Models Of Word Recognition Research Articles

Two streams deep neural network for handwriting word recognition

Handwritten word recognition is one of the hot topics in automatic handwritten text recognition that received a lot of attention in recent years. Unlike character recognition, word recognition deals with considerable variations in word shape and written style. This paper proposes a novel deep model for language-independent handwritten word recognition. The proposed deep structure has two parallel stages for jointly learning character and word-level information. In the character-level stage, a weakly character segmentation method is performed and then applies a series of Long short-term memory (LSTM) layers for character-level representation. The word-level stage employs a series of convolutional layers for the shape and structure representation of the word. These representations are then concatenated and followed by a series of fully connected layers for jointly learning the words and the character-level information. Since the character segmentation is language independent and error-prone, the proposed deep structure only applies weakly separation scheme and does not rely on any character segmentation algorithm. Thus, it effectively utilizes character level representation without bounding on any language model. In the proposed methodology, we use two new data augmentation strategies based on a psychological assumption to increase the model generalization performance. Experimental results on five public datasets including Arabic, English and German languages demonstrate that the proposed deep model has a superior performance to the state-of-the-art methods.

Syllable effects in beginning and intermediate European-Portuguese readers: Evidence from a sandwich masked go/no-go lexical decision task.

Reading is one of the most important milestones a child achieves throughout development. Above the letter level, the syllable has been shown to play a relevant role at early stages of visual word recognition in adult skilled readers. However, studies aiming to examine when, during reading acquisition, the syllable emerges as a functional sublexical unit are scarce, and the studies conducted so far have led to inconsistent results. In this work, beginning and intermediate European-Portuguese (EP) developing readers performed a sandwich masked lexical decision task in which CV (e.g., RU.MOR[rumour]) and CVC (e.g., CIS.NE[swan]) first-syllable EP words were preceded either by syllable congruent (e.g., rum.ba-RU.MOR, cis.ra-CIS.NE), syllable incongruent (e.g., rum.ba-RU.MOR, ci.ser-CIS.NE), unrelated (e.g., va.cra-RU.MOR, zar.vo-CIS.NE) pseudowords primes, or identity (e.g., ru.mour-RU.MOUR, cis.ne-CIS.NE) primes. Results showed reliable syllable effects only for intermediate readers and for CV and CVC words alike. Findings are discussed attending to current models of visual word recognition.

The role of letter features on the consonant-bias effect: Evidence from masked priming

Recent research has shown an advantage of consonants at early stages of visual word recognition (C-bias), although the locus of this effect remains elusive. Here we examine whether the C-bias is affected by the consonant letters' features. Skilled readers performed a masked priming lexical decision task in which target words containing only either consonants without any ascending/descending features (flat words, canino[canine]) or consonants with ascending/descending features (non-flat words, palito[toothpick]) were preceded by briefly (50 ms) presented primes that could preserve the same consonants of the targets (cenune-CANINO, pelute-PALITO), the same vowels of the targets (raxizo-CANINO, fajibo-PALITO), or, as controls, unrelated (ruxuze-CANINO, fejube-PALITO) and identity primes (canino-CANINO, palito-PALITO). The case in which prime-target pairs were presented was also manipulated (lower-upper vs. upper-lower). Results showed that in both case conditions flat words were recognized faster than non-flat words. Evidence for the C-bias was observed both for flat and non-flat words in the lower-upper condition, in which a vowel inhibitory priming effect was also observed for non-flat words. In the upper-lower condition, however, the C-bias was restricted to flat words. These findings suggest that letter features play a role in the C-bias and ask for amendments in current models of visual word recognition.

