Master Thesis
Parallel language activation during bilingual word processing: Evidence from
word production both within and out-of-context.
Student
name : Snoeks, D.
address : Prins Hendriklaan 91
zip code and residence : 2051 JB Overveen telephone number : 0643423115 student ID Card number : 0577456
e-mail address : daansnoeks@gmail.com
date : April, 2015
Supervisor(s UvA)
specialization : De Groot, A.M.B. & Starreveld, P.A.
Table of contents
Abstract 3
Introduction 4
Bilingual word recognition 4
Bilingual word recognition in context 6
Bilingual word production 8
Bilingual word production in context: Starreveld et al. (2014) 11 The present study: bilingual word production in various types of context 13
Method 14
Participants 14
Research design 15
Research strategy 16
Materials and apparatus 17
Procedure 20
Results 22
General discussion 25
Does the effect of cognate status depend on the type of context used? 25 Differences between Starreveld et al. (2014) and the present study 25
Ignoring the context 27
Improving the methods used 28
Conclusion 30
Parallel language activation during bilingual word processing: Evidence from word production both within and out-of-context.
Abstract
By means of a picture-naming experiment the present study examined whether the extent to which bilinguals co-activate the non-target language during word production is contextually dependent and, if it is, it will also identify the aspect(s) of context that are necessary for the existence of such a dependency. Different groups of Dutch-English bilinguals performed a picture-naming task in English (i.e. their second language). The pictures' names were either Dutch-English cognate words or non-cognate words and the cognate effect served as a marker of activation of the non-target language. The experiment directly compared the (size of the) cognate effect that was obtained with pictures presented in various types of context. Even though a cognate effect was generally found, the size of this effect was not contextually dependent. This result does not correspond with earlier research, which suggests that the extent to which bilinguals co-activate the non-target language is contextually dependent. Disagreement was most likely caused by the fact that participants in the present experiment ignored the context.
Parallel language activation during bilingual word processing: Evidence from word production both within and out-of-context.
Introduction
The current study examined whether the extent to which bilinguals co-activate the non-target language during bilingual word production is contextually dependent. This would, however, be an oddly specific starting point for a discussion. Therefore, the present discussion will begin by examining whether bilingual language processing is either selective or language-nonselective (i.e. co-activation of information in both of a bilingual's linguistic subsystems is known as language-nonselective lexical access, whereas exclusive activation of information in the contextually appropriate system is known as language-selective lexical access.). In doing this a distinction between word recognition and word production will be made. Initially, the question of whether a written word causes activation in both of a bilingual's linguistic subsystems
simultaneously or whether activation is restricted to the contextually appropriate subsystem will be addressed. This will be done by describing several studies that examine bilingual word recognition both within and out-of-context. Afterwards, the focus will be on word production and on the question of whether or not during the process of generating a word, simultaneous activation occurs in the contextually inappropriate subsystem. This will, once again, be done by describing several studies that address bilingual word production both within and out-of-context. Finally, an outline of the present study will be given.
Bilingual word recognition
Certain words, such as interlexical homographs or cognates, share either meaning, orthography or phonology between languages. These types of words can be used in order to assess whether word recognition is either a language-selective or a language-nonselective process. Experiments in which
these types of words are used commonly assume that, if word recognition is language-nonselective, then the processing of words that share a between-language relation should differ from words that do not share such a relation. For instance Dijkstra, Van Jaarsveld and Ten Brinke (1998) compared the speed at which recognition of interlexical homographs and matched non-homographic controls occurs. Interlexical homographs are words with the same orthographic form, but with different meanings in a bilingual’s two languages (e.g. the English word coin means corner in French). Dijkstra et al. (1998) described two separate experiments. In both of these experiments Dutch-English bilinguals were asked to categorize letter strings as either words or nonwords in Dutch-English. However, the stimuli that were used differed between these experiments. In the first experiment all nonwords letter strings were nonwords in both Dutch and in English, whereas in the second
experiment the nonwords letter strings could be actual Dutch words. Dijkstra et al. (1998) found that, in the first experiment, interlexical homographs and matched non-homographic controls were responded to equally fast. In the second experiment, however, response times were longer for interlexical homographs than for matched non-homographic controls. Dijkstra et al. (1998) argued that the more elements of the non-target language (i.e. Dutch) that are included in the stimulus set, the larger the level of activation of the non-target language system will be. Thus, because Dutch words were included as nonwords in the second experiment, the non-target language system has been activated more. This non-target language activation must be suppressed, which causes a delay in word recognition in the target language. The fact that in the first experiment homographs and controls were processed equally fast suggests that the non-target language was not sufficiently activated to interfere with word recognition.
Additionally Dijkstra, Grainger and Van Heuven (1999) described experiments in which they examined whether the speed at which word recognition occurs differed between cognate words and control words. Cognate words are translation pairs that largely or completely share either phonology or orthography between a bilingual’s two languages (e.g. the Dutch-English translation pair muis-mouse), whereas non-cognate words are translation pairs that do not share any amount of
phonology or orthography (e.g. the Dutch-English translation pair bijl-axe). Dijkstra et al. (1999) report an experiment in which Dutch-English bilinguals performed a progressive demasking task. This is a task in which the presentation of a target word is alternated with that of a mask. During this process of alternation, the target presentation time increases while that of the mask decreases. The participants were instructed to push a button as soon as the target word is identified. The obtained results indicated that response times for cognate words were shorter than for matched controls. This suggests that whenever words exhibit orthographic and semantic overlap between languages, this would lead to facilitation of the word recognition process. The authors concluded that this pattern of results would not correspond with a view in which word recognition was language-selective.
The studies presented thus far generally show an influence of the non-target language on target word recognition, which suggests that word recognition processes are language-nonselective. However, since words are not normally encountered in isolation, it might be more interesting to investigate whether co-activation of elements of the non-target language also occurs during word recognition within a sentence context.
