Language switching when writing: The role of phonological and orthographic overlap

Abstract

While language switching of bilinguals has been investigated extensively in the spoken domain, there has been little research on switching while writing. The factors that impact written language switching may differ from those that impact language switching while speaking. Thus, the study’s goal was to test to what extent phonological and/or orthographic overlap impacts written language switching. In four experiments (NExp.1 = 34; NExp. 2 = 57; NExp. 3 = 39; NExp. 4 = 39), German–English bilinguals completed a cued language switching task where responses had to be typed. To-be-named translation-equivalent concepts were selected to be similar phonologically, orthographically or neither. Participants switching between languages while writing was facilitated by both phonological and orthographic overlap. Maximum orthographic overlap between translation-equivalent words with dissimilar pronunciations facilitated switching to the extent that no switch costs could be observed. These results imply that overlapping orthography can strongly facilitate written language switching and that orthography’s role should be considered more thoroughly in models of bilingual language production.

Bilingualism is very common worldwide (Grosjean, 2010). When producing or comprehending one language, bilinguals have been found to also co-activate words from the other, nontarget language (e.g., Hermans et al., 1998; Meade et al., 2018; Spivey & Marian, 1999). As a result, control processes may be needed to limit cross-lingual competition. One situation in which competition may be particularly strong is when switching from one language to another. Although there has been a wealth of research on language switching in both comprehension and production, we still do not fully understand what types of processes support it (Declerck & Koch, 2022; Declerck & Philipp, 2015a).

The effect of switching on language production is typically studied in the spoken domain (see Iniesta et al., 2021; Wong & Maurer, 2021, for exceptions); that is, participants have to switch between speaking in their first language (L1) and their second language (L2). However, some bilinguals may predominantly switch when writing (e.g., in the work place, when replying to customers’ emails in both English, the L2, and German, the L1). Written language switching may be impacted by other factors that spoken language switching, such as orthographic overlap between translation-equivalent words instead of phonological one. The goal of this study therefore was to investigate how phonological and orthographic overlap impact language switching when writing.

Language switching (when speaking)

Language switching is not just an important behaviour in its own right; it also has been used extensively to study language control. Language control is the process by which one ensures that language production occurs in the target language (e.g., Declerck & Philipp, 2015a; Green, 1998). It is typically studied in the so-called language switching paradigm where bilinguals are presented with a stimulus (e.g., picture) that they have to name in their L1 or L2 (e.g., Christoffels et al., 2007; Declerck et al., 2017; Green, 1998; Meuter & Allport, 1999). Trials can be repetition trials, where the same language has to be produced as on a previous trial, or switch trials, where a different language than on the preceding trial has to be produced.

In general, performance on switch trials is worse than on repetition trials (i.e., slower, more erroneous; Declerck & Koch, 2022). Switch costs are then calculated by taking the difference in performance between the two trial types. These performance costs are often considered to be the result of inhibition (e.g., Green, 1998): When using one language, activation of the competing language has to be suppressed. This inhibition then persists into a following trial, which then may complicate activation of the previously suppressed language on a switch trial. One prediction consistent with this theoretical account is that switch costs can be asymmetrical in unbalanced bilinguals (Meuter & Allport, 1999): It is harder to complete a switch towards L1 because L1, as the more dominant language, must be inhibited to a greater extent when completing the proceeding L2 naming trial. In contrast, on a L1 trial, L2 does not require as much inhibition and can therefore be activated more easily on a subsequent trial. However, such asymmetrical switch costs are not always found and may be more dependent on moderator variables such as participant characteristics than previously thought (Declerck et al., 2012; Declerck & Koch, 2022; Declerck & Philipp, 2015a; Gade et al., 2021; Slevc et al., 2016).

Moreover, mixing costs can also be calculated by including nonmixed language blocks in the paradigm. One can then compare performance on repetition trials in mixed language blocks with performance in nonmixed ones. In general, performance for both languages is found to be worse in mixed than nonmixed blocks. These mixing costs have often been seen as evidence for the use of proactive language control: During nonmixed blocks, activation of the nontarget language can be proactively inhibited to reduce interference; at the same time, the target language can be activated in advance (Ma et al., 2016). In mixed language contexts, however, no proactive inhibition can occur, while both languages may be activated, thus leading to overall more interference. In addition, L1 is sometimes found to be more affected by mixing it with L2 than vice versa, sometimes to the extent that naming performance is worse for L1 than for L2 during mixed presentation (reversed language dominance effect or L1 slowing, for example, Christoffels et al., 2007). Interestingly, previous research has shown that L1 slowing and asymmetrical switch costs are rarely observed within the same data set (e.g., Christoffels et al., 2007; Mosca & Clahsen, 2016; Mosca & de Bot, 2017; Peeters & Dijkstra, 2018).