The temporal dynamics of first and second language processing: ERPs to spoken words in Mandarin-English bilinguals

The dynamics of bilingual spoken word recognition remain poorly characterized, especially for individuals who speak two languages that are highly dissimilar in their phonological and morphological structure. The present study compared first language (L1) and second language (L2) spoken word processing within a group of adult Mandarin-English bilinguals (N = 34; ages 18–25). Event-related potentials (ERPs) were recorded while participants completed the same cross-modal matching task separately in their L1 Mandarin and L2 English. This task consisted of deciding whether spoken words matched pictures of items. Pictures and spoken words either matched (e.g., Mandarin: TANG2-tang2; English: BELL-bell), or differed in word-initial phonemes (e.g., Mandarin: TANG2-lang2; English: BELL-shell), word-final phonemes (e.g., Mandarin: TANG2-tao2; English: BELL-bed), or whole words (e.g., Mandarin: TANG2-xia1: English: BELL-ham). Each mismatch type was associated with a pattern of modulation of the Phonological Mapping Negativity, the N400, and the Late N400 that was distinct from those of the other mismatch types yet similar between the two languages. This was interpreted as evidence of incremental processing with similar temporal dynamics in both languages. These findings support models of spoken word recognition in bilingual individuals that adopt an interactive-activation framework for both L1 and L2 processing.

The time course of the lowercase advantage in visual word recognition: An ERP investigation

Previous word identification and sentence reading experiments have consistently shown faster reading for lowercase than for uppercase words (e.g., table faster than TABLE). A theoretically relevant question for neural models of word recognition is whether the effect of letter-case only affects the early prelexical stages of visual word recognition or whether it also influences lexical-semantic processing. To examine the locus and nature of the lowercase advantage in visual word recognition, we conducted an event-related potential (ERP) lexical decision experiment. ERPs were recorded to words and pseudowords presented in lowercase or uppercase. Words also varied in lexical frequency, thus allowing us to assess the time-course of perceptual (letter-case) and lexical-semantic (word-frequency) processing. Together with a lowercase advantage in word recognition times, results showed that letter-case influenced early perceptual components (N/P150), whereas word frequency influenced lexical-semantic components (N400). These findings are consistent with those models of written word recognition that assume that letter-case information from the visual input is quickly mapped onto the case-invariant letter and word units that drive lexical access.

Computational modelling of an auditory lexical decision experiment using jTRACE and TISK

ABSTRACT We present a series of computational simulations of the auditory lexical decision task using the jTRACE and TISK models of spoken word recognition. Simulation 1 replicates high accuracy in word recognition and similar performance of these models using the small, default dictionary. Simulation 2 expands the set of words and phonemes, leading to issues in representing certain phonemes in jTRACE. Simulation 3 expands the lexicon of competitors and we find that TISK struggles to select the target word as the winner. Finally, Simulation 4 shows that the decision criteria employed leads to many false positives when pseudowords are presented to the model. None of the model estimates of the time cycle when the winner should be selected predicted participant response latency in the auditory lexical decision task. We discuss these findings and offer suggestions as to what a contemporary model of spoken word recognition should be able to do.

Jalapeno or jalapeño: Do diacritics in consonant letters modulate visual similarity effects during word recognition?

AbstractPrior research has shown that word identification times to DENTIST are faster when briefly preceded by a visually similar prime (dentjst; i↔j) than when preceded by a visually dissimilar prime (dentgst). However, these effects of visual similarity do not occur in the Arabic alphabet when the critical letter differs in the diacritical signs: for the target the visually similar one-letter replaced prime (compare and is no more effective than the visually dissimilar one-letter replaced prime Here we examined whether this dissociative pattern is due to the special role of diacritics during word processing. We conducted a masked priming lexical decision experiment in Spanish using target words containing one of two consonants that only differed in the presence/absence of a diacritical sign: n and ñ. The prime-target conditions were identity, visually similar, and visually dissimilar. Results showed an advantage of the visually similar over the visually dissimilar condition for muñeca-type words (muneca-MUÑECA < museca-MUÑECA), but not for moneda-type words (moñeda-MONEDA = moseda-MONEDA). Thus, diacritical signs are salient elements that play a special role during the first moments of processing, thus constraining the interplay between the “feature” and “letter” levels in models of visual word recognition.