Bilingual word recognition in context
A study by Elston-Güttler, Gunter and Kotz (2005) provided relevant information on the way context might modulate word recognition in bilinguals. German-English bilinguals were shown English sentences in which the final word was either a German-English interlexical homograph or an unrelated English prime. This sentence was followed by an English target word that expressed the German meaning of the homograph. Participants were instructed to perform English lexical decisions on the targets and researchers collected both behavioural responses and event-related brain potentials (ERPs) to the critical stimuli. In order to clarify this, consider the German-English homograph gift (which means poison in German, but present in English). Participants could be shown the sentence the woman gave her friend an expensive gift, which would be followed by the
target word poison. Alternatively participants could be shown the sentence the woman gave her
friend an expensive item, which would also be followed by the target word poison. Thus, in the
former scenario the prime gift is followed by the target poison, whereas in the latter scenario the prime item is followed by the target poison. If word recognition in context is language-nonselective, then the homograph’s German meaning should be processed in addition to its English meaning. This then implies that, because the target is related to the German meaning of the prime, semantic priming effects will be observed (i.e. faster response times for related targets than for unrelated targets). Elston-Güttler et al. (2005) showed that a pattern consistent with language-nonselectivity can only be produced under very specific circumstances. In all but one of the conditions, neither the behavioural responses nor the event-related brain potentials (ERPs) differed between related and unrelated targets. Thus, word recognition in context appears to be a primarily language-selective process.
Schwartz and Kroll (2006) also examined word recognition within context, but they
employed a different design, a different task and (partially) different stimuli. In this study the targets to which the participants responded were either homographs, cognates or control words and these appeared in the middle of English sentences. If a target word appeared in red, then participants should read this word out loud. The words preceding and following the target did not require an overt response. The target words were embedded in two types of sentence conditions: high and low constraint. Constraint was operationalized as the extent to which the sentence frame preceding the target word biased that word. Both balanced and unbalanced Spanish-English bilinguals (i.e. whose Spanish was clearly stronger than English) were tested in this experiment. Balanced bilinguals responded equally fast, and made the same amount of errors, to both homographs and controls in both the high- and low-constraint conditions (i.e. no homograph effect was found). However, for unbalanced bilinguals a homograph effect that interacted with sentence constraint was found. This implied that, for unbalanced bilinguals, the stronger language (i.e. Spanish) was co-activated with the weaker language (i.e. English) whenever the context would not indicate the homograph’s
intended meaning. Effects of cognate facilitation (faster responses to cognates than to controls) persisted in low-constraint sentences, but were eliminated in the high constraint sentences for both balanced and unbalanced bilinguals.
To summarize, evidence from sentence context studies suggests that both highly and weakly constraining sentence context prevents co-activation of an interlexical homograph’s representation in the non-target language. Additionally, highly constraining sentence context prevents co-activation of a cognate’s representation in the non-target language. These findings indicate that word
recognition within context is language-selective. In contrast, in a weakly constraining sentence context cognates were named faster than their non-cognate controls, a finding that suggests language-nonselective processing.
Bilingual word production
The evidence presented thus far suggests that bilingual word recognition out-of-context is generally language-nonselective. Additionally, when the to-be-recognized words are presented within a sentence context, then the exact circumstances dictate whether word recognition is language selective or language-nonselective. However, this does not necessarily indicate that word production functions in a similar fashion, as there are substantial differences between word
recognition and word production. Most of the studies that assess whether bilingual word production is either language-selective or language-nonselective have examined word production in isolation. The task that is most frequently used in these studies is (some version of) the picture-naming task. The reason for this is that the outcome of the picture analysis process resembles the output of the mental conceptualization process in natural speech production, and this output is thought to set off the remainder of the production process in the same way in picture naming and common speech production (de Groot, 2011). Several experiments, in which picture-naming was used in order to assess bilingual word production, will be discussed here.
Costa, Caramazza and Sebastián-Gallés (2000) explored the effect of the cognate variable in picture naming. By comparing the responses for naming both pictures that represent cognate and non-cognate words Costa et al. (2000) showed that pictures with cognate names were named faster than pictures with non-cognate names. This so-called cognate effect (i.e. faster response times for cognates than for non-cognates) was observed both with naming in the dominant language and with naming in the non-dominant language. Costa et al. (2000) explained the occurrence of a cognate effect by describing what happens when Spanish-Catalan bilinguals perform a picture-naming task in which pictures have to be named in Spanish (Figure 1). At a certain point in time a Spanish-Catalan bilingual will be presented with a picture of a cat. This will iniatially activate the semantic representation of a cat. This semantic representation is shared between the bilingual's two
languages. Next, the semantic activation is passed on to a lexical level. Here it activates the appropriate lexical nodes, one for each language, which implies that both gato and gat become activated. Finally, the activation from both lexical nodes reaches the sublexical level, which
contains phonological units and is once again shared between a bilingual's two languages. Because the Spanish-Catalan cognate pair gato-gat share a substantial amount of phonological units (i.e. /g/, /a/, /t/), some of these phonological units not only receive activation from the contextually appropriate lexical node, but from the contextually inappropriate lexical node as well (Costa et al., 2000).This additional activation should ensure that cognate words are named faster than
non-cognate words.These findings show that lexical nodes in the non-target language are activated during word production, which suggests that bilingual word production is language-nonselective.
Another piece of evidence comes from studies in which the picture-word interference paradigm was used. In this paradigm the picture is accompanied by a distracter word and the influence of this distracter on picture-naming performance is determined. Hermans, Bongaerts, De Bot and Schreuder (1998) used a picture-word interference paradigm in order to examine whether, when bilinguals were asked to name pictures in their second language, the representations of the corresponding translations in their first language were also activated. In order to examine this Dutch-English bilinguals were asked to name pictures in English. These pictures (e.g. a picture of a peacock) were presented alongside a distracter that was phonologically similar to the translation of the pictures name (e.g. power for pauw, where pauw is the Dutch translation of peacock). Hermans et al. (1998) showed that these phonological translation distracters slowed down picture naming (i.e. presenting power alongside a picture of a peacock slows down picture naming). These results indicate that, when bilinguals have to name pictures in their second language, the representation of the corresponding translation in their first language is active as well. In other words, the results suggested that bilingual word production is language-nonselective.