The role of phonological overlap during spoken language switching

When speaking a word, several different language processing stages are completed (e.g., Levelt et al., 1999): First, the abstract concept that one plans to convey is formed. Second, at the lemma level, syntactic information is added to the previously nonlinguistic concept. This is followed by a third level, the phonological one, where sound representations are added. The final level is the production of the concept via the respective articulators. These different stages can be serial and discrete (as described here), but do not have to be (Jescheniak & Schriefers, 1998). At the lemma level, representations are thought to compete with each other; in bilinguals, this is considered to happen nonselectively with words from both languages being activated (e.g., de Groot, 1992; Green, 1998). Traditionally, there was a general assumption that language control operates at the lemma level (Costa & Santesteban, 2004; Green, 1998; Meuter & Allport, 1999); as a result, the impact of co-activation at other levels due to overlap was not considered in detail. In general, the hypothesis is that depending on whether activation of both languages at a specific level leads to the selection of the correct word or not, one can observe a delay or facilitation (e.g., in RTs or error rate) in language production (Gollan & Kroll, 2001). Overlap in translation-equivalent words leads to co-activation and therefore competition between L1 and L2; how quickly and at what level this competition is resolved depends on a number of factors, such as the type of word, the task that is being performed, how strongly the nontarget word is activated as well as language proficiency (e.g., Dijkstra et al., 2010; Kroll et al., 2006).

One frequent way to examine the role of phonological overlap at different levels of bilingual language processing is to look at cognates (but also see Declerck & Philipp, 2015b; Goldrick et al., 2014 for other investigations on the impact of phonology). Cognates are here defined as translation-equivalent words with a similar etymological background, which often coincides with a large phonological overlap as for example in the word pairs BOOK/BUCH or SUN/SONNE [English/German] (Costa et al., 2000; Hoshino & Kroll, 2008). In a study by Declerck et al. (2012), German–English bilinguals completed a cued switching paradigm with digits, noncognates, cognates, and semantically related words. While in the noncognate stimulus set and the semantic set, translation-equivalent words did not sound similarly (e.g., HORSE/PFERD [English/German]), translation-equivalent words in the digit and cognate stimulus set overlapped significantly (e.g., NINE/NEUN or MAN/MANN). Declerck et al. (2012) found that digits and cognates were named more easily and caused smaller switch costs than noncognates and semantically related pictures, suggesting these differences can be attributed to phonology.

Similar results have since been reported by Li and Gollan (2018), though the cognate facilitation effect on switch costs appears to be limited to the nonmixed presentation of cognates (vs mixed with, for example, noncognates). In a detailed analysis, they revealed that cognate facilitation was dependent on presentation order: Cognates produced smaller switch costs when a picture was presented for the first time; however, such facilitation disappeared or even reversed in mixed blocks when a picture was repeated more often. Li and Gollan (2018) thus argued that phonological co-activation across languages initially facilitates selection of the correct word but that over time activation feeds back to the lexical level, where words compete for selection. Their findings highlight that activation at different levels during bilingual language production can interact in complex ways.

However, cognates are frequently spelled similarly, as phonological and orthographic overlap is often correlated in alphabetic languages due to the existence of grapheme-phoneme-correspondence (GPC) regularities, that is, (quasi-)regular mappings between phonology and orthography. As a result, we currently do not have a good understanding of the extent to which co-activation is caused by phonological and/or orthographic overlap in spoken language switching (but see Hoshino & Kroll, 2008, for evidence of cognate facilitation in Japanese-English bilinguals).

The role of phonological and/or orthographic overlap during written language switching

Although the beginning stages of written production are thought to be shared with spoken production (formation of abstract concept; addition of syntactic information; Perret & Laganaro, 2013), there is less agreement about what happens after. The phonological mediation hypothesis (e.g., Geschwind, 1974; Luria, 1970) states that writing always also involves some type of inner speech (likely because writing is acquired after speaking); that is, when writing, people are often thought to first activate the phonological code which subsequently can be converted into orthographic information. Evidence for the phonological mediation hypothesis comes from findings that naming an object by writing often takes longer than by speaking, as predicted if phonological information needed to be accessed first before orthographic information could be activated (e.g., Bonin et al., 2002; but also see Perret & Laganaro, 2013). In contrast, the orthographic autonomy hypothesis states that phonological and orthographic information can be accessed independently (Rapp et al., 1997). Here, activating a word’s phonology is not required to retrieve its orthography (even though optional phonological mediation is still possible). Evidence for orthographic autonomy comes from a patient that can write a word without accessing its spoken form (Rapp et al., 1997). As described in the previous section, there is considerable evidence for phonological co-activation of translation-equivalent words across languages during bilinguals’ spoken language production. The phonological mediation hypothesis suggests that similar phonological activation may occur during written language production, while the orthographic autonomy hypothesis indicates that orthographic similarity of translation-equivalent words may be more critical in bilinguals’ writing.

Evidence for orthographic co-activation across languages during bilinguals’ written language production comes from Muscalu and Smiley (2019): Here, Romanian-English bilinguals translated cognates and noncognates from their L2 to L1 by typing them. Even as cognates were produced more quickly, participants made more cross-language orthographic interference errors for them: although overlap facilitated language selection at the lexical level, it impeded selection when the full word was actually being written. Moreover, Iniesta et al. (2021) found that Spanish-English bilinguals spelled translation-equivalent words more correctly when the words were more similar orthographically.

Research into orthographic overlap on language switching also comes from the lexical decision task. For example, Thomas and Allport (2000) investigated the effect of language-specific orthography (operationalised here as language-specific letter clusters) on language switching in language comprehension where English-French bilinguals had to judge if a stimulus was an existing word or not in their L1 or L2. They found no effect of orthography on switch cost, concluding that switch costs are independent of the influence of orthographic features