Exploring orthographic neighborhood size effects in a computational model of Chinese character naming

Highlights

• Effects of orthographic neighborhood size in Chinese character naming were examined.

• Inconsistent characters with large neighborhood size inhibited the naming latencies.

• Two PDP models of Chinese character naming were developed to explore the effect.

• Consistency and neighborhood size interact with the division of labor in the system.

• The semantic pathway contributes to the inhibitory effect of neighborhood size.

Abstract

Orthographic neighborhood (N) size effects have been extensively studied in English consistently producing a facilitatory effect in word naming tasks. In contrast, several recent studies on Chinese character naming have demonstrated an inhibitory effect of neighborhood size. Response latencies tend to be inhibited by inconsistent characters with large neighborhoods relative to small neighborhoods. These differences in neighborhood effects between languages may depend on the characteristics (depth) of the mapping between orthography and phonology. To explore this, we first conducted a behavioral experiment to investigate the relationship between neighborhood size, consistency and reading response. The results showed an inhibitory effect of neighborhood size for inconsistent characters but a facilitatory effect for consistent characters. We then developed two computational models based on parallel distributed processing principles to try and capture the nature of the processing that leads to these results in Chinese character naming. Simulations using models based on the triangle model of reading indicated that consistency and neighborhood size interact with the division of labor between semantics and phonology to produce these effects.

Introduction

Orthographic neighborhood size is one of the key lexical variables that affect word response latencies during visual word recognition. The most widely used measure of neighborhood size (denoted by the statistic of N) is defined as the number of words that could be created by changing one letter in a target word (Coltheart, Davelaar, Jonasson, & Besner, 1977). For example, dog has a number of orthographic neighbors such as jog, dot, dig, log, and doe. Many studies have explored the effects of orthographic neighborhood across a range of tasks including naming and lexical decision (Andrews, 1989, Andrews, 1992, Balota et al., 2004, Carreiras et al., 1997, Coltheart et al., 1977, Forster and Shen, 1996, Grainger, 1990, Sears et al., 1995). While the focus of this paper is naming, it is useful to consider both lexical decision and naming data together as these can constrain the theoretical explanations of the neighborhood effect. In lexical decision, the findings concerning the neighborhood size effect appear to be somewhat mixed (Balota et al., 2004). Andrews, 1989, Andrews, 1992 reported a facilitatory effect of neighborhood size in the lexical decision task, in particular for low frequency words. However, other studies have found that response latencies for words having high frequency neighbors tend to be prolonged in comparison with words having low frequency neighbors (Carreiras et al., 1997, Grainger, 1990, Grainger et al., 1989). This has been referred to as the neighborhood frequency effect. Despite this, facilitation in lexical decision has been reported when both neighborhood size and neighborhood frequency are considered in the same experiment (Forster and Shen, 1996, Sears et al., 1995). Balota et al. (2004) examined the neighborhood size effect by conducting multiple regression analyses on a group of younger readers and another group of older readers in both naming and lexical decision tasks. They showed that younger readers’ lexical decision performance was facilitated by neighborhood size particularly for low frequency words, which is consistent with Andrews, 1989, Andrews, 1992. However, in older and slower readers the lexical decision performance was inhibited by neighborhood size. These results suggest that the neighborhood size effect in lexical decision may depend on decision strategies and the processing speed of the subjects.

The effect of neighborhood size in naming tasks is much more consistent, showing a robust facilitatory effect across many studies (Andrews, 1989, Andrews, 1992, Balota et al., 2004, Carreiras et al., 1997, Grainger, 1990, Peereman and Content, 1995, Peereman and Content, 1997, Sears et al., 1995), particularly when words are low in frequency. This has been supported by studies using either a factorial design (Andrews, 1989, Andrews, 1992, Sears et al., 1995) or a regression technique (Balota et al., 2004); and the effect also has been found in different alphabetic languages such as Dutch (Grainger, 1990), French (Peereman & Content, 1995) and Spanish (Carreiras et al., 1997).

One interpretation of the neighborhood size effect proposed by Andrews (1989) is based on the interactive activation theory of word recognition (McClelland & Rumelhart, 1981). Since a target word and its neighbors only differ in one letter, when the target word is presented, the word nodes for the neighboring words would be activated early in processing along with the target word. The activations in turn feedback to facilitate the activations of the constituent letter nodes. The feedback activations are particularly helpful for naming low frequency words. On this view, the facilitation results from lexical contribution to the orthographic activation, although as argued by Peereman and Content (1995), lexical activation of neighbors could contribute to phonological computation rather than orthographic processing. One problem with this account is that it would predict the same facilitation for lexical decision as naming, but as we have indicated the data for lexical decision is much more complex.