A third piece of evidence that suggests that word production might be language-nonselective comes from studies in which participants had to perform a task called phoneme monitoring. For instance, in a study by Colomé (2001), Catalan-Spanish bilinguals were presented with pictures and they had to provide the picture's Catalan name. However, instead of being asked to name each picture out loud they had to generate its name covertly and to then monitor the internally generated name for the presence of a certain specified phoneme. Thus, the participants’ task was to indicate whether or not the phoneme occurred in the picture’s name in Catalan. The participants were presented with three different types of trials. The first type of trials were those in which the specified phoneme was present in the Catalan name of the picture. The second type of trials were those in which the specified phoneme was absent in both the Catalan and Spanish name of picture.
The third, and final, type of trials were those in which the specified phoneme was absent in the picture's Catalan name, but present in the picture's Spanish name. If the picture’s name is encoded phonologically in both of the bilingual's languages, then the latter type of trials would be hard to dismiss, because the specified phoneme would be activated and cause interference. Colomé showed that response times were indeed longer for this latter type of trial, thereby providing evidence of language-nonselective word production.
The evidence presented in this section suggests that isolated word production in bilinguals, as assessed by means of (different versions of) the picture-naming task, is most likely language-nonselective.
Bilingual word production in context: Starreveld et al. (2014)
Even though word production hardly ever occurs in isolation, there are very few studies that have examined word production in context. A study provided by Starreveld et al. (2014) is one of the first studies that examines whether co-activation of the non-target language occurs during word
production in a sentence context. Starreveld et al. (2014) report multiple experiments in which Dutch-English bilinguals performed a picture naming task either in English or in Dutch. Pictures' names were either Dutch-English cognate words or non-cognate words. With this set-up, the cognate effect can be used as a marker that indicates whether elements of the non-target language are active.
In the first of these experiments, in which pictures were presented in isolation, a cognate effect was observed both when pictures were named in English and when pictures were named in Dutch. This effect was larger when pictures were named in English. These results correspond well with Costa et al. (2000) and suggest that isolated word production is language-nonselective. In their second experiment Starreveld et al. (2014) embedded the same pictures that were used in the first experiment within a visually presented sentence context. Each of these pictures was presented twice in each language condition, once in a high-constraint sentence context and once in a low constraint
sentence context. In a so-called high-constraint condition the pictures' names were highly
predictable from the prior sentence context, whereas in the low-constraint condition they were not. The results showed that a cognate effect was found whenever Dutch-English participants named pictures in a sentence context in English. This effect was found in both high-constraint and low-constraint sentences. However, when the same task was performed in Dutch, a cognate effect was observed only when the picture’s name could not be predicted based on prior sentence context. This suggests that, if the participant's weaker language is the target language, then word production in a sentence context is language-nonselective. But, if the participant's stronger language is the target language, then word production is only language-nonselective when the lead-in sentences are not predictive of the upcoming pictures.
Finally, in order to assess the influence of language and sentence context on the cognate effect directly, the data from their first experiment (i.e. word production in isolation) and parts of data of their second experiment (i.e. the low-constraint condition) were combined. Analysis of this dataset showed that pictures were named faster in Dutch than in English, pictures were named faster when presented in a low-constraint context than when presented in isolation, pictures with cognate names were named faster than pictures with non-cognate names and - most importantly – the size of the cognate effect when pictures were named in isolation and in English was significantly larger than the size the cognate effect in all other conditions (i.e. Dutch naming in isolation, Dutch naming in context and English naming in context). This final finding implies that, whenever a bilingual's weaker language is the target-language, embedding the to-be-named pictures in a low-constraint sentence context leads to a significant reduction in the size of the cognate effect. Thus, embedding these pictures in a low-constraint sentence context significantly reduces the extent to which word production is language-nonselective. Starreveld et al. (2014) explained this finding in terms of differential activation of the memory structures in the language system. They hypothesized that a sentence context may boost the activation of all elements in a target language system. This boost may be caused by a 'language cue' (La Heij, 2005), which is activated more when pictures are
presented in a context than when they are presented in isolation. This would then lead to a more highly activated target language system, which in turn leads to a reduction in the size of the cognate effect.
The present study: bilingual word production in various types of context
The present study builds upon the study conducted by Starreveld et al. (2014) by means of identifying the aspects of a sentence context that are necessary in order for context to be able to reduce the size of the cognate effect. However, because it is not feasible to assess the influence of every aspect of sentence context, it was chosen to investigate whether either grammar, phonology or semantics (or any combinations hereof) was responsible. The present study encompassed an
experiment in which several groups of Dutch-English bilinguals performed a picture-naming task. The pictures' names were either cognates (25%) or non-cognates (75%). These pictures could be presented either in isolation or within different types of sentence context. In some of these sentence contexts certain aspects of normal sentence context were missing. By comparing the size of the cognate effect for the various types of context used, the types of context that cause a reduction in the size of the cognate effect can be identified. Next, by carefully analysing the contents of these types of context, the responsible aspects can be identified. If the hypotheses provided by Costa et al. (2000) and Starreveld et al. (2014) are correct, then this experiment can be expected to produce three distinctive effects. Firstly, a cognate effect should be obtained. Pictures with cognate names should be named faster than pictures with non-cognate names. Secondly, an effect of context should also be obtained. Pictures presented within a sentence context should be named faster than pictures presented in isolation. Thirdly, the size of the cognate effect should depend upon the type of context that will be used. The cognate effect associated with pictures presented in isolation should be larger than the cognate effect associated with pictures presented within a sentence context. However, it is currently not possible to predict if the remaining context conditions (detailed below) affect the cognate effect. The reason for this is that, as far as we know, the present study is among the first to
investigate which aspects of a sentence context are necessary in order for context to be able to reduce the size of the cognate effect.
Method
Participants
In total seventy-nine students from the University of Amsterdam (UvA) participated in the current experiment. Fifty-two (65.82%) of them were female and they were on average 21.35 (SD: 2.21) years old. Participants responded to an advertisement, which was posted both within the university and online, that asked first-year, Dutch-English bilingual students (who had Dutch as their native language) to sign up. According to Starreveld et al. (2014) these students share the following characteristics: They started training in English when they were about 10 years old and
subsequently received several hours of training each week until the age of eighteen. Furthermore, they are required to read English texts regularly after they have enrolled in university. This ensured that their English is usually well developed. It is, however, not as strongly developed as their Dutch, which they have used throughout their entire lives.
