The Effects of Semantic Context and Phonological Similarity on Bilingual Lexical Access Thomaz Freire Offrede – S3619699
tom.offrede@gmail.com
Master’s Thesis
Research Master Language and Cognition (Linguistics) Faculty of Arts
Supervisors: Dr. Simone A. Sprenger and Dr. Jacolien C. van Rij-Tange Submission date: February 16, 2020
A
BSTRACT
Speaking in a non-native language (L2) can be more challenging than speaking in one’s mother tongue (L1). Although this phenomenon is well established in the study of late bilingualism, the source of the added difficulty when accessing words in an L2 has not yet been completely understood. Whereas some evidence suggests that it is the early stages of lexical access that are more effortful in the L2—i.e., the processing of semantic information—other research indicates that the locus of the increased cognitive load is in later operations of phonological encoding. Further, some researchers suggest that the entire operation is more demanding in the L2, as compared to the L1. The present study addressed this issue by testing Dutch–English late bilinguals in a picture-naming task while their pupil dilation (PD), a measure of cognitive effort, was recorded. These participants named concrete objects in both languages while two experimental manipulations took place: (a) the target words and their translation equivalents in the unused language were either cognates or non-cognates, and (b) the pictures were inserted in blocks that were either categorically homogeneous (only images belonging to one category; e.g., animals) or heterogeneous (with mixed categories). The results of the analysis were not straightforward: although naming latencies were significantly shorter in the L2 than in the L1, pupil sizes were only slightly larger in the L2 (indicating higher cognitive effort), and only for a few moments in time. Further, PD during L1 speech did not seem to be modulated much by the experimental conditions, although, in the L2, pupils seemed to be smaller during semantically homogeneous blocks. This might indicate that there was a priming effect of semantic category during L2. Limitations of the study and its statistical analysis are discussed.
A
CKNOWLEDGMENTS
Although the name of this thesis’s author is mine, this has probably been the most collaborative project I have ever worked on. Thus, some thanks are in order! Firstly, I could not have asked for a better supervisor than dr. Simone Sprenger. No bit of this work would look the same if it wasn’t for your open ears, great theoretical and practical insights, and incredible writing feedback. Not only did you provide immense academic support for this thesis—from its very conception to its final stages—but you also made me actively look forward to our Friday meetings, which always felt like a friendly gathering way more than a school obligation. I cannot wait to continue collaborating with you in the (hopefully not too distant) future! Further, I cannot express how grateful I am that dr. Jacolien van Rij joined our team, even when her schedule seemed prohibitive. Your input has been vital throughout the entire creation and conduction of the experiment, and your statistical knowledge, patience, and kindness made our analysis happen much more smoothly than anyone could hope for with GAMMs. Next step: deconvolution!
This project could not have been possible without the amazing friends and colleagues who proactively helped me when my Dutch skills weren’t up to the task, who helped me find participants when it seemed like there were no more Dutchies to test, and who even participated in the experiment themselves. I am not citing any names because that would take another 90 pages—that is how many generous people I have crossed paths with. Still, Amélie la Roi deserves a shoutout for being so thoughtful and never hesitating to go out of her way whenever I needed help.
Working on something you are passionate about can also take a toll on you if you forget to take care of yourself (who would have imagined!?). I am inexpressibly privileged to have an amazing circle of loving, beautiful friends and family who have been present to remind me to slow down sometimes. They were ready to hear me and say a few soothing words whenever things seemed overwhelming. Again, almost no names cited, as that would take an extra 100 pages (that is how lucky I am) and I have been trying minimalism. Three people have been at the core of it, though. Sam, you couldn’t have been more patient and caring this entire time; thank you for the amazing companionship. Mãe e pai, espero que vocês entendam o quão grato eu sou por vocês sempre terem me ensinado que ter sucesso significa ser feliz, e que ser rico em dinheiro não se compara a ser rico em amor. Muito obrigado por todo o investimento emocional e prático que vocês têm feito por mim por tantos anos. Eu gostaria de conseguir retribuir tudo isso, mas seria uma tarefa impossível.
Finally, thank you to you who are reading this thesis. I hope you enjoy it, that you can take something from it, and that we can keep the conversation going. I will say it one more time: science is a collective activity.
T
ABLE OF
C
ONTENTS
Section 1: Introduction ... 5
1.1 Interactivity of the L1 and L2 systems ... 5
1.2 The Present Study ... 10
Section 2: Theoretical Background ... 13
2.1 Lexical Access ... 13
2.1.1 A Model of Lexical Access ... 13
2.1.2 The Time Course of Lexical Access ... 15
2.2 Semantic Context Effects on Lexical Access ... 20
2.2.1 Semantic Facilitation ... 20
2.2.2 Semantic Interference ... 21
2.3 Phonological Similarity Effects on Lexical Access ... 26
2.3.1 Cognate Facilitation ... 26
2.3.2 Cognate Interference ... 29
2.4 Bilingual Lexical Access ... 31
2.4.1 Differences between L1 and L2 Processing ... 32
2.4.2 Models of Bilingual Lexical Access ... 33
2.4.2 The Present Study ... 36
2.5 Pupil Dilation as a Measure of Cognitive Effort ... 37
Section 3: Method ... 42 3.1 Participants ... 42 3.2 Materials ... 44 3.2.1 Speech Production ... 44 3.2.2 Language Proficiency ... 46 3.2.3 Language Background ... 47 3.3 Apparatus ... 47 3.4 Procedure ... 47
3.5 Data Processing and Analysis ... 48
Section 4: Results ... 51
4.1 Behavioral Data: Reaction Times... 51
4.2 Physiological Data: Pupil Dilation ... 53
4.2.1 Language and Semantic Context... 54
4.2.2 Language and Phonological Similarity ... 58
Section 5: Discussion ... 62
5.1.1 Cognitive Effort in L1 and L2 ... 62
5.1.2 Semantic Context and Lexical Access ... 63
5.1.3 Phonological Similarity and Lexical Access ... 65
5.2 GAMM Models ... 66
5.3 Conclusion ... 67
References ...68
Appendix I: Distribution of English and Dutch Proficiency Scores ...86
Appendix II: Distribution of Self-Reported Use of Dutch and English ... 87
Appendix III: Stimulus Set ...89
1
I
NTRODUCTION
Most of the world’s population is composed of bilinguals (Grosjean, 1982). Indeed, Romaine (2006) indicates that there are roughly 6,700 languages spoken by humankind, and they are distributed across only 200 nation-states; this entails that the world is largely constituted by multilingual societies. With the growing awareness of the prevalence of bilingualism, in the past few decades, there has been an increasing interest in the research of bilingualism from a psycholinguistic perspective. This field investigates, basically speaking, the cognitive processes underlying how language is comprehended and produced by humans (Altmann, 2001). One issue of interest in the psycholinguistic study of bilingualism is the cognitive process that enables bilinguals to access words for speech production in the target language while suppressing the unused language (Goldrick, Ferreira, & Miozzo, 2014). More specifically, it is important to know whether or not this process is the same in a person’s first (L1) and second (L2) language, and, further, how the languages are connected: it is known, for example, that simply knowing an L2 influences how the L1 is processed cognitively (e.g., van Hell & Dijkstra, 2002). It is also well established that, in most cases, accessing words in an L2 is less efficient than it is in an L1, which makes L2 speech slower and more hesitant (Vieira, 2017). Nonetheless, it is not yet well understood what exactly makes speaking an L2 more difficult; in other words, which step, or steps, of the word retrieval process is more cognitively demanding in an L2. These are the primary questions that this study will aim to answer.