The Effect of the Non-task Language When Trilingual People Use Two Languages in a Language Switching Experiment.

This study investigated the effect of non-task language in a language switching experiment. Non-task language refers to participants’ languages (regardless of proficiency level) that are not used in any trials throughout the experiment. We recruited 60 Tibetan-Chinese-English trilinguals (12th-grade high school students with a median age of 17) to perform a lexical decision (word vs. non-word) task in only two of their languages. We repeated the experiment three times to present each language pair once. In each experiment, the participants were divided into two groups that significantly contrasted with each other in their non-task language while remaining comparable in the two task languages. Response time (RT) and error rate (ER) have been examined to evaluate task performance. The interaction between task performance and the participants’ proficiency in the non-task language was also examined. The results showed anull effect of language switching. In addition, the effect of the non-task language was not found. These results were interpreted with reference to the main models of bilingual visual word recognition and the role of orthography specificity.

Measuring Lexical Quality: The Role of Spelling Ability.

The construct of 'lexical quality' (Perfetti Scientific Studies of Reading 11, 357-383, 2007) is widely invoked in literature on word recognition and reading to refer to a systematic dimension of individual differences that predicts performance in a range of word identification and reading tasks in both developing readers and skilled adult populations. Many different approaches have been used to assess lexical quality, but few have captured the orthographic precision that is central to the construct. This paper describes, evaluates, and disseminates spelling dictation and spelling recognition tests that were developed to provide sensitive measures of the precision component of lexical quality in skilled college student readers - the population that has provided most of the benchmark data for models of word recognition and reading. Analyses are reported for 785 students who completed the spelling tests in conjunction with standardized measures of reading comprehension, vocabulary, and reading speed, of whom 107 also completed author recognition and phonemic decoding tests. Internal consistency analyses showed that both spelling tests were relatively unidimensional and displayed good internal consistency, although the recognition test contained too many easy items. Item-level analyses are included to provide the basis for further refinement of these instruments. The spelling tests were moderately correlated with the other measures of written language proficiency, but factor analyses revealed that they consistently defined a separate component, demonstrating that they tap a dimension of variability that is partially independent of variance in reading comprehension, speed, and vocabulary. These components appear to align with the precision and coherence dimensions of lexical quality.

The Role of Surface and Underlying Forms When Processing Tonal Alternations in Mandarin Chinese: A Mismatch Negativity Study.

Phonological alternation (sound change depending on the phonological environment) poses challenges to spoken word recognition models. Mandarin Chinese T3 sandhi is such a phenomenon in which a tone 3 (T3) changes into a tone 2 (T2) when followed by another T3. In a mismatch negativity (MMN) study examining Mandarin Chinese T3 sandhi, participants passively listened to either a T2 word [tʂu2 je4] /tʂu2 je4/, a T3 word [tʂu3 je4] /tʂu3 je4/, a sandhi word [tʂu2 jen3] /tʂu3 jen3/, or a mix of T3 and sandhi word standards. The deviant in each condition was a T2 word [tʂu2]. Results showed an MMN only in the T2 and T3 conditions but not in the Sandhi or Mix conditions. All conditions also yielded omission MMNs. This pattern cannot be explained based on the surface forms of standards and deviants; rather these data suggest an underspecified or underlying T3 stored linguistic representation used in spoken word processing.

The origins of backward priming effects in logographic scripts for four-character words

Abstract Yang, Chen, Spinelli, and Lupker (2019) reported a large masked priming effect in a Chinese lexical-decision task using prime-target pairs in which the primes were presented in a backward (right-to-left) orientation (e.g., 说来的总-总的来说) (a “backward” priming effect). The question addressed here is whether this effect is truly an orthographic priming effect or is, to some degree, morphologically/meaning- or syllabic/phonologically-based. Five experiments, two involving phonologically-related primes and three involving meaning-related primes, produced no evidence that either of those factors contributed to the backward priming effect, implying that it truly is an orthographic effect. As backward priming effects do not emerge in English, these results suggest that the orthographic coding process is quite different for Chinese versus English readers. Specifically, they support the conclusion that the orthographic coding process for Chinese readers codes character positions in a quite flexible fashion. Issues concerning the generalizability of current models of orthographic coding in alphabetic languages, as well as implications for models of Chinese word recognition, are discussed.