Participants were subsequently assigned to conditions in a quasi-random fashion (i.e. they were alternately assigned to the conditions as they came to the laboratory). This meant that the first participant went into the first condition, the second participant into the second condition and so forth. This resulted in the following distribution of participants over the various conditions (to be described in more detail in the materials and apparatus section): 191 in the without context condition, 12 in the sentence context condition, 12 in the scrambled context condition, 12 in the pseudowords condition, 11 in the semantic anomaly condition and 12 in the nonwords condition.2
1 More participants were assigned to the without context condition, because this was the only context condition that
immediately functioned as intended (i.e. at the start of the experiment the code in the other context conditions still contained some bugs that needed to be removed).
2 It was initially assumed that a minimum of at least twenty-five participants per condition would be required in order
to make statistically meaningful comparisons. However, due to an apparent lack of interest among potential participants in the experiment, it was not possible to test the required number of participants.
Three experimenters were involved in data collection. Each of them gathered portions of the data associated with each individual condition, so that potential experimenter effects would be evenly distributed across the various conditions (i.e. Experimenter 1 – condition 1 / 2 / ... / 6, Experimenter 2 – condition 1 / 2 /... / 6, Experimenter 3 – condition 1 / 2 /... / 6).
After completing the experiment participants rated their reading and speaking abilities in Dutch and English in an exit questionnaire. They were asked to use a scale from 1-10, where a score of 10 indicated that their self-perceived level of proficiency was outstanding, while a score of 1 indicated that they thought it was very poor. The average reading ability scores of the participants were 9.48 (SD: 0.78) and 8.21 (SD: 0.98) for Dutch and English, respectively. The corresponding average speaking abilities were rated with a 9.48 (SD: 0.61) and a 7.60 (SD: 0.89). These scores confirmed that participants were more fluent in Dutch than they were in English.
Participants received either course credit (1 RC) or a financial compensation (€10.00) for their efforts.
Research design
The present study employed a 2x6 design with both a within- and a between-subject variable. The independent variables were cognate status (with two levels: cognate and non-cognate) and context condition (with six levels: without context, sentence context, scrambled context, pseudowords, semantic anomaly and nonwords). In the participant analysis cognate status was a within-subject variable and context condition a between-subject variable, whereas in the item analysis cognate status was a between-item variable and context condition was a within-item variable.The dependent variable was response time, which was measured in milliseconds.
Research strategy
In total, six different context conditions were created (Table 1). Two of these (i.e. without context and sentence context) closely resembled the ones used by Starreveld et al. (2014). They were
included in order to replicate the findings by Starreveld et al. (2014). The remaining four conditions (i.e. scrambled, pseudowords, semantic anomaly and nonwords) were created in order to examine which aspects of context might cause a reduction in the size of the cognate effect. These conditions were each constructed in such a way that they lacked specific components of normal context (Table 2). A distinction between the following components of context was made: grammar, phonology and whether the context is meaningful (on both a word and a sentence level).
Table 1: Various Experimental conditions
Context condition Description Cognates
Without context Pictures only “(ANCHOR)” Sentence context Subset of sentences used by
Starreveld et al. (2014) “In the middle of the square was an (ANCHOR) with a thick chain attached to it. ”
Scrambled Random word order “The square in middle an of was the (ANCHOR) chain with thick to it attached a.”
Pseudowords Most words have been replaced by
pseudowords “In the miggle of the squank wiss an (ANCHOR) wime a thack chail attinced to it.”
Semantic anomaly Replaced words in order to make
sentences meaningless “In the story of the square was an (ANCHOR) with a thick rain attached to it.”
Nonwords All words have been replaced by
nonwords “Uq ssr naehte hd ofw liatdi enm ea(ANCHOR) iatk e hiohc ahnit ctcaitdw th ta.”
Note: Additional details about the various conditions can be found in the 'Materials and apparatus' section.
Table 2: Components of context that are present in the various experimental conditions
Context condition Grammar Phonology Semantics (word)
Without context - - -Sentence context + + + Scrambled - + + Pseudowords + + -Semantic anomaly + + + Nonwords - -
-Note: A plus sign (+) indicates the presence and a minus sign indicates the absence (-) of a component.
By manipulating the type of context in which pictures are presented it will be possible to identify the types of context that modify the (size of the) cognate effect. Next, by carefully analysing the contents of these types of context, we can identify the aspects of context that are necessary for such a modification to occur. In order to clarify this, consider the following (abstract) example:If we observe that a sum of parts causes an effect (i.e. part A + part B + part C → effect
causes the effect to occur (e.g. part A + part B → effect Z?). If effect Z still occurs in this scenario, then we can deduce that part C did not contribute to effect Z and that part A, part B, or part A + B are necessary for the effect to occur. On the other hand, if effect Z does not appear anymore then part C must be vital to its occurrence.Thus, if we apply the same logic to the cognate effect and the various context conditions, then we could argue that the components of context that cause a
modification of the cognate effect can be identified by carefully analysing the contents of the conditions in which such a modification occurs. For example: when we compare the sentence context condition with the semantic anomaly condition we can see that the only difference between the two is the fact that the latter has no meaning on the sentence level. Therefore, any difference in the size of the cognate effect between these two conditions has to be caused by the lack of meaning on the sentence level.
Materials and apparatus
Picture stimuli: The current experiment uses the same pictures that were used by Starreveld et al. (2014). This implies that a hundred pictures of objects were used in total, which were obtained from a corpus of black-and-white line drawings that was made available by Székely, D'Amico, Devescovi, Federmeier, Herron, Iyer, Jacobson and Bates (2003). Fifty of these served as so-called critical pictures. Twenty-five of these critical pictures depicted Dutch-English cognates (e.g.
anker-anchor) and these functioned as the material for the cognate condition. The other twenty-five
critical pictures represented Dutch-English non-cognates (e.g. bril-glasses) and these functioned as the material for the non-cognate condition. The remaining fifty pictures served as non-critical pictures. They represented Dutch-English non-cognates and functioned as fillers. Furthermore, these pictures were also presented in warm-up trials and after an error was made. Both the pictures themselves and the ratio of pictures that represent cognates, non-cognates and fillers (25, 25, 50) was kept constant across all conditions. For more detailed information about the pictures that were used, please consult Starreveld et al. (2014).