1.1 INTERACTIVITY OF THE L1 AND L2 SYSTEMS
As Grosjean (1989) puts it, a bilingual person is not “two monolinguals in one body”. This means that the languages they speak are not separate entities in their mind; they interact with one another, largely sharing the same neural and cognitive systems. A large body of behavioral as well as physiological evidence indeed reveals that, when a bilingual individual uses one language, their other language(s) are also activated, at least to some extent. Colomé (2001), for instance, asked Catalan–Spanish bilinguals to judge whether given phonemes belonged to the Catalan name of pictures of objects. These phonemes belonged either to the Catalan name of each picture, to the
Spanish name, or to none. Even though the task was only carried out in Catalan, Colomé observed that, when the target phoneme was present in the Spanish word, even if it was not in the Catalan word, participants would take longer to reject the phoneme. This is taken as evidence that, even in monolingual settings, the non-target language is activated and competes with the selected language. A similar pattern was found by Freeman, Blumenfeld, and Marian (2016) with Spanish–English bilinguals. They exploited the Spanish feature according to which words cannot start with an s+ consonant cluster (e.g., the English word study would be impossible in Spanish), which sometimes lead Spanish native speakers to add an /e/ to the beginning of such s+ syllables when pronouncing English words (e.g., estudy). This vowel addition strategy is called epenthesis. Freeman et al. tested whether Spanish–English bilinguals also apply Spanish phonotactic constraints during English listening comprehension. They aurally presented them with English words that either started or not with an s+ consonant cluster (e.g., stable). Then, these participants did a lexical decision task in which some of the non-words began with an /e/ or /es/ (e.g., elopevent and estimagle). When the participants listened to English words beginning with an s+ consonant cluster, they made decisions about English non-words that started with /e/ or /es/ significantly faster than did monolingual English speakers. The authors thus argue that, when listening only to English, their L2, these bilinguals also activated—and were primed by—their L1’s phonotactic constraints, which suggests language co-activation. In a related line of research, van Heuven, Dijkstra, and Grainger (1998) tested Dutch– English bilinguals on a series of progressive demasking and lexical decision tasks while manipulating the degree of orthographic neighborhood density in each language; that is, each of the target English and Dutch words had either few or many orthographic neighbors, either in the target or non-target language. The participants’ performance in the experiment was affected by within-language neighborhood density as well as by between-language differences in orthographic neighbors. For instance, when deciding whether a stimulus was an English word or not, reaction times were slower if this stimulus had many (as opposed to few) Dutch orthographic neighbors. Van Hell and Dijkstra (2002) further demonstrated that, for this effect to take place, the bilingual needs to have at least a high enough proficiency level in the L2. Dijkstra and van Heuven (2003) created the BIA+ model of visual word recognition to account for these (and other) findings; this model takes into account factors that influence bilingual word recognition such as task demands, semantic features, and lexical and sublexical phonological and orthographic features.
This type of parallel language activation is not restricted to phonological or orthographic processing. Kantola and van Gompel (2011), for example, primed Swedish–English speakers with different syntactic structures that expressed events containing ditransitive verbs (e.g., The bartender gave some beer to the queen vs. The bartender gave the queen some beer) and then asked them to complete other fragmented sentences of the same type. Replicating previous findings (cf. Pickering & Ferreira, 2008), the participants tended to produce the same type of syntactic structure as the one they had previously been exposed to. Crucially, though, this effect took place even when the language used for priming and that used for production were different. This would happen because, after a certain L2 proficiency level, abstract syntactic representations cease to be language-specific and become shared across languages (e.g., Bernolet, Hartsuiker, & Pickering, 2013). These results are taken as evidence that both languages are not separate, but are in constant interaction.
In addition to behavioral evidence, physiological data also validates the idea of language co-activation. Two examples are the event-related potential (ERP) studies of Spalek, Hoshino, Wu, Damian, and Thierry (2014) and of Thierry and Wu (2007). ERPs concern data with high temporal resolution involving deflections in the cortical electrical activity in response to a given experimental stimulus or manipulation (Patel & Azzam, 2005). Thierry and Wu (2007) asked Chinese–English speakers to indicate whether pairs of written or spoken English words were semantically related. Some of the Chinese translation equivalents of these word pairs are spelled with a character that is repeated across words, while some pairs do not use any shared characters. For example, the spelling of the Chinese translation of train is “火车”, and, of ham, “火腿.” Note that, even though these two words are unrelated, they share the first character of their orthography; whether or not two words are semantically related is not necessarily linked to whether they share a character. In Thierry and Wu’s study, even though the Chinese words were never overtly accessed, the bilinguals showed a reduction in the amplitude of their N400 ERP component when exposed to English word pairs whose translation had character repetition, as compared to pairs without repetition. In a similar line of investigation, Spalek et al. (2014) demonstrated that L1 phonological information is accessed even in the absence of any explicit or implicit cues for L1 activation. They tested German–English bilinguals who had been living in an English-speaking environment for at least several months with a picture-naming task. While only speaking their L2 English, these speakers said the name of a color and object presented to them; the initial phonemes of the names of the color and the object were
either the same in English, but not in German (e.g., blue bird–blauer Vogel); or in German, but not in English (e.g., red skirt–roter Rock). A facilitation effect occurred not only when the phonemes overlapped in the target L2 English, but also when they were the same in the unused L1 German. Specifically, there was a significant reduction in the mean amplitude of the electroencephalography (EEG) signal in a 300-350 ms time window after stimulus onset: even though the L1 was never produced, its phonology was still processed covertly. These data are compatible with Thierry and Wu’s (2007) findings: even in monolingual settings, bilinguals still activate different modules of their unused languages.