Semantic priming effects can be modulated by crosslinguistic interactions during second-language auditory word recognition

AbstractThe current study investigates how second languageauditoryword recognition, in early and highly proficient Spanish–Basque (L1-L2) bilinguals, is influenced by crosslinguistic phonological-lexical interactions and semantic priming. Phonological overlap between a word and its translation equivalent (phonological cognate status), and semantic relatedness of a preceding prime were manipulated. Experiment 1 examined word recognition performance in noisy listening conditions that introduce a high degree of uncertainty, whereas Experiment 2 employed clear listening conditions, with low uncertainty. Under noisy listening conditions, semantic priming effects interacted with phonological cognate status: for word recognition accuracy, a related prime overcame inhibitory effects of phonological overlap between target words and their translations. These findings are consistent with models of bilingual word recognition that incorporate crosslinguistic phonological-lexical-semantic interactions. Moreover, they suggest an interplay between L2-L1 interactions and the integration of information across acoustic and semantic levels of processing in flexibly mapping the speech signal onto the spoken words, under adverse listening conditions.

When does rotation disrupt letter encoding? Testing the resilience of letter detectors in the initial moments of processing.

Previous research has reported that both letter and word identification are slower when the stimuli are presented at rotations above 45° than when presented in their canonical horizontal view. Indeed, influential models of word recognition posit that letter detectors in the visual word recognition system are disrupted by rotation angles above 40° or 45° (e.g., Local Combinations Detector model; Dehaene, Cohen, Sigman, & Vinckier, 2005, Trends in Cognitive Sciences). However, recent experiments have shown robust masked identity/form priming effects for 90° rotated words, thus calling into question this assumption. Here we aimed to isolate the degree to which letter detectors are disrupted when manipulating letter rotation in three masked identity priming letter match experiments. Probes and targets were always presented in the canonical upright position, whereas forwardly masked primes were rotated in different angles. The rotation angles were 0° versus 45° (Experiment 1), 22.5° versus 67.5° (Experiment 2), and 45° versus 90° (Experiment 3). Results showed a sizeable masked identity priming effect regardless of the rotation angle, hence demonstrating that letter detectors are not disrupted by rotations smaller than 90° in the early moments of letter processing. This pattern suggests that letter detectors are more resilient to changes in visual form than predicted by the LCD model.

Context Effect on L2 Word Recognition: Visual Versus Auditory Modalities.

This study tested the effect of constraining sentence context on word recognition time (RT) in the first and second language. Native (L1) and nonnative (L2) speakers of English performed self-paced reading and listening tasks to see whether a semantically-rich preceding context would lead to the activation of a probable upcoming word prior to encountering it. The pre-access prediction model (e.g., Altmann and Kamide in Cognition 73(3):247-264, 1999; McClelland and Rumelhart in Psychol Rev 88:375-407, 1981) posits that when the preceding context is semantically high-constraining, the perceptual system anticipates a probable upcoming word prior to encountering it. In contrast, the post-access model (e.g., Fodor in The modularity of mind, MIT Press, Cambridge, 1983; Forster in Q J Exp Psychol Sect A 33(4):465-495, 1981; Traxler et al. in J Psycholinguist Res 29(6):581-595, 2000; Van Petten and Luka in Int J Psychophysiol 83(2): 176-190, 2012) suggests that it is only after a word is encountered that a subsequent process integrates it into the preceding context. The integration process is easier and faster when the word is more congruent with the preceding context. In line with these two models of visual word recognition, auditory word recognition is modeled by the TRACE model (McClelland and Elman in Cogn Psychol 18(1):1-86, 1986), which suggests that a spoken word is influenced by the preceding sentence context, whereas models such as the cohort model (Marslen-Wilson in Cognition 25(1-2):71-102, 1987) support a post-access integration process. Some studies observe a facilitative effect of sentence context on L2 word recognition (e.g., Kamide et al. in J Mem Lang 49(1):133-156, 2003), while others find no effect of sentence context in L2 (e.g., Ito et al. in Lang Cogn Neurosci 32(8):954-965, 2017; Martin et al. in J Mem Lang 69(4):574-588, 2013). In the present experiments, the RTs of native English speakers (L1) and non-native English speakers (L2) were collected in both visual and auditory word recognition in semantically high-constraint sentences and semantically low-constraint sentences. A linear, mixed-effects model shows that both groups of participants are faster to recognize a word when it is preceded by a semantically high-constraining context. This result is observed in both the visual and the auditory modalities, lending some support for a mechanism facilitating the access of target words based on sentence context in both L1 and L2.