Without context: In this condition pictures appeared in isolation (e.g. a picture of an
(ANCHOR), see Table 1 for another example).
Sentence Context: In this condition pictures were embedded in well-formed, meaningful sentences (see Table 2). These sentences were a subset of sentences used by Starreveld et al. (2014). This subset included the sentences that Starreveld et al. (2014) labelled as 'low constraint', which implied that the pictures' names were not highly predictable from the prior sentence context. In addition, half of their filler material was used as well. Sentences which contained critical pictures had a similar grammatical structure: The first part of the sentence was a main clause in which the final word was replaced by the picture that had to be named. The sentence then continued with a subordinate clause, another main clause, or a prepositional clause (e.g. In the middle of the square
was an (ANCHOR) with a thick chain attached to it, see Table 1 for another example). Sentences
which contained filler pictures had a somewhat similar structure. However, in these sentences pictures did not necessarily appear in the middle region, but could also appear elsewhere in the sentence. This was done in order to prevent participants from anticipating the occurrence of the picture, while ignoring the remaining parts of the sentence. Spreading the appearance of the picture over multiple sentence regions would discourage such a strategy.
Scrambled: In this condition the sentences, in which the pictures were embedded, had a largely random word order(e.g. The square in middle an of was the (ANCHOR) chain with thick to
it attached a, see Table 1 for another example). In order to achieve this the sentences used in the
sentence context condition were cut into two parts, namely, the part that precedes the picture and the part that follows the picture. Afterwards, both parts were entered independently in a automatic word scrambler (www.thecanadianteacher.com/tools/games/sentence/) and the resulting parts were joined together again. Finally, these sentences were checked manually (and, if necessary, corrected) for both the presence of identical words that appeared in succession (e.g. “the the”) and to see whether there were no parts of the new sentence that were identical to the original. This procedure ensured
that the sentences that were used in this context condition were grammatically incorrect and no longer had any meaning (see Table 2).
Pseudowords: Pictures were once again embedded in modified versions of the sentences used in the sentence context condition. These modified sentences were comprised of a mix of normal words and pseudowords. Pseudowords are words that are phonologically similar to normal words, but they lack meaning. All words that make up the sentences in the sentence context condition - except articles, personal pronouns, possessive pronouns and small prepositions (small was defined as three letters or less) – were replaced by pseudowords (e.g. In the miggle of the squank wiss an
(ANCHOR) wime a thack chail attinced to it, see Table 1 for another example). This implied that the
sentences that were used in this context condition had no meaning, but still possessed (some form of) normal grammar and phonology (see Table 2).Pseudowords were automatically created by a multilingual pseudoword generator known as 'Wuggy' (Keuleers & Brysbaert, 2010). Afterwards, three independent evaluators assessed whether the generated pseudowords sound like actual English words. If two out of three evaluators thought this was not the case, then these pseudowords were replaced by manually generated pseudowords.
Table 3: Kappa statistics for pairs of raters
Rater 1 2 3
1 - 0,15 0,38
2 - 0,08
3
-An inter-rater reliability analysis based upon the Kappa statistic was performed in order to determine consistency among raters. The Kappa statistic is a statistical measure that signals agreement between two raters. For designs with three or more raters Light (1971) suggests computing Kappa for all pairs of raters (Table 3)and subsequently calculating the mean of these estimates in order to generate an overall index of agreement. These three Kappa statistics yielded an overall index of agreement of 0.20, which signaled only slight agreement between raters (Landis &
raters. It is entirely possible that a different set of results would have been obtained when native English speakers had been used to assess the phonology of the generated pseudowords.
Semantic anomaly: In this condition pictures were embedded in sentences. However, the sentences that were used were again modified versions of the ones used in the context condition. In each sentence two words were replaced by other, semantically non-fitting, words (e.g. In the story
of the square was an (ANCHOR) with a thick rain attached to it, see Table 1 for another example).
This modification ensured that these sentences became meaningless (see Table 2). Preferably, nouns were replaced, but if this was not possible, adjectives or verbs were replaced.
Nonwords: In this condition the sentences in which pictures were embedded contained nonwords only (e.g. Uq ssr naehte hd ofw liatdi enm ea (ANCHOR) iatk e hiohc ahnit ctcaitdw th
ta, see Table 1 for another example). In order to create nonwords the sentences that were used in the
sentence context condition were initially cut into two parts (in the same way as in the scrambled condition). These parts were then entered individually in an automatic letter scrambler
(www.bluestwave.com/toolbox_letter_scrambler.php). Afterwards, the resulting sentence parts were joined together again. It was ensured that the length of each nonword corresponded with the length of the original word. Next, the resulting sentences were checked for the existence of known English and Dutch words and, if these were found, they were replaced by manually generated nonwords . Finally, single-letter English words were consistently replaced by a single letter nonword (i.e. “I” and “a” were always replaced by “n” and “e” respectively). This procedure resulted in sentences that were grammatically incorrect, had no audible phonology, and were entirely meaningless (see Table 2).
Apparatus: The experiment was performed using Presentation software (version 9.90,
www.neurobs.com). Pictures were presented on a fast cathode ray tube monitor running on a refresh rate of 70 Hz. Response times were collected using a voice key and measured up to the nearest millisecond. Response time was defined as the time duration between picture onset and the moment at which a participant initiated a vocal response.
Procedure
Participants were tested in a quiet, well lit room and were seated in front of a computer monitor. All communication with the participants was in English. Once participants were comfortably seated they were first handed an instruction sheet. If participants did not have any questions or objections after reading the instructions they were then handed an informed consent form, which they had to sign in order to partake in the experiment. Next, in order to familiarize themselves with the experimental materials, participants were handed a booklet which contained all the pictures (and their names) that were used in the experiment. Participants were asked to thoroughly study the booklet and to only use the names that were provided. When participants felt they had sufficiently studied the booklet, the picture-naming phase of the experiment began.