In accordance with the research supporting parallel language activation, a growing body of evidence suggests that there is a vast overlap between the brain regions implicated in the lexico-semantic processing of the L1 and L2, as is shown by a number of literature reviews (e.g., Abutalebi, 2008; Fabbro, 2001; Golestani et al., 2006; Gómez-Ruiz, 2010; Indefrey, 2006; Perani & Abutalebi, 2005). This overlap, evidently, is not absolute, and it is modulated by language proficiency—although not by the age at which the L2 was acquired (e.g., Chee, Tan, & Thiel, 1999; Perani et al., 2003; Poldrack et al., 1999). When processing the L2, speakers with lower L2 proficiency engage mostly the same brain areas as those dedicated to L1 lexical processing; however, these regions are activated either more strongly or more extensively for the L2. The higher the speaker’s proficiency, though, the more the L2 activation resembles the L1 activation; that is, highly proficient L2 speakers do not activate the relevant brain areas much more for the L2 than they do for the L1. In other words, the less autonomous a language is, the more brain areas are engaged in lexical processing (Gómez-Ruiz, 2010). One explanation for this pattern is the neural efficiency hypothesis (Grabner, Stern, & Neubauer, 2003): less skilled individuals (in this case, less proficient speakers) require more neural activity to accomplish a given task; more skilled individuals, on the other hand, make more efficient use of their brain resources, displaying less cortical activity. Alternatively, as Abutalebi (2008) suggests, the greater activation during processing of the less fluent L2 can be caused by the further engagement of areas linked to (non-linguistic) executive control processes. Most of the evidence gathered about lexico-semantic processing in L1 and L2, though, seems to corroborate Green’s (2003) convergence hypothesis, which states that, as L2 proficiency increases, the same neural structures as those used by native speakers start to be involved, overriding qualitative differences between L1 and L2 processing.
Despite the evidence presented thus far pointing towards (partly) integrated linguistic systems, there are also data that suggest language independence. Gerard and Scarborough (1989), for example, tested Spanish-English bilinguals in a lexical decision task in which the words were cognates, cognates, or homographic non-cognates. Cognates are translation equivalents in two languages that have a similar form, such as the English word house and its Dutch counterpart huis ‘house’. Homographic non-cognates, or false friends, are words spelled the same way in both languages but with different meanings—e.g., red in English and Spanish (in the latter, it means net). Gerard and Scarborough’s speakers seemed to access their vocabulary of each language independently: when performing in English, their response latencies and error patterns were not distinguishable from those of monolinguals English speakers (monolingual Spanish speakers were not tested). Moreover, regardless of the current language of testing, they were not influenced by the cognate status of the stimulus words, and were only sensitive to the frequency of the words in the target language. In another lexical decision experiment with participants from the same population, Scarborough, Gerard, and Cortese (1984) divided the test into two parts: the second section of the lexical decision task included words that had already been presented to the speakers in the first part, either in the same or in the other language. They observed that words which were repeated benefited from a facilitation effect in terms of accuracy and reaction times, but, crucially, only when they were presented in the same language. In other words, the authors argue that these bilinguals were not accessing the translation of the words they were presented with, which would suggest that they process their languages separately.
Moreover, Alexandrov, Sams, Lavikainen, Reinikainen, and Näätänen (1998) provided electrophysiological evidence to reveal a potential difference between L1 and L2 semantic processing. They recorded Finnish–English bilinguals’ electroencephalography (EEG) data during a listening comprehension task in which the end of each sentence was either expected (e.g., People eat bread) or unexpected (e.g., People eat scientists). These participants were either sober or under alcohol intoxication. The authors observed that non-target words elicited N100 components of smaller amplitude in the intoxicated as compared to the sober condition, and, crucially, that this effect was significantly stronger in the L2 (the N400 component was delayed by alcoholization, but this effect was not language-specific). They suggest that this difference occurred because, as had been noted by Alexandrov and Alexandrov (1993), acute alcohol ingestion primarily affects brain mechanisms underlying more recently
learned behaviors—in this case, the late-learned L2. Crucially, although strong evidence corroborates the idea that there is interaction and an overlap between a bilingual’s languages, there are still differences between the processing of the two linguistic systems.
1.2 THE PRESENT STUDY
As was argued in the previous section, when a bilingual individual speaks one language, their other language(s) also become activated to some degree. Importantly, though, this does not mean that these speakers are uncapable of controlling which language they use; in fact, most of the time, communication happens successfully without much apparent interference from one language on the other—perhaps, for example, through a mechanism of inhibitory control (Green, 1986, 1998). Furthermore, a large body of research has demonstrated that lexical access is slower and, thus, more effortful in the L2 than in the L1 (e.g., Ivanova & Costa, 2008). Crucially, although different accounts have been advanced to explain this, there is, as of yet, no consensus on the mechanisms behind this (see section 2.4.1 below). The current study aims to examine this issue: it will investigate how L1 and L2 lexical access takes place in spoken language production, targeting both semantic and phonological processes as possible bottlenecks in L2 lexical access. These two processes were chosen because, as will be detailed in section 2.1, whereas semantic operations are thought to take place in the early stages of lexical access, phonological encoding processes occur at later periods (Indefrey & Levelt, 2004).
Both of these types of processes—semantic and phonological—have indeed been demonstrated to modulate word retrieval. Concerning semantic modulation, for instance, Starreveld, De Groot, Rossmark, and Van Hell (2014) showed that bilingual speakers were considerably faster to name pictures when the stimuli were inserted in a meaningful sentential context than when they were not. In addition, when speakers are asked to name pictures that all belong to the same semantic category (in what is called a semantic blocking paradigm), their response latencies are significantly slower than when the target answers do not belong to one unique category (e.g., Damian, Vigliocco, & Levelt, 2001; Kroll & Stewart, 1994). In an ERP study, Aristei, Melinger, and Abdel Rahman (2011) demonstrated that this interference effect happens around 250 ms after the onset of the presentation of the stimulus picture, that is, around the time window when, according to Indefrey and Levelt (2004) and Indefrey (2011), lemma retrieval takes place (see section 2.1.2). In fact, the present study will use the semantic blocking
paradigm as a way to manipulate the context of speech production: as will be explained in the following sections, this type of design (potentially) creates lexical competition and can thus inform us on semantic operations of lexical selection. Specifically in the case of bilinguals, it is not yet known how this type of manipulation affects each language—whether the lexical competition effect is the same, larger, smaller, or otherwise different in each language. This would allow for the identification of any potential differences in the cognitive effort exerted in an L1 and L2 during semantic processing.
Another possible locus of L1–L2 interference is the phonological stage of word retrieval; specifically, given that the two (or more) languages of a bilingual interact, it is reasonable to assume that the degree of similarity between the phonological forms of translation equivalents plays a role in how words are accessed, making it either more or less costly. Indeed, many studies have demonstrated that cognate words (i.e., words whose phonological forms are very similar between two given languages) are processed more easily than non-cognates, both in language production and comprehension (e.g., Blumenfeld, Bobb, & Marian, 2016; Costa, Santesteban, & Caño, 2005; Hoshino & Kroll, 2008; Leacox, Wood, Sunderman, & Schatschneider, 2016). This depends, however, on many contextual factors, such as the language (L1 or L2) being spoken. Under some conditions, such as in situations of language mixing, the cognate status of a word can fail to have any influence on lexical access, or it can even produce an interference effect (Colomé & Miozzo, 2010; Sudarshan & Baum, 2019). The mechanisms behind these effects are further explored in section 2.3. Importantly, the different kinds of cognate effects in L1 and L2 can provide insight on the relatively later (Indefrey & Levelt, 2004) operations of word retrieval in bilinguals and further inform us on the source of the additional difficulty in L2 lexical access.