An alternative hypothesis is that the neighborhood size effect is related to phonological computation (Peereman and Content, 1995, Peereman and Content, 1997). According to this view, the effect is not limited to orthographic processing; rather it can be attributed to the variability of phonological properties among orthographic neighbors. Evidence for this view comes from a study by Peereman and Content (1997), in which they examined the influence of different types of orthographic and phonological neighbors on naming. The results showed that when the orthographic neighbors were also phonological neighbors (i.e., they are phonographic neighbors), the facilitation in naming was the strongest compared with other types of neighbors. Thus, they argued that the extent to which orthographic neighborhood size could accelerate phonological computation is dependent on the similarity between the phonological codes of neighboring words and the target word. This finding is corroborated by the results reported in a multiple regression study on four large English naming datasets (Adelman & Brown, 2007), where the number of phonographic neighbors was a stronger predictor than the conventional neighborhood size in accounting for naming data.

The phonological computation account of the neighborhood size effect is supported by most current theories of reading (Adelman & Brown, 2007). According to the dual-route cascade (DRC) models (Coltheart et al., 2001, Perry et al., 2007), the facilitatory effect in naming is expected because the models allow the processing to activate orthographic neighbors of word stimuli in the orthographic lexicon, which in turn activates phonological entries and phonemes. Phonetic activation generated from the lexical route along with that generated from the non-lexical route would speed the naming latencies. Within the parallel distributed processing (PDP) models (Chang et al., 2012, Harm and Seidenberg, 2004, Plaut et al., 1996, Seidenberg and McClelland, 1989) the effect emerges as the system adjusts its connection weights following exposure to the shared orthographic structure of the neighboring words.

While the facilitatory effect of neighborhood size in naming is robust in alphabetic languages, recent studies in Chinese have showed a contradictory pattern of neighborhood size effects (Li et al., 2011, Zhao et al., 2012), where the orthographic neighbors of phonetic radicals tend to increase naming latencies. To our knowledge, there are no studies in alphabetic languages reporting an inhibitory effect of neighborhood size in naming; this effect seems to be reported only in studies based on Chinese characters. In Chinese, over 80% of characters are phonograms, which consist of a semantic radical (usually on the left) and a phonetic radical (usually on the right) (Zhou, 1978). In general the semantic radical provides some information relating to meaning, while the phonetic radical provides some information about pronunciation. The neighborhood size of phonetic radicals is defined as the number of characters that share the same phonetic radical. Two relevant measures are the orthography-to-phonology consistency of a character, indicating whether the pronunciation of a character agrees with other characters containing the same phonetic radical, and regularity, which is defined as whether a character is pronounced the same as its phonetic radical under the constraint that the phonetic radical is pronounceable (Fang, Horng, & Tzeng, 1986). These definitions are based on similar concepts to those used in English (Coltheart, 1978, Glushko, 1979). Despite the fact that Chinese has a very different orthographic system from alphabetic languages, most typical reading effects such as frequency effects (Balota et al., 2004, Forster and Chambers, 1973, Hue, 1992, Lee et al., 2005) and regularity or consistency effects (Glushko, 1979, Lee et al., 2005, Taraban and McClelland, 1987) tend to have a similar pattern across English and Chinese, and those effects also have been simulated by computational models based on the same general learning principles in Chinese (Hsiao and Shillcock, 2004, Yang et al., 2009). It remains unclear how a more language-specific effect (i.e., the inhibitory effect of neighborhood size) seen in Chinese emerges in the reading system.

On the basis of the orthographic structures of Chinese phonograms, two different types of orthographic neighbors can be defined: semantic radical neighbors and phonetic radical neighbors (Feldman & Siok, 1999). Of particular interest here is the phonetic radical neighborhood size, also known as phonetic combinability, because it is directly linked to phonology. Throughout this paper we will use orthographic neighborhood size in Chinese to refer to the neighborhood size of phonetic radicals, unless stated otherwise.

Several studies in Chinese character reading have examined the effects of orthographic neighborhood size and consistency simultaneously because they are closely related to phonetic radicals (Hsu et al., 2014, Hsu et al., 2009, Li et al., 2011, Zhao et al., 2012). Li et al. (2011) found an inhibitory effect of neighborhood size for inconsistent characters while a null effect was observed for consistent characters. However, when the high frequency neighboring characters in the inconsistent condition were removed, the effect became facilitatory. They suggested that the neighboring characters of a target character might accelerate activation at the orthographic level; however, any high frequency neighbors of the target would cause interference at the phonological processing stage, resulting in an inhibitory effect. This interpretation is partly consistent with the phonological computation account (Adelman and Brown, 2007, Peereman and Content, 1997), suggesting that the effect of orthographic neighborhood size is not limited to the orthographic level but it is also dependent on the stage of phonological processing. Further evidence for this comes from a study of event-related potentials (ERPs) conducted by Hsu et al. (2009). They examined the effects of orthographic neighborhood size and consistency in Chinese character reading. They demonstrated that characters with large neighborhoods facilitated the earlier stages of orthographic (N170) and phonological processing (P200) relative to characters with small neighborhoods. They also elicited larger negativity at the later stage of semantic processing (N400), suggesting an increase of semantic competition for high neighborhood characters. Their results suggest the effect of orthographic neighborhood size is widespread throughout the reading system.