In the picture-naming phase participants were asked to name the pictures in English as quickly and as accurately as possible. This phase started with the presentation of twenty warm-up trials that allowed participants to familiarize themselves with the upcoming task. The pictures that were used in these warm-up trials were selected at random from the non-critical (i.e. filler)
materials. After the warm-up trials had been completed, all critical and non-critical pictures (a hundred in total) were presented in four blocks. Participants were allowed to take small breaks between blocks. The sequence in which the pictures were presented was randomized anew for each participant. Furthermore, after participants had made an error or when the voice key malfunctioned an additional trial was presented (these additional trials always used a randomly selected non-critical picture).
The picture-naming trials themselves differed between conditions, because pictures in the without context condition were presented in isolation, whereas the pictures in the remaining five conditions were presented within a context. A description of picture-naming trials in the without context condition will be provided first, followed by a description of what the trials looked like for the other five conditions.
Picture-naming trials in the without context condition began with the appearance of a fixation point (i.e. a “+” sign) in the center of the screen, which remained on the screen for 500 ms.
Afterwards a blank screen was presented for 500 ms, followed by the presentation of the picture. The picture remained on the screen either until the participant provided a response or until 2500 ms had gone by. The experimenter subsequently coded the response, thereby indicating whether the response was correct, false, or whether the voice key malfunctioned. Finally a blank screen was presented for another 500 ms, after which the next trial began.
In the context conditions participants were faced with a so-called self-paced reading task (SPRT). In this task a sentence is presented word by word and participants can decide for
themselves when they would like to move on to the next word. Picture-naming trials began with the appearance of a fixation point (i.e. the '+' sign) in the middle of the screen. After participants
pressed the space bar the fixation point was replaced by the first word, pseudoword, or nonword of a sentence. When they pressed the button again, the word on the screen was replaced by the next word of the sentence. This continued until a picture, instead of a word, appeared. At that point the participant named the picture as fast and as accurately as possible. The experimenter subsequently coded the response, thereby indicating whether the response was correct, false, or whether the voice key malfunctioned. The coding of the response triggered either the appearance of the next word of the sentence or, if the picture was located at the very end of the sentence, the start of the next trial. An empty screen was presented for 1000 ms between the presentation of two sentences.
Finally, after the picture-naming phase had been completed, participants were handed a questionnaire which asked them (among other things) to rate their language proficiency, to indicate whether they had ever been diagnosed with dyslexia, and to provide information regarding self-perceived task performance. The whole experiment took about 45 minutes to complete.
Results
Data collected for both 'warm-up' and filler pictures were not analyzed. Next, response times (RTs) from incorrect responses and from trials in which the voice key malfunctioned were removed. Furthermore, RTs shorter than 300 milliseconds or longer than 2000 milliseconds were removed as well. In total these exclusions accounted for 8.56%, 4.56%, and 0.86% of the data. The remaining RTs were used for the calculation of the means for the various conditions.
An analysis of variance (ANOVA) was performed on the mean RTs per participant per condition, with cognate status (two levels: cognate and non-cognate) as a within-subjects variable and condition (six levels: without context, sentence context, scrambled, pseudowords, semantic anomaly and nonwords) as a between-subjects variable. The corresponding item analysis was also performed on the mean RTs per item per condition, with condition (six levels: without context, sentence context, scrambled, pseudowords, semantic anomaly and nonwords) as a within-items variable and cognate status (two levels: cognate and non-cognate) as a between-items variable. If the assumption of sphericity was violated, then a Greenhouse-Geisser correction was applied. Note that, because cognate status is a within-subjects variable in the participant analysis, the means that enter this analysis are averaged over pictures. This implies that the nuisance variance caused by differences in response speeds for different pictures within a condition is excluded. However, because cognate status is a between-items a variable in the item analysis, the means that enter this analysis are averaged over participants. This means that the nuisance variance caused by differences in response speeds for different pictures cannot be excluded. Therefore, the power to detect a cognate effect is much higher in the subject analysis than in the item analysis.
The effect of cognate status was significant, F1(1, 73) = 113.22, MSE = 276,631, p < .001 and
F2(1,48) = 9.94, MSE = 525,772, p = .003. Participants responded faster in the cognate condition (M
= 836.37 ms) than in the non-cognate condition (M = 922.63 ms) (see Table 4). The effect of context condition was not significant in the participant analysis, F1(5,73) = 0.84, MSE = 15,996, p =
the sentence context condition (M = 858.28 ms) responded the fastest, followed – respectively – by the scrambled condition (M = 859.98 ms), pseudoword condition (M = 869.27 ms), without context condition (M = 871.83 ms), semantic anomaly condition (M = 897.45 ms) and finally the nonwords condition (M = 922.56 ms) (see Table 4). The fact that the results of the item and participant
analyses do not match with regard to this effect, implied that it remains uncertain whether response times for naming pictures in general depend upon the type of context in which they are presented. The interaction between cognate status and condition was not significant, F1(5,73) = 0.76, MSE =
1,868, p = .58 and F2(4.02;192.85) = 0.98, MSE = 6,422, p = .42. This implies that the effect of
cognate status was similar for all levels of the context condition variable. This result does not match with the expectations. Based on previous research (Starreveld et al., 2014) it was expected that the effect of cognate status would be significantly smaller in the sentence context condition than in the without context condition. The fact that no interaction was found between cognate status and context condition suggests that the effect of cognate status does not depend on the type of context in which pictures were embedded.