In order to address these issues, the present study aims to examine the time course of lexical access during speech production in both L1 and L2, in the same speakers, and the effect of semantic context and phonological similarity on said process. It will attempt to answer the following questions:
1. (How) does the time course of linguistic encoding during picture naming differ between an L1 and an L2?
2. Assuming that there is a difference between L1 and L2 lexical access, is this distinction located in earlier or in later stages of the process?
The operationalization of these questions and the method employed to answer them will be further detailed in sections 2 and 3.
The remainder of this thesis is organized as follows. Section 2 will present an overview of the theoretical background of relevant concepts discussed in this study: the time course and theoretical models of the lexical access process, effects of semantic context and phonological similarity between translation equivalents on said process, implications of bilingualism for word retrieval, and the use of pupil dilation as a measure of cognitive effort. Section 2 will also elaborate on the hypotheses and predictions concerning our research questions. In section 3, the method of data collection, processing, and analysis will be described. Section 4 will expose the results obtained, and, finally, section 5 will present a discussion of those findings in relation to the questions posed and literature reviewed.
2
T
HEORETICAL
B
ACKGROUND
This section will discuss relevant work concerning how lexical access takes place during speech production. Specifically, it will analyze the time course of the lexical access process and consider how it is influenced by the semantic context of the utterance and, in the case of bilingual speech, by the phonological similarity between the target word and its translation equivalent. Then, it will discuss how bilingual word retrieval may be different from monolingual lexical access. Finally, a brief overview of the use of pupil dilation as a measure of cognitive effort will be presented.
2.1 LEXICAL ACCESS
Many of the utterances produced by humans are relatively complex and consist of a combination of phrases. At the basis of such speech, though, is the retrieval of individual words, which, in itself, already entails intricate cognitive processes. This section will discuss one influential model of lexical access and the time course of the substages of this operation.
2.1.1 A Model of Lexical Access
It is generally assumed that lexical access involves at least three stages: (a) conceptualization of the intended message; (b) retrieval of the lemma, containing semantic and syntactic information of the word; and (c) access to the phonological form of the word (Costa, 2005). As Jescheniak and Schriefers (1998) explain, three main types of models of lexical access have been proposed: discrete serial models (e.g., Levelt et al., 1991), cascade models (e.g., Peterson & Savoy, 1998), and interactive-activation models (e.g., Dell & O’Seagdha, 1992). Serial models assume a strict hierarchy between the process’s substages: activation flows in one direction only and, although more than one lemma may be activated by a given concept, only the selected one will further activate its corresponding phonological form. Cascade models follow similar assumptions, but they also argue that even non-selected lemmas can activate their word forms. Finally, interactive models conceive the conceptual-lexical system as a network of interconnected nodes in which activation can spread in any direction; hence, not only
do lemmas activate their phonological forms, but the forms also send activation back to the lemmas (Jescheniak & Schriefers, 1998).
Figure 2.1. Outline of the model developed by Levelt, Roelofs, and Meyer for lexical access in speech production. Adapted from Levelt, Roelofs, and Meyer (1999).
One of the most influential models of lexical access was put forward by Levelt, Roelofs, and Meyer (1999), based on experimental data and the computational implementation WEAVER++. The model argues that accessing a word for speech production begins with the formulation of a concept to be expressed. Then, by retrieving information from the mental lexicon, lexical selection produces a lemma, containing its relevant syntactic features. Still drawing from the mental lexicon, processes of morphophonological encoding and syllabification give way to a representation of the word in terms of all its phonological segments, organized in syllables. The next substage in the WEAVER++ is phonetic encoding: it uses items from the syllabary to transform the phonological word into a gestural score. The syllabary contains overlearned syllabic scores that override the need for re-computation of the gestures involved in producing phonemes that co-occur frequently. Finally, an
articulatory system executes the phonological word’s gestural score. This latter system is comprised of the organs involved in the physical production of sounds (lungs, larynx, vocal tract) as well as the neural components that control the transformation of the gestural score into speech. The model is summarized in Figure 2.1.
2.1.2 The Time Course of Lexical Access
Indefrey (2011) provides a meta-analysis, based on Indefrey and Levelt’s (2004) overview, of neuroimaging and behavioral studies that provide insights about the time course (and neural correlate distribution, although this is not the focus of this work) of the word retrieval process. He bases his estimations on a large body of evidence as well as on Levelt et al.’s (1999) model of lexical access. In their meta-analyses, Indefrey and Levelt (2004) and Indefrey (2011) compare empirical evidence from several studies that investigate lexical access using different experimental paradigms. The authors base their estimations on the assumption that the average response latency in picture naming tasks is 600 ms. This number was derived from the average RTs reported in three picture naming studies in which stimulus pictures were presented repeatedly to the participants: Damian et al. (2001), with a mean RT of 567 ms; Jescheniak and Levelt (1994), with a mean RT of 680 ms; and Levelt, Praamstra, Meyer, Helenius, and Salmelin (1998), with a mean RT of 591 ms ((567 + 680 + 591) / 3 = 612.7 ms).
Studies have tackled the issue of the time needed for access of conceptual information using different types of tasks—e.g., by comparing the naming latency of objects with distinct semantic properties or by requiring a response from participants (or absence thereof) based on a decision related to the target’s semantic features, such as its state of animacy (Indefrey, 2011). Abdel Rahman and Sommer (2008), for instance, taught participants new words describing novel objects under two conditions: they gave either in-depth, complex definitions of the objects, or simpler, minimal ones. These participants did not take any longer to name the objects of each condition, but the availability of more complex knowledge was reflected on their ERPs as early as 120 ms after object presentation, suggesting an early availability of conceptual information. Furthermore, four studies have tested participants on a go/no-go task based on whether the presented image was an animal or not (Thorpe, Fize, & Marlot, 1996), if it was an animal or an object (Schmitt, Münte, & Kutas, 2000; Rodriguez-Fornells, Schmitt, Kutas, & Münte, 2002), or if it was a living or a non-living object (Zhang & Damian, 2009). They noted that the ERPs elicited by each response condition (go or no go) started to diverge between 150 and 264 ms, depending on the study, which indicates
that conceptual knowledge has already been accessed by then. Finally, another study tested healthy and anomic individuals on a simple picture naming task (Laganaro, Morand, & Schnider, 2009). These authors found that, during speech planning, the ERPs of patients with semantic impairment—but not of those with phonological impairment—started to diverge from the ERPs of healthy controls as early as at 90 ms post-stimulus. Indefrey (2011) states that the median latency of these studies (and of others that he reviews) is 200 ms, and that this must thus be a fair estimate of the time necessary for conceptual preparation.