According to the phonological computation account (Peereman & Content, 1997), one might expect a similar effect of neighborhood size in both English and Chinese, if considering the orthographic activation of phonology alone. However, as agreed by most current theories of reading, there are two different pathways active during reading (Coltheart et al., 2001, Plaut et al., 1996). Within the PDP models of reading (Harm and Seidenberg, 2004, Plaut et al., 1996) there is a phonological pathway from orthography to phonology and a semantic pathway from orthography to phonology via semantics. The division of labor between pathways is greatly shaped by the nature of the orthographic systems (Yang, Shu, McCandliss, & Zevin, 2013). In English, the mappings between orthography and phonology are mostly consistent, which contrasts with the arbitrary mappings between orthography and semantics. So learning the mappings in the phonological pathway is much faster and the connection weights are optimized for the consistent spelling-to-sound mappings. Thus, consistent words can utilize the phonological pathway very efficiently for their pronunciations. While inconsistent words can utilize the phonological pathway, at the same time they may also partly rely on the semantic pathway for their pronunciations (Plaut et al., 1996).

Chinese has less transparent mappings between orthography and phonology but more regular mappings between orthography and semantics compared with those in English. Moreover, in English, inconsistent words still have many subcomponents that are shared among words. For instance, the pronunciation of an inconsistent word, pint, can benefit from the pronunciation of pant because they share the same onset and coda and it is only the vowel section that is inconsistent. In Chinese most inconsistent characters do not share any phonetic components with their phonetic radicals (e.g., 灑 /sa3/ – 麗 /li4/). However, some characters may share either onset or rime. For example, the inconsistent character 結 /jie2/ shares the same onset (i.e. /j/) with its phonetic radical 吉 /ji2/ while the inconsistent character 妒 /du4/ shares the same rime (i.e. /u/) with its phonetic radical 戶 /hu4/. However, in Chinese this partial information derived from the phonetic radical is not very helpful in determining the pronunciation of inconsistent characters (Chen, Shu, Wu, & Anderson, 2003). Overall, these properties might suggest that the semantic pathway would play a more important role in Chinese character naming than it does in English. This might explain the inhibitory effect of neighborhood size seen in inconsistent Chinese characters (Li et al., 2011). Specifically, a performance cost observed for inconsistent words with large neighborhood may be due to the conflict between orthographic activation of phonology from the phonological pathway and semantic activation of phonology from the semantic pathway.

Although a number of theoretical models of Chinese reading have been proposed in the literature (Perfetti et al., 2005, Perfetti and Tan, 1999, Taft and Zhu, 1997), only recently have large-scale computational models been developed (Hsiao and Shillcock, 2004, Yang et al., 2009, Yang et al., 2013). These provide explicit details about the connections between the core processing layers within the system and allow for effective evaluation of reading effects. In particular, a recent parallel distributed processing model of Chinese character naming by Yang et al. (2009) has demonstrated that the same statistical learning principles can be applied to both English and Chinese. They developed a computational model of Chinese character naming on the basis of the previous models in English (Harm & Seidenberg, 1999), with revised representations to represent Chinese orthography and phonology. The model was able to capture the pattern of interaction between frequency and consistency seen in skilled Chinese readers (Hue, 1992, Lee et al., 2005). Analyses of internal representations revealed that phonetic radicals emerged as critical processing units over learning. These findings show that the processing of orthography-to-phonology conversion for Chinese has strong similarities to English. However, the relationship between neighborhood size of phonetic radicals and consistency has not been investigated in their model.

The aim of the present study was to develop computational models based on the parallel distributed processing framework and to test whether the division of labor between phonological and semantic pathways is the key to accounting for the inhibitory effect of neighborhood size observed in Chinese character naming. We investigated this by conducting a behavioral naming task in order to replicate previous results of Li et al. (2011). We then developed two computational models of Chinese character naming based on previous models (Chang et al., 2012, Plaut et al., 1996, Yang et al., 2013): one contained only the phonological pathway from orthography to phonology and the other one contained both the phonological and semantic pathways by providing both orthographic and semantic inputs to phonology. We also incorporated a visual processing stage into models, which allowed orthographic representations to be developed over the course of learning (Chang et al., 2012). We expected that both models would account for the typical interaction between frequency and consistency shown in the previous Chinese reading models (Hsiao and Shillcock, 2004, Yang et al., 2009). However, if the semantic pathway contributes to the emergence of the inhibitory effect of neighborhood size, we would expect that only the model including a semantic pathway would show the inhibitory effect of phonetic neighbors as seen in skilled Chinese readers.