Table 4: Participant mean reaction times (ms) per condition and error percentages (in parentheses)
Context condition Cognate Non-cognate Average
Without context 822.62 (6.32) 921.05 (12.63) 871.83 (9.47) Sentence context 813.39 (6.00) 903.17 (12.67) 858.28 (9.33) Scrambled 826.80 (4.67) 893.17 (9.33) 859.98 (7.00) Pseudowords 838.92 (4.00) 899.63 (10.00) 869.27 (7.00) Semantic anomaly 849.54 (6.55) 945.36 (14.55) 897.45 (10.55) Nonwords 873.03 (4.62) 972.09 (10.77) 922.56 (7.69) Average 836.37 (5.42) 922.63 (11.70) 879.50 (8.56)
In order to check whether speed-accuracy trade-offs may have occurred, the corresponding error analyses were also performed. They showed a significant main effect of cognate status, F1(1,73) =
42.10, MSE = 1,519, p < .001 and F2(1,48) = 4.94, MSE = 2,854, p = .03. More errors were made
while naming pictures with non-cognate names (M = 11.70%) than while naming pictures with cognate names (M = 5.42%). The effect of context condition was not significant, F1(5,73) = 0.91,
errors made did not depend on the type of context in which pictures were presented. Finally, the interaction between cognate status and condition was not significant, F1(5,73) = 0.19, MSE = 7, p
= .97 and F2(3.76;180.41) = 0.36, MSE = 26, p = .82. This means that the effect of cognate status
was similar for all levels of context condition (i.e. the way errors were distributed over the two levels of cognate status was similar for all levels of context condition). Thus, we conclude that speed-accuracy trade-offs were absent in the data.
General discussion
Does the effect of cognate status depend on the type of context used and, if so, which aspect(-s) of context is (are) responsible for such a dependency?
The present study examined whether the effect of cognate status depends on the type of context in which pictures were presented. The analyses revealed that the (size of the) cognate effect was similar for all the different types of context used. Therefore, the effect of cognate status was independent of the type of context used. This finding contradicts earlier research (Starreveld et al., 2014), in which context did modify the effect of cognate status (i.e. a smaller cognate effect was found for pictures embedded in a sentence). Furthermore, the fact that no support was found for the notion that context modifies the effect of cognate status, implied that the second part of the research question could not be answered (i.e. it makes no sense trying to attribute an effect to aspects of context when no effect is found in the first place).
Differences between Starreveld et al. (2014) and the present study
The current results disagree with the study done by Starreveld et al. (2014). How is it possible that the results of two similar studies seem to contradict each other? Disagreement between the two studies was most likely caused by the fact that the present study was unable to sufficiently
ignored the context, then the context condition variable no longer functioned as intended. This means that it would no longer be possible to distinguish between the various levels of the context condition variable, because the task would be identical for all levels of this variable (i.e. press a button until a picture appears and, when it does, name the picture as fast and accurately as possible). A variable that does not function as intended creates a situation in which there is a difference between what was intended to be measured and what was actually measured (i.e. not without context vs. sentence context vs. scrambled etc. but rather without context vs. without context vs. without context etc.). This would in turn imply that the meaning that was attributed to the results of certain analyses might have been incorrect.
Furthermore,Starreveld et al. (2014) included a sentence constraint variable (two levels: high- and low-constraint) that was not included in the present experiment. In the high-constraint condition the pictures' names were highly predictable from the prior sentence context (e.g. One
usually eats soup with a (SPOON), because it is a lot trickier with a fork), whereas in the
low-constraint condition they were not (e.g. She gave the guest a (SPOON) so he could eat his pudding). High- and low-constraint sentences were presented in random order, but because especially in the high-constraint sentences reading the context helps participants with the task at hand (i.e. picture naming) this might have caused participants to assign more value to the context as a whole. In contrast, in the present experiment lead-in sentence fragments were never predictive of the upcoming pictures, thus reading the context provided participants with no additional benefit. Additionally, Starreveld et al. (2014) used comprehension questions in order to encourage
participants to pay attention to the sentence contexts (and to evaluate whether they indeed did so) whereas the present study did not (detailed below). These differences suggest that participants in the current study may have ignored the context.
Finally,disagreement between the two studies could also have been due to differences in the number of participants per condition. Starreveld et al. (2014) used (on average) twenty-three
recruit a number of participants per condition that was (at least) equal to the number used by Starreveld et al. (2014). However, it proved to be a lot harder than anticipated to find suitable participants, because students showed very little interest in signing up for the present study (See footnote 2 on page 15). This, in turn, reduced the chance that the present study would detect an effect of the critical variables.
Ignoring the context
As mentioned previously, the fact that the results of the present study did not agree with Starreveld et al. (2014) might have been caused by the fact that participants in the present study paid no attention to the context.In order to prevent these problems from occurring, participants should not have been able to ignore the context, as differences between the various levels of context condition disappear when the context is ignored. Starreveld et al. (2014) used three different methods in order to prevent this from happening, whereas the present study only used two. Firstly, in both studies participants were told that it was very important for them to read the context carefully. Secondly, both studies varied the location of the picture within a (filler) sentence. This implied that the occurrence of a picture could not be accurately predicted over the course of the experiment, which ensured that participants had to continuously pay attention to the context. As far as preventing the context from being ignored is concerned, the main difference between the two studies was that Starreveld et al. (2014) used comprehension questions, whereas the present study did not. Starreveld et al. (2014) used yes/no comprehension questions regarding an immediately preceding filler sentence in order to encourage participants to pay attention to the sentence context. Such questions could not be used in the current study, because the sentences that make up the context are generally meaningless. In other words, the information that is needed to accurately answer comprehension questions was not available in the context. Furthermore, comprehension questions could not be used in the present experiment, because they would have contaminated the context manipulation with variation in sentence processing. Consider, for instance, a situation in which participants first read a
sentence in the scrambled context condition and afterwards they would read a comprehension question that has a normal sentence structure. This is a situation that repeats itself whenever there is a new comprehension question to be answered. Therefore, participants would continuously be alternating between reading scrambled and normal sentences. This implies that the context in the scrambled context condition could no longer be regarded as truly scrambled, as the scrambled context had become 'contaminated' with normally structured questions. This contamination would make it difficult to interpret the results, because it can no longer be said which part of this
contaminated context would be responsible for potential effects (i.e. effects could be caused by either the scrambled part, the normal part or a combination of both). At this point it should be clear as to why comprehension questions could not be used in the current experiment. Thus, in order to prevent the context from being ignored, alternative strategies should be considered for use in future experiments.