Once enough conceptual information is available, lemma retrieval can take place. In a simple picture naming task, Costa, Strijkers, Martin, and Thierry (2009) analyzed speakers’ ERPs in relation with how many items from the same semantic category had been previously presented at each point in time. In other words, they manipulated the amount of lexical competition the participants underwent (this notion will be further discussed in section 2.2). Between 200 and 380 ms post-stimulus, the amplitude of their ERP waveforms increased with the number of items from the same category that had been presented, that is, with higher levels of lexical competition. Indeed, this type of competition has been demonstrated to take place during semantic processing of the word (e.g., Belke, Meyer, & Damian, 2005; and see section 2.2). Aristei et al. (2011) conducted a similar experiment with a combination of the semantic blocking and the picture-word interference paradigms, and also found that their semantic manipulations started to have an effect on ERPs at approximately 200 ms. This evidence corroborates the estimation that lemma retrieval does begin around 200 ms, when conceptual information is available. To investigate the duration of the process of lemma retrieval, which comprises both semantic and syntactic information, it is relatively common to study the processing of syntactic information (specifically, syntactic gender). Because this discussion goes beyond the scope of the present review, as we are not investigating syntactic effects on lexical access, we refer the reader to Indefrey’s (2011) overview. What is essential to know, however, is that syntactic gender information seems to be available around 75 ms after conceptual information has been accessed. This estimate is an upper boundary of the lemma retrieval stage, since, for syntactic gender information to be accessible, the lemma already needs to have been retrieved (Indefrey, 2011).
The next step in word access is the encoding of the word’s form. In Levelt et al.’s (1999) model, this consists of phonological code retrieval, syllabification, and phonetic encoding. Morgan, van Elswijk, and Meyer (2008) conducted an experiment that gives
a (partial) indication of when phonological access takes place. They displayed pictures that their participants should name. On some trials, the image was replaced by a written word either at 150 or at 350 ms after picture onset; when this happened, the subjects were instructed to name the word instead. The authors observed that, when the word was phonologically related to the picture’s name (e.g., the picture of a bottle and the word bottom), it was named faster than when it was unrelated (bottom–keen); crucially, however, this effect occurred when the image had been seen for 350 ms, but not if it had only been seen for 150 ms. This corroborates the aforementioned time course estimations and indicates that 150 ms is not enough time for a word’s phonological form to be activated. In addition, a few studies provide data that informs the first of those stages by testing participants on a go/no-go task contingent on the first phoneme of the target picture’s name (Rodriguez-Fornells et al., 2002; Guo, Peng, Lu, & Liu, 2005, as cited in Indefrey, 2011; Zhang & Damian, 2009). In other words, participants must press a button if the target word’s first phoneme matches a predefined phoneme (e.g., /t/–table), or refrain from pressing it if it does not (/m/– table). They report that the ERP waveforms of the two conditions started to differ between 300 and 456 ms, depending on the study—this is the moment by which the first phoneme of the word is available.
Once the phonological code has been retrieved, it is mapped onto syllables— although, as Indefrey (2011) notes, there is no reason to believe that the syllabification process only starts after the entire code has been accessed. A few studies provide fair estimates of the duration of syllabification. Van Turennout, Hagoort, and Brown (1997), for instance, conducted two ERP experiments in which they showed pictures to their participants and instructed them to either press a button or not under two conditions: based on either the initial or the final phoneme of the corresponding word. The authors identified a lateralized readiness potential (LRP) component which started to build up on every trial, but, on no-go trials, decreased after a few milliseconds. The onset of the LRP reduction was at approximately 40 ms when the initial phoneme was being monitored, and at around 120 ms when the word’s final phoneme was attended to. Although this time difference cannot be taken strictly, as the measurements came from different pools of participants, the authors argue that these values indicate that words of an average of 1.5 syllable and 4.5 phonemes take about 80 ms to be syllabified. Furthermore, Wheeldon and Levelt (1995) asked Dutch participants to covertly translate aurally presented English words and to monitor the first phoneme of either the first or second syllable (all words were disyllabic), pressing a button if a given
phoneme was present in it. They found a 123 ms difference in button-press latency between first and second syllable monitoring. Interestingly, they also observed that this latency was not modulated by the number of phonemes, that is, syllables were monitored as a cluster, and not in terms of individual phonemes. Finally, Schiller (2006) showed his participants pictures and tested them on a go/no-go task based on whether the first or the second syllable of each picture’s name was stressed. Not only were their RTs on go trials faster in stress-initial words than in stress-final words (885 and 971 ms, respectively; a significant difference of 86 ms), but their ERPs also showed a distinction. When the response decision was contingent on initial stress, the peak of their ERP waveforms, which Schiller interprets as being related to inhibitory control, was around 475 ms after picture presentation. When the final syllable was being monitored, on the other hand, this peak latency was of about 533 ms (58 ms difference). Indefrey’s (2011) conclusion from these studies is that syllabification takes around 20 ms per phoneme, and about 55 ms per syllable. Importantly, the evidence presented also points to the view that syllabification happens in a left-to-right manner, that is, the first syllable of the word is the first one to be processed, and so forth.
Considering that syllabification starts at approximately 355 ms after picture onset and that it lasts for around 100 ms, the transformation of this phonological word into an abstract motor plan—the gestural score—would begin taking place at 455 ms. Again, Indefrey (2011) points out that this is an upper boundary, as phonetic encoding can start as soon as the first syllable has been phonologically encoded. This stage lasts until the articulatory execution of the word, that is, until the estimated 600 ms for the overt naming onset. Figure 2.2 summarizes Indefrey’s (2011) estimations for the time course of lexical access.
Figure 2.2. Estimated onset times and durations, in milliseconds, of operations involved in spoken word encoding, according to Indefrey (2011). The time required for access of a percept
The estimates of durations of the operations implicated in lexical retrieval proposed by Indefrey (2011) are informative about the sequential order of the lexical access operations and their proportional duration in relation to each other. Nevertheless, they should not be taken too strictly. Firstly, these values are based on the assumption that word planning happens in an interval of approximately 600 ms. This assumption only holds true under some circumstances, namely, in picture naming tasks in which each target stimulus is presented repeatedly. Clearly, this is not the average real-world scenario in which natural word production takes place. Additionally, the 600-ms time window comprises not only word planning, but also cognitive processes related to visual perception and object recognition (the lead-in process, as Indefrey and Levelt, 2004, call it, that serves as conceptual input). Secondly, specific lexical retrieval events are arguably influenced by a number of factors—related to the environmental context in which speech takes place, to individual characteristics of the speaker, and to word features—which may have differential impacts on each operation. This is demonstrated by the fact that no two of the previously discussed studies obtained the exact same observations. Furthermore, the 600-ms estimate was obtained from studies with monolingual speakers; speech is known to take longer in bilinguals than in monolinguals (see section 2.4.1), so this estimation does not hold true for many (or most) real-world settings. Finally, as Strijkers and Costa (2011) argue, the studies used to determine the time course estimations mostly employ indirect measures of latency and do not tap purely into the substages of word retrieval. Indeed, many of them base their evidence on go/no-go tasks, which requires further metalinguistic decision-making processes, instead of basic low-level linguistic computations that happen in realistic situations. These decision-making processes are modulated not only by the linguistic features manipulated experimentally, but also by task-related factors. For instance, many of the experiments discussed asked participants to identify whether a word belonged to one or another semantic category or if it started or not with a given letter. However, perceiving objects as belonging to classes such as animal or tool is a natural way in which humans categorize the world (e.g., Caramazza & Mahon, 2003; Warrington & McCarthy, 1987) and thus happens easily; classifying them as starting with such or such sound is not. Hence, Indefrey’s (2011) time course estimates serve the purpose of a working frame and guideline.