Zooming in on zooming out: Partial selectivity and dynamic tuning of bilingual language control during reading

Prominent models of bilingual visual word recognition posit a bottom-up nonselective view of lexical processing with parallel access to lexical candidates of both languages. However, these accounts do not accommodate recent findings of top-down effects on the relative global activation level of each language during bilingual reading. We conducted two eye-tracking experiments to systematically assess the degree of accessibility of each language in different global language contexts. When critical words were presented overtly in Experiment 1, code switches disrupted reading early during lexical processing, but not as much as pseudowords did. Participants zoomed out of the target language with increasing exposure to language switches. In Experiment 2, a monolingual language context was created by presenting critical words covertly as parafoveal previews. Here, code-switched words were treated like pseudowords, and participants remained zoomed in to the target language throughout the experiment. Switch direction analyses confirmed and extended these interpretations to provide further support for the role of global language control on lexical access, above and beyond effects due to proficiency differences across languages. Together, these data provide strong evidence for dynamic top-down adjustment of the degree of language selectivity during bilingual reading.

A tale of two frequency effects: Toward a verification model of L2 word recognition

AbstractThis study examined the activation of first language (L1) translations in second language (L2) word recognition in a lexical decision task. Test materials included English words that differed in the frequency of their Chinese translations or in their surface lexical frequency while other lexical properties were controlled. Chinese speakers of English as a second language of different proficiencies and native speakers of English were tested. Native speakers produced a reliable lexical frequency effect but no translation frequency effect. English as a second language speakers of lower English proficiency showed both a translation frequency effect and a lexical frequency effect, but those of higher English proficiency showed a lexical frequency effect only. The results were discussed in a verification model of L2 word recognition. According to the model, L2 word recognition entails a checking procedure in which activated L2 words are checked against their L1 translations. The two frequency effects are seen to have two different loci. The lexical frequency effect is associated with the initial activation of L2 lemmas, and the translation frequency effect arises in the verification process. Existing evidence for verification in L2 word recognition and new issues this model raises are discussed.

Longitudinal Speech Recognition in Noise in Children: Effects of Hearing Status and Vocabulary.

Objectives: The aims of the current study were: (1) to compare growth trajectories of speech recognition in noise for children with normal hearing (CNH) and children who are hard of hearing (CHH) and (2) to determine the effects of auditory access, vocabulary size, and working memory on growth trajectories of speech recognition in noise in CHH.Design: Participants included 290 children enrolled in a longitudinal study. Children received a comprehensive battery of measures annually, including speech recognition in noise, vocabulary, and working memory. We collected measures of unaided and aided hearing and daily hearing aid (HA) use to quantify aided auditory experience (i.e., HA dosage). We used a longitudinal regression framework to examine the trajectories of speech recognition in noise in CNH and CHH. To determine factors that were associated with growth trajectories for CHH, we used a longitudinal regression model in which the dependent variable was speech recognition in noise scores, and the independent variables were grade, maternal education level, age at confirmation of hearing loss, vocabulary scores, working memory scores, and HA dosage.Results: We found a significant effect of grade and hearing status. Older children and CNH showed stronger speech recognition in noise scores compared to younger children and CHH. The growth trajectories for both groups were parallel over time. For CHH, older age, stronger vocabulary skills, and greater average HA dosage supported speech recognition in noise.Conclusion: The current study is among the first to compare developmental growth rates in speech recognition for CHH and CNH. CHH demonstrated persistent deficits in speech recognition in noise out to age 11, with no evidence of convergence or divergence between groups. These trends highlight the need to provide support for children with all degrees of hearing loss in the academic setting as they transition into secondary grades. The results also elucidate factors that influence growth trajectories for speech recognition in noise for children; stronger vocabulary skills and higher HA dosage supported speech recognition in degraded situations. This knowledge helps us to develop a more comprehensive model of spoken word recognition in children.

