as processing speed or knowledge of grapheme-phoneme correspondences. Cognitive models suggest automaticity could be achieved in different components of word recognition (e.g., by memorizing familiar words or by tuning the mappings between orthography, phonology or semantics). However, the contributions of each path have not been assessed. This study developed a new measure of automaticity to overcome these limitations and relates automaticity to standard outcomes. Subjects were 58 middle-school students (mean age = 13.2 years ± 8 months) with average to below-average reading comprehension. To assess automaticity with an accuracy-based measure, backward masking was used: On half the trials, items were presented for 90 ms and replaced by a nonlinguistic mask; on the other half it was presented unmasked to assess children’s knowledge of the word. This was instantiated in 3 experimental tasks developed to maximize reliance on different reading mappings. Automaticity, particularly in a task stressing meaning, predicted reading fluency over and above knowledge of the relevant grapheme-phoneme mappings. Automaticity in tasks involving nonwords also predicted fluency, suggesting the possibility of automaticity in orthography to phonology mappings. Variation in automatic word recognition did not predict reading comprehension or decoding. This link between automatic written word recognition and fluency has important implications for how automaticity may be targeted to improve reading outcomes. (PsycINFO Database Record (c) 2019 APA, all rights reserved)

Parietotemporal Stimulation Affects Acquisition of Novel Grapheme-Phoneme Mappings in Adult Readers.

Neuroimaging work from developmental and reading intervention research has suggested a cause of reading failure may be lack of engagement of parietotemporal cortex during initial acquisition of grapheme-phoneme (letter-sound) mappings. Parietotemporal activation increases following grapheme-phoneme learning and successful reading intervention. Further, stimulation of parietotemporal cortex improves reading skill in lower ability adults. However, it is unclear whether these improvements following stimulation are due to enhanced grapheme-phoneme mapping abilities. To test this hypothesis, we used transcranial direct current stimulation (tDCS) to manipulate parietotemporal function in adult readers as they learned a novel artificial orthography with new grapheme-phoneme mappings. Participants received real or sham stimulation to the left inferior parietal lobe (L IPL) for 20 min before training. They received explicit training over the course of 3 days on 10 novel words each day. Learning of the artificial orthography was assessed at a pre-training baseline session, the end of each of the three training sessions, an immediate post-training session and a delayed post-training session about 4 weeks after training. Stimulation interacted with baseline reading skill to affect learning of trained words and transfer to untrained words. Lower skill readers showed better acquisition, whereas higher skill readers showed worse acquisition, when training was paired with real stimulation, as compared to readers who received sham stimulation. However, readers of all skill levels showed better maintenance of trained material following parietotemporal stimulation, indicating a differential effect of stimulation on initial learning and consolidation. Overall, these results indicate that parietotemporal stimulation can enhance learning of new grapheme-phoneme relationships in readers with lower reading skill. Yet, while parietotemporal function is critical to new learning, its role in continued reading improvement likely changes as readers progress in skill.

Orthographic neighborhood effects in recognition and recall tasks in a transparent orthography.

The number of orthographic neighbors of a word influences its probability of being retrieved in recognition and free recall memory tests. Even though this phenomenon is well demonstrated for English words, it has yet to be demonstrated for languages with more predictable grapheme-phoneme mappings than English. To address this issue, 4 experiments were conducted to investigate effects of number of orthographic neighbors (N) and effects of frequency of occurrence of orthographic neighbors (NF) on memory retrieval of Brazilian Portuguese words. One hundred twenty-four Brazilian Portuguese speakers performed first a lexical-decision task (LDT) on words that were factorially manipulated according to N and NF, and intermixed with either nonpronounceable nonwords without orthographic neighbors (Experiments 1A and 2A), or with pronounceable nonwords with a large number of orthographic neighbors (Experiments 1B and 2B). The words were later used as probes on either recognition (Experiments 1A and 1B) or recall tests (Experiments 2A and 2B). Words with 1 orthographic neighbor were consistently better remembered than words with several orthographic neighbors in all recognition and recall tests. Notably, whereas in Experiment 1A higher false alarm rates were yielded for words with several rather than 1 orthographic neighbor, in Experiment 1B higher false alarm rates were yielded for words with 1 rather than several orthographic neighbors. Effects of NF, on the other hand, were not consistent among memory tasks. The effects of N on the recognition and recall tests conducted here are interpreted in light of dual process models of recognition. (PsycINFO Database Record

A functional investigation of RAN letters, digits, and objects: How similar are they?