2.2 SEMANTIC CONTEXT EFFECTS ON LEXICAL ACCESS
As was mentioned in section 1, the lexicalization process is modulated by environmental, linguistic, and intrapersonal factors. One such factor is the semantic context of the utterance, herein considered as any meaning-related constraints on the word production event, be they related to real-life situations or experimental settings (Abdel Rahman & Melinger, 2009). This section will explore two types of effects produced by different contexts: a facilitation and an interference effect.
2.2.1 Semantic Facilitation
Some types of semantic context induce a facilitation effect on word retrieval, as is the case, for example, when the related words have an associative connection (e.g., duck and lake; vampire and blood). Costa, Alario, and Caramazza (2005), for instance, observed this effect: using the picture-word interference (PWI) paradigm, they asked participants to name pictures that had a distractor word superimposed on them, which could be either unrelated to the concept of the image (e.g., PARROT–car) or linked to it in a part–whole relationship (e.g., BUMPER–car). The participants’ naming latencies were significantly shorter when there was a conceptual association (around 696 ms) than when there was not (719 ms). De Zubicaray, Hansen, and McMahon (2013) replicated this finding in a similar experiment, but using thematically related words (e.g., POPCORN–movie) as opposed to part–whole related concepts. In their study, the mean RT in the thematic association condition was 787 ms; in the unrelated condition, 803 ms.
One of the explanations proposed for this effect is that related words prime the lexical node of the target item more than do unrelated distractors; this priming would then speed up the process of lexicalization (Costa et al., 2005). This notion ensues from the idea of lexical access through spreading activation between different representational levels (see, e.g., de Zubicaray et al., 2013; Dell & Sullivan, 2004). This framework predicts that, when one node is activated, it sends a proportion of its activation to all of the nodes that are linked to it. The proportion of activation that is sent out to each connected node varies, depending on the strength of the connection (Dell, 1986; Dell & O’Seaghdha, 1991; Roelofs, 1992). As a concrete example, consider that an individual’s goal is to say the word whale aloud. The lexicalization process will begin with the enhanced activation of the concept whale—or, more precisely, of at least enough conceptual information that allows for the continuation of lexicalization, such as the features “animacy” and “animal” (Indefrey, 2011). This activation then spreads
to other similar concepts, such as fish and sea, and to its corresponding lemma, containing semantic and syntactic information. Outside of the semantic scope, activation then spreads further to the relevant phonological form (again, in addition to other potentially similar words, such as way and wait; Dell & O’Seagdha, 1992). At each of these stages, the selected node is the one with the highest level of activation (Roelofs, 1992). Thus, when it comes to the aforementioned semantic facilitation effect, the spreading-activation framework suggests that the activation of, say, the concept blood will spread to related concepts, such as red and vampire, priming them and facilitating access to such words.
2.2.2 Semantic Interference
In other types of contexts, in contrast, the relationship between the target word and the other concept(s) activated can create an interference effect which makes lexical selection more effortful. This occurs, for example, when the words’ relationship is not one of association, but of categorical identity: e.g., kangaroo, octopus, and butterfly, all three concepts belonging to the category animals. Indeed, the two previously mentioned studies (Costa, Alario, et al., 2005; de Zubicaray et al., 2013) also used experimental conditions with categorically related word distractors (e.g., GUITAR– violin) and found an interference effect on those trials: the mean naming latencies in this condition were 748 ms in Costa, Alario, et al. (as opposed to 719 ms in the unrelated condition) and 833 ms in de Zubicaray et al. (as opposed to 803 ms in the unrelated condition). These values are illustrated in Figure 2.3.
Several studies have replicated and extended these findings—most of them, but not all, using the PWI and/or the semantic blocking paradigms (see, e.g., Belke, 2008; Belke, Brysbaert, Meyer, & Ghyselinck, 2005; Friederici, Meyer, Maess, Levelt, & Damian, 2002; Python, Fargier, & Laganaro, 2018; Shao, Roelofs, Martin, & Meyer, 2015). For example, Vigliocco, Lauer, Damian, and Levelt (2002) replicated the semantic interference effect in a task where Dutch participants were presented with English words and needed to respond with their translations. It has also been demonstrated that the effect manifests in non-fluent aphasic patients in terms of either more error rates in comparison with fluent aphasics and healthy controls (Schnur, Schwartz, Brecher, & Hodgson, 2006) or an exaggerated slowing down of RTs (Biegler, Crowther, & Martin, 2008). Vigliocco, Vinson, Damian, and Levelt (2002) also provided valuable insights regarding the nature of the interference effect. In their picture naming task, they did not simply group the stimuli into semantically homo- or
heterogeneous contexts; using measures of semantic relatedness between concepts that were obtained empirically, they created blocks of pictures that were taxonomically similar (e.g., body parts), closely related (e.g., body parts and clothing items), distantly related (e.g., body parts and vehicles), or unrelated. The semantic interference effect manifested in a gradient manner: naming latencies in the categorically identical blocks were longer than in the closely related ones, which, in turn, were slower than in the distantly related condition. This indicates that the effect is not either present or absent, depending on strict category boundaries; rather, it is modulated by the number of semantic features shared by the concepts in question (Vigliocco, Vinson, et al., 2002). Another interesting observation is that the interference effect is persistent over time, even when the categorically related targets are interspersed with unrelated filler words, as demonstrated by Damian and Als (2005). Indeed, Howard, Nickels, Coltheart, and Cole-Virtue (2006) found that the only factor that predicted naming latency was the word’s ordinal position (a finding also obtained, as previously noted, by Costa et al., 2009); even a lag of as many as eight fillers between the target words did not reduce the interference effect. In addition, the effect was cumulative: it increased RTs at about 30 ms with each new presentation of a target item.
Figure 2.3. Mean reaction times in Costa et al.’s (2005) and de Zubicaray et al.’s (2013) studies in semantic contexts containing associatively related, taxonomically related, and unrelated words.
The Lexical Competition Account
The most commonly accepted explanation for semantic interference is that lexical selection happens through competition. It is also derived from the notion of spreading activation and it maintains that, when more than one concept or lexical unit are activated, they will compete for selection, thus slowing down the lexicalization process (Damian et al., 2001; see also Roelofs, 2018). Damian et al. (2001) argue that there are two possible confounds that could dispute the lexical competition hypothesis: the semantic interference effect could be due (a) to visual confusability, as pictures of same-category objects often look similar and it might be more cognitively demanding to differentiate them, or (b) to non-linguistic conceptual conflict, which would mean that the locus of the effect is not related to linguistic processing. The authors addressed these issues by carrying out two experiments. In one, they conducted a simple picture naming task while minimizing the visual similarity between the pictures of each category; in the other, they asked participants to name German words written on the screen either with or without their gender-marked determiner, in a design that is not expected to produce any interference effects. Both experiments were conducted in accordance with the semantic blocking paradigm, that is, some blocks of trials only contained concepts that were categorically homogeneous, whereas others had unrelated concepts. Damian et al. (2001) replicated the semantic interference effect in the first experiment, demonstrating that visual confusability does not explain the phenomenon. The second experiment also produced the interference effect, but only when speakers needed to retrieve the grammatical gender of the words. The authors argue that this finding refutes the confound related to non-linguistic conceptual conflict, as the interference effect was obtained when the participants accessed the words’ lemmas (to retrieve gender information), but not when they only accessed their phonological representation. As a conclusion, they sustain that lexical competition is indeed the mechanism underlying the interference effect observed. The lexical competition account is also supported by Crowther and Martin’s (2014) finding that, across participants, better capacity to inhibit distractors, as measured through a verbal Stroop task (Stroop, 1935), was negatively correlated with the strength of interference in homogeneous blocks.