Multiple routes to word recognition: evidence from event-related potentials.

We used event-related potentials to determine whether lexical access during semantic processing is achieved solely by the letter-based route, or by both a letter-based and word-based route. Participants determined whether words were related or unrelated to a prespecified category. To disrupt the word-based route (i.e., disrupt the processing of overall word shape), we manipulated case type. We measured the N170, assumed to be an index of holistic processing, and the N400, an index of semantic activation. Surprisingly, mixed-case words elicited a larger N170 effect than either consistent lowercase words (Experiment 1) or consistent uppercase words (Experiment 2). The N400, meanwhile, was unaffected by case mixing. In contrast, LEET words (e.g., T4BL3 instead of TABLE), which preserve overall word shape but distort letter shape, increased the N400 but did not reduce the N170 (Experiment 3). The results indicate that the N170 is in fact not a reliable index of holistic word processing. Implications for word recognition models are discussed.

Bilinguals apply language-specific grain sizes during sentence reading

Languages differ in the consistency with which they map orthography to phonology, and a large body of work now shows that orthographic consistency determines the style of word decoding in monolinguals. Here, we characterise word decoding in bilinguals whose two languages differ in orthographic consistency, assessing whether they maintain two distinct reading styles or settle on a single ‘compromise’ reading style. In Experiment 1, Welsh-English bilinguals read cognates and pseudowords embedded in Welsh and English sentences. Eye-movements revealed that bilinguals dynamically alter their decoding strategy according to the language context, including more fixations during lexical access for cognates in the more consistent orthography (Welsh) than in the less consistent orthography (English), and these effects were specific to word (as opposed to pseudoword) processing. In Experiment 2, we compared the same bilinguals’ eye movements in the English sentence reading context to those of monolinguals’. Bilinguals’ eye-movement behaviour was very similar to monolinguals’ when reading English, suggesting that their knowledge of the more consistent orthography (Welsh) did not alter their decoding style when reading in English. This study presents the first characterisation of bilingual decoding style in sentence reading. We discuss our findings in relation to connectionist reading models and models of bilingual visual word recognition.

Speech predictability can hinder communication in difficult listening conditions

In difficult listening situations, such as in noisy environments, one would expect speech intelligibility to improve over time thanks to noise adaptation and/or to speech predictability facilitating the recognition of upcoming words. We tested this possibility by presenting normal-hearing human listeners (N = 100; 70 women) with sentences and measuring word recognition as a function of word position in a sentence. Sentences were presented in quiet and in competition with various masker sounds at individualized levels where listeners had 50% probability of recognizing a full sentence. Contrary to expectations, recognition was best for the first word and gradually deteriorated with increasing word position along the sentence. The worsening in recognition was unlikely due to differences in word audibility or word type and was uncorrelated with age or working memory capacity. Using a probabilistic model of word recognition, we show that the worsening effect probably occurs because misunderstandings generate inaccurate predictions that outweigh the benefits from accurate predictions. Analyses also revealed that predictions overruled the potential benefits from noise adaptation. We conclude that although speech predictability can facilitate sentence recognition, it can also result in declines in word recognition as the sentence unfolds because of inaccuracies in prediction.