Improving the methods used
Can it be ensured that participants attentively read a sentence context which they do not necessarily understand? One possible solution for this problem would be to replace the comprehension
questions that were used by Starreveld et al. (2014) by delayed match-to-sample tasks. In delayed match-to-sample tasks participants are first shown an initial stimulus that should be kept in mind. Next, after a brief delay, they are shown several other comparison stimuli (one of which matches the initial stimulus). Participants have to indicate, by means of a forced-choice response, which of the comparison stimuli matches the initial stimulus. As delayed match-to-sample tasks are included in the present experiment in order to prevent participants from ignoring the context it should be obvious that the context itself is the initial stimulus. Comparison stimuli can be made in a myriad of different ways by altering the appearance of the context (for specific examples see Table 5). At random positions across the experiment (but always after a picture-naming trial has been
have to indicate which one of these matches the initial stimulus. It could be argued that the number of errors while executing this task, reflects the extent to which participants have observed and remembered the initial stimulus. A high number of errors would indicate that participants were inattentive while viewing the initial stimulus, whereas a low number of errors would indicate that participants were attentive while viewing the initial stimulus. In this respect delayed match-to-sample tasks function in a similar manner as comprehension questions, but the main difference is that participants do not need to understand an initial stimulus in order to provide correct answers.
Table 5: Various examples of potential initial + comparison stimuli in delayed match-to-sample tasks Initial stimulus Correct comparison
stimulus
Incorrect comparison stimulus
“In the miggle of the squank wiss an (ANCHOR) wime a thack chail attinced to it.”
“wime a thack chail attinced
to it” “wime a thack chail attincedto it”
“In the miggle of the squank wiss an (ANCHOR) wime a thack chail attinced to it.”
“wime a thack chail attinced
to it” “wime a thack chail attinced to it” “In the miggle of the squank wiss
an (ANCHOR) wime a thack chail attinced to it.”
“wime a thack chail attinced
to it” (presented in black) “wime a thack chail attinced to it” (presented in red) In the miggle of the squank wiss an
(ANCHOR) wime a thack chail attinced to it.”
“wime a thack chail attinced
to it” “with a thick chain attached to it.”
However, even though the use of delayed match-to-sample tasks makes it possible to assess whether or not participants have paid attention to an incomprehensible sentence context, it simultaneously causes a different problem to occur. In order to answer these delayed match-to-sample tasks correctly participants have to pay attention to the appearance of the context. This implies that adding these tasks to the experiment causes a switch in participants' focus from the contents of context to the appearance of the context. Thus adding delayed match-to-sample tasks might cause confounding variables to affect the results as the current study no longer merely assesses the effects of language processing on word production, but it assesses the effects of perception and memory on word production as well. It is up to the individual researcher to decide whether this is worth the risk.
Conclusion
To summarize, the current study found that the effect of cognate status did not depend on the type of context in which pictures were embedded. Additionally, because the present study was unable to confirm that the effect of cognate status is contextually dependent, it was also unable to attribute such a dependency to either grammar, phonology or semantics (or any combinations hereof). The former finding did not correspond with earlier research in which it was found that the effect of cognate status diminishes when pictures are presented within a sentence context. Disagreement between these studies was most likely caused by the fact that participants in the current study ignored the context. If the context was ignored, then the context condition variable no longer functioned as intended. This, in turn, might have led to the drawing of incorrect conclusions (i.e. accurate conclusions regarding the effects of context cannot be drawn, because only a single type of context was effectively used in the present study). Even though the present research has its flaws (which has been demonstrated by participants' ability to ignore the context), the research strategy itself is valid. However, in order to prevent the context from being ignored, certain adjustments will have to be made. Delayed match-to-sample tasks could be included in order to assess whether participants have paid attention to an incomprehensible context, but the inclusion of these tasks introduces a new set of problems that will have to be addressed.
References
Colomé, A. (2001). Lexical activation in bilinguals' speech production: Language-specific or language-independent? Journal of Memory and Language, 54, 324-345.
Costa, A., Caramazza, A., & Sebastián-Gallés, N. (2000). The cognate facilitation effect:
implications for models of lexical access. Journal of Experimental Psychology: Learning,
De Groot, A. M. B. (2011). Language and Cognition in Bilinguals and Multilinguals: an
Introduction. Psycholy Press ltd.
Dijkstra, T., Van Jaarsveld,H., & Ten Brinke, S. (1998). Interlingual homograph recognition: Effects of task demands and language intermixing. Bilingualism: Language and Cognition, 1, 51-66. Dijkstra, T., Grainger, J., & Van Heuven, W. J. B. (1999). Recognition of cognates and interlingual
homographs: the neglected role of phonology. Journal of Memory and Language, 41, 496-518.
Elston-Güttler, K. E., Gunter T. C., & Kotz, S. A. (2005). Zooming into L2: Global language context and adjustment affect processing of interlingual homographs in sentences. Cognitive
Brain Research, 25, 57-70.
Hermans, D., Bongaerts, T., De Bot, K., & Schreuder, R. (1998). Producing words in a foreign language: Can speakers prevent interference from their first language? Bilingualism:
Language and Cognition, 1, 213-229.
Keuleers, E., & Brysbaert, M. (2010). Wuggy: a multilingual pseudoword generator. Behavior
Research Methods, 42, 627-633.
La Heij, W. (2005). Selection processes in monolingual and bilingual access. In J. F. Kroll & A. M. B. De Groot (eds.), Handbook of bilingualism: Psycholinguistic approaches, 289-307, New-York: Oxford University Press.
Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data.
Biometrics, 33, 159-174.
Light, R.J. (1971). Measures of response agreement for qualitative data: Some generalizations and alternatives. Psychological Bulletin, 76, 365–377.
Schwartz, A. I., & Kroll, J. F. (2006). Bilingual lexical activation in sentence context. Journal of
Starreveld, P. A., De Groot, A. M. B., Rossmark, B. M. M., Van Hell, J. G. (2014) Parallel language activation during word processing in bilinguals: Evidence from word production in sentence context. Bilingualism: Language and Cognition, 1-19.
Székely, A., D'Amico, S., Devescovi, A., Federmeier, K., Herron, D., Iyer, G., Jacobson T., & Bates, E. (2003). Timed picture naming: Extended norms and validation against previous studies.