Belke, Meyer, and Damian (2005) provide further evidence in support of such notions. In one of their experiments, they presented participants with pictures in a classical semantic blocking paradigm; however, they presented the same images repeatedly in eight cycles. Semantic interference occurred as expected, but, crucially,
only after the first presentation cycle. This is in line with the spreading-activation account described above: the target concepts must be activated, and their category established, before there can be competition for selection. In a second experiment in the Belke, Meyer, et al. (2005) study, the blocks of pictures were not presented sequentially, but simultaneously; that is, all the objects appeared at once on the screen. Additionally, the speakers’ eye movements were tracked. The researchers observed that, in semantically homogeneous blocks, not only were response latencies slower, but the objects were also fixated on for longer, as compared to heterogeneous blocks. However, semantic context did not modulate the participants’ eye-speech lag, that is, the lag between the offset of gaze fixation and speech onset. This observation is important because it has been demonstrated that, although gaze duration is contingent on processes of object recognition and name retrieval, gaze-to-speech lag only depends on post-lexical and articulatory operations (see Meyer, 2004).
There have also been alternative accounts proposed concerning the locus—as well as the origin—of the interference effect; for instance, Belke (2013) and Roelofs (2018) both argue that the origin of the effect is conceptual and that its locus is lexical. These fine-grained details transcend the scope of the present discussion, and the interested reader is referred to such works (see also Navarrete, Mahon, & Caramazza, 2010; Oppenheim, Dell, & Schwartz, 2010). Nonetheless, what is largely agreed on is that the interference effect reflects early stages of the word retrieval process.
The Swinging Lexical Network Account
Abdel Rahman and Melinger (2009) point out that the lexical competition explanation as described above fails to account for the different facilitation and interference effects presented thus far. In fact, in an earlier study, they observed both a facilitation and an interference effect in the naming of pictures that were associatively related (Abdel Rahman & Melinger, 2007). They have thus put forward the swinging lexical network account in order to explain such contradictory findings. This account does not go against the lexical competition view; rather, it proposes an extension to it. The swinging lexical network account sustains two assumptions. Firstly, for a semantic interference effect to be manifested, there must be activation of a cohort of lexical competitors of a large enough size—that is, more than one competitor must be activated, although the authors admittedly do not know how many items in the cohort are enough. This assumption is derived from notions such as the Luce ratio (Luce, 1959), according to which the probability of a lemma being selected is contingent on
the state of its activation in relation to the sum of the activation of all its competitors (Abdel Rahman & Melinger, 2009). That is to say, the more items are activated, the longer it takes to select a target. The second assumption of the account states that semantic context has a differential impact on two distinct stages of the word retrieval process. These effects are (a) a facilitation at the conceptual level, through semantic priming, since concepts receive (more or less) activation, but do not need to be actively selected, and (b) an interference effect at the lexical level, as one target item must be chosen from among a set of candidates. This way, the resulting effect of semantic context on any given speech event is the interaction of the facilitative and the interfering effects. In fact, Abdel Rahman and Melinger (2009) specifically predict that interference will only happen under two circumstances, namely, when a cohort of interrelated lexical items are activated and/or when conceptual facilitation is largely reduced. The authors do not specify the mechanisms that control the interaction between the two effects.
The swinging lexical network account assumes that the semantic categories that form the cohort can either be well established and known categories (e.g., types of fruit) or, in accordance with Barsalou (1983), ad hoc categories. The latter are constructed in order for the individual to achieve a particular goal. For instance, an ad hoc category called “things that can fall on your head” (a category Barsalou, 1983, used in one of his experiments) might include elements such as an apple from a tree, an avalanche, and a piano—items that do not normally co-occur. Indeed, Abdel Rahman and Melinger (2011) provided evidence indicating that semantic interference can occur when the categories in question are constructed ad hoc. In a set of experiments, their participants repeatedly named pictures that were organized in three types of blocks: categorically homogeneous, heterogeneous, or containing seemingly unrelated items but that could be grouped into a category (e.g., rice, a camera, and an altar for the category “wedding”). As expected, the two latter categories did not produce any interference effects. However, when the authors added a title to the beginning of the task describing the block’s theme, naming latencies became slower even for the ad hoc categories. These titles grouped the concepts into meaningful contexts and created a more strongly activated set of competing units.
In sum, it is generally accepted that semantic contexts containing taxonomically related words cause an interference effect on word retrieval, as they create lexical competition due to concomitantly activated lemmas.
2.3 PHONOLOGICAL SIMILARITY EFFECTS ON LEXICAL ACCESS
In bilingual settings, another factor which might affect language processing and, more specifically, speech production, is the similarity in form between the target word and its translation equivalent in the unused language. One simple way to measure the phonological distance between words in two languages is to determine their cognate status. As explained in section 1, cognates are translation equivalents whose phonological form is similar in both languages (e.g., English bear and Dutch beer). The production of cognates has been demonstrated to be both easier and more difficult than that of non-cognates. These findings are discussed in the present section.
2.3.1 Cognate Facilitation
The most commonly found cognate status effect on language production is one of facilitation: cognate words are reported to be produced both faster (e.g., Christoffels, Firk, & Schiller, 2007; Costa, Caramazza, & Sebastian-Galles, 2000; Ivanova & Costa, 2008; Starreveld et al., 2014; Strijkers, Costa, & Thierry, 2010) and more accurately (e.g., Leacox et al., 2016; Rosselli, Ardila, Jurado, & Salvatierra, 2014) than non-cognates. In addition, Hoshino and Kroll (2008) and Gollan, Forster, and Frost (1997) have demonstrated that this facilitation effect is persistent even across languages that use different scripts (English and Japanese in the former study and English and Hebrew in the latter). Blumenfeld et al. (2016) asked their participants, who were either monolingual speakers of English or Spanish–English bilinguals, to produce as many English words as they could, within one minute, that either started with a given letter (e.g., the letter S) or that belonged to a certain category (e.g., countries). The bilingual participants produced more cognates than their monolingual counterparts in both tasks, and, interestingly, they also produced more cognates in the letter than in the category task. This indicates that it is easier to spontaneously retrieve cognates than non-cognates.