Although rapid automatized naming (RAN) of letters, digits, and objects are popular tasks and have been used interchangeably to predict academic performance, it remains unknown if they tap into the same neural regions. Thus, the purpose of this study was to examine the neural overlap across different RAN tasks. Fifteen university students were assessed on RAN digits, letters, and objects using functional magnetic resonance imaging (fMRI). Results showed a common neural pattern that included regions related to motor planning (e.g., cerebellum), semantic access (middle temporal gyrus), articulation (supplementary motor association, motor/pre-motor, anterior cingulate cortex), and grapheme–phoneme mapping (ventral supramarginal gyrus). However, RAN digits and letters showed many unique regions of activation over and above RAN objects particularly in semantic and articulatory regions, including precuneus, bilateral supramarginal gyrus, nucleus accumbens and thalamus. The only region unique to RAN objects included bilateral fusiform, a region commonly implicated in object processing. Overall, our results provide the first neural evidence for a stronger relationship between RAN letters and digits than when either task is compared to RAN objects.

The Modulation of Visual and Task Characteristics of a Writing System on Hemispheric Lateralization in Visual Word Recognition—A Computational Exploration

Through computational modeling, here we examine whether visual and task characteristics of writing systems alone can account for lateralization differences in visual word recognition between different languages without assuming influence from left hemisphere (LH) lateralized language processes. We apply a hemispheric processing model of face recognition to visual word recognition; the model implements a theory of hemispheric asymmetry in perception that posits low spatial frequency biases in the right hemisphere and high spatial frequency (HSF) biases in the LH. We show two factors that can influence lateralization: (a) Visual similarity among words: The more similar the words in the lexicon look visually, the more HSF/LH processing is required to distinguish them, and (b) Requirement to decompose words into graphemes for grapheme-phoneme mapping: Alphabetic reading (involving grapheme-phoneme conversion) requires more HSF/LH processing than logographic reading (no grapheme-phoneme mapping). These factors may explain the difference in lateralization between English and Chinese orthographic processing.

Word count analysis of Malay language textbooks for the purpose of developing a Malay reading remedial programme

Abstract Cross-linguistic studies showed that young readers are sensitive to the orthographic and phonological structures of the languages in use. In our attempt to develop a Malay reading remedial programme, we conducted a word count analysis on Year 1 and Year 2 Malay textbooks. The results provided empirical evidence to describe the orthographic and phonological features of Malay, which functions to inform the cross-linguistic adaptations needed for a Malay-reading remedial programme. The results revealed that words with single-letter grapheme are common, and grapheme–phoneme mappings are transparent in the Malay language. However, different from English, monosyllabic words are few in the Malay language. The under-representation of monosyllabic words necessitates the use of two-syllabic words as the primary set of word stimuli for a reading remedial programme. It is also shown that the Malay language is more complex at the syllabic and morphological levels, which we argue will affect word reading acqui...

Learning to read aligns visual analytical skills with grapheme-phoneme mapping: evidence from illiterates

Learning to read puts evolutionary established speech and visual object recognition functions to novel use. As we previously showed, this leads to particular rearrangements and differentiations in these functions, for instance the habitual preference for holistic perceptual organization in visual object recognition and its suppression in perceiving letters. We performed the experiment in which the differentiation between holistic non-letter processing and analytic letter processing in literates was originally shown (van Leeuwen and Lachmann, 2004) with illiterate adults. The original differentiation is absent in illiterates; they uniformly showed analytic perception for both letters and non-letters. The result implies that analytic visual perception is not a secondary development resulting from learning to read but, rather, a primary mode of perceptual organization on a par with holistic perception.

Orthographic Traces in Romanian and Japanese Loanwords: Enriching Phonological Representations

This paper presents a formal account of the influence of orthography in the adaptation of Romanian loanwords from French and Japanese loanwords from English. It agues that, in the course of adaptation, the accompanying presence of a written representation does play a part in shaping the phonological content of borrowed words. To explain such orthographic manifestations in loanwords, a grammatical mechanism is devised in which underlying input representations are composed of linguistic information emanating from both the native perceptual system and the grapheme-phoneme mapping procedure. Cast in the framework of Optimality Theory (Prince and Smolensky, 1993/2004), the bulk of the analysis rests in determining how the grammar evaluates output forms resulting from such amalgamated inputs. Theoretical implications of such a proposal are also discussed, in particular as it concerns the nature of input coding and representation. In short, phonological representations are assumed to embrace the segmental richness imparted by both speech and print.

Masked onset priming effect for high-frequency words: Further support for the speech-planning account

Two experiments investigated whether priming due to a match in just the onset between a masked prime and target is found with high-frequency target words. Forster and Davis (1991, Exp. 5) reported that the masked onset priming effect was absent for high-frequency words and used the finding to argue that the effect has its locus in the grapheme-phoneme mapping process that operates serially within the nonlexical route. Experiment 1 used primes that were unrelated to targets and found a masked onset priming effect of equal size for high-frequency and low-frequency target words. Experiment 2 used form-related primes as used by Forster and Davis, and again found that the effect of onset mismatch was not dependent on target word frequency. These results are interpreted in terms of an alternative view that the masked onset priming effect has its origin in the process of preparing a speech response.