There have been different suggestions regarding the origin of the cognate effect (Costa, Santesteban, et al., 2005). It has been advanced, for instance, that cognates have a larger conceptual overlap than do non-cognate translation equivalents (van Hell & de Groot, 1998) and that cognate words have a similar lexical-morphological structures in both languages, which would enable bilinguals to re-use the same representation (Kirsner, Lalor, & Hird, 1993). However, Costa, Santesteban, et al. (2005) indicate a number of theoretical—and empirical—flaws with these views. For example, if the origin of the effect was semantic, then we would also need to assume
that, say, the words iron and ironing share more conceptual features than the words broom and sweeping do, simply because of their closer phonological form. Further, if the origin was morphological, it would be difficult to explain how one morpheme gives way to two distinct phonological realizations—and it would be even hard to identify the component morphemes of cognate words in the first place. This discussion fundamentally amounts to one idea: the distinction between cognates and non-cognates is formal in nature—i.e., what differentiates them is their phonological similarity—so semantic and morphological dimensions cannot explain the effects of a word’s cognate status. Indeed, the most likely explanation is that the cognate effect originates at a phonological-sublexical level (Costa, Santesteban, et al., 2005).
The authors argue that, during lexical retrieval, phonological information of cognate words is activated from two sources: the target word and its translation equivalent. One mechanism through which this could happen would require the assumption that the lexical selection process is bi-directional. In this scenario, the activation of a lexical representation would spread to its corresponding segments, and said segments would then spread the activation back to all words connected to them in either language. Thus, because the Dutch word salade is formally closer to its English equivalent salad than paddestoel is to its counterpart mushroom, there would then be stronger activation flowing between the former pair than between the latter. A second possible mechanism, which does not preclude the first suggestion, is related to the frequency effect (Oldfield & Wingfield, 1965). This would mean that the time it takes for a speaker to name a word depends on when was the last time said word was used: more recently processed words are easier to activate. In the case of bilinguals, as Costa, Santesteban, et al. (2005) explain, this recency effect takes both languages into account, so, the more recently the Dutch word skelet was used, the more activation will be sent to its English translation skeleton when the latter is the target word for production. In other words, according to this view, cognates are treated by the bilingual linguistic system as higher-frequency words. Indeed, Sherkina-Lieber (2004) demonstrated that Russian–English bilinguals subjectively perceived Russian–English cognates in English as more frequent than English monolinguals did, even though there were no differences between the bi- and monolinguals’ frequency perceptions when it came to non-cognate words.
In line with these explanations, a bilingual’s proficiency in each of their languages has also been demonstrated to modulate the cognate effect. In the aforementioned Blumenfeld et al. (2016) study, individuals who were more proficient
in their less-dominant Spanish produced more cognates in both languages and, interestingly, the cognates they used in their L1 tended to have lower frequency. According to the authors, this suggests that greater L2 knowledge may increase the functional frequency of L1 cognate words, as they explain that Spanish–English cognates generally have lower frequency, based on monolingual corpora in each language. Moreover, in a series of five experiments, Poarch and van Hell (2012) asked native speakers of German who spoke English at various levels of proficiency, and some of which who also spoke a third language, to name pictures in German and in English. Cognate words produced a facilitation effect (faster naming latencies) in both languages only when the participant’s English proficiency was high enough; even then, the effect was always smaller in their dominant German. In addition, children who were learning English as an L2 did not show any signs of cognate facilitation. Rosselli et al. (2014) obtained a similar finding: testing Spanish–English bilinguals who had different degrees of language dominance on a picture naming task, they observed that cognates were overall named more accurately than non-cognates. Crucially, the more balanced bilinguals benefited from the cognate effect similarly in both languages; the non-balanced group, on the other hand, showed much stronger cognate facilitation in their less-fluent L2.
According to Poarch and van Hell (2012) and to Rosselli et al. (2014), each language benefits differently from cognate statuses because of the manner in which activation spreads. Specifically, they argue that activation of phonological forms is always stronger in the dominant than in the non-dominant language. Hence, when lemma selection takes place in the non-dominant language, thus spreading activation on to its translation equivalent in the dominant language, the dominant phonological form will spread onto the form of the non-dominant language. When lemma selection occurs in the dominant language, a similar process happens; however, because the dominant phonological form is already activated at its strongest, the activation of the non-dominant representation does not boost the other one as much. Of course, in balanced bilinguals, the activation threshold is (more) similar in both languages, so the spread of activation is more likely to flow equivalently in both directions.
Other research provides further insights with respect to the interaction between relative language proficiency and cognate facilitation. Christoffels et al. (2007) tested unbalanced German–Dutch bilinguals on a picture-naming task under two conditions: the languages were either blocked (only one language being used) or mixed. These researchers also collected ERP data, which will be discussed in the following section. In
the blocked condition, the same effect was found as described above—namely, the cognate facilitation effect was larger in the participants’ non-dominant L2 than in their L1. In contrast, when the languages were mixed and constantly had to be switched, naming cognates was faster in the L1 than in the L2. The authors explain this pattern in accordance with Kroll, Bobb, and Wodniecka’s (2006) suggestion. Although in normal (non-mixed) contexts, the dominant language is less susceptible to influence from the L2, this changes in a mixed-language scenario where both languages are highly active. In the latter case, the speaker lowers their overall L1 activation (or, alternatively, increases their L1 activation threshold), which would lead the two languages to be equivalently activated; or they activate their L2 even more strongly. Hence, the spread of activation from the less to the more dominant language may be reversed. This is related to the asymmetrical switching costs phenomenon (Meuter & Allport, 1999), whereby switching from an L2 to an L1 is more costly than from an L1 to an L2.
2.3.2 Cognate Interference
As demonstrated above, it has been consistently observed that cognate words, as compared to non-cognates, are processed more easily by bilinguals, which reflects in the time spent on lexical access as well as in the accuracy of their output. Nevertheless, not only are there contexts in which cognates produce a diminished or no facilitation effect—e.g., when L2 proficiency is not sufficiently high (Poarch & van Hell, 2012)—but, in some cases, cognates can also effectively hamper word retrieval. This was the case, for instance, in Colomé and Miozzo’s (2010) study. They showed Spanish–Catalan bilinguals two overlapping pictures and asked them to name one of them while ignoring the other. Interestingly, when the name of the distractor image was a Spanish–Catalan cognate, naming latencies were slower than when it was a non-cognate. Furthermore, Sudarshan and Baum (2019) studied the relationship between the cognate effect and individual differences in inhibitory control. Their highly proficient French–English bilinguals named pictures while listening to cognate and non-cognate distractor words that were presented at different moments in relation to the stimulus image onset—at stimulus-onset asynchrony (SOA) times of -300, -150, 0, and 150 ms. The participants with higher inhibitory control capacity (as measured through a non-verbal version of the Simon task; Simon & Rudell, 1967) were not affected by the cognate status of the distractors. In contrast, those with lower inhibition capacity displayed two patterns of influence: at the late SOAs, they named cognates more quickly than they did non-cognates; at earlier SOAs, though, they suffered a cognate interference effect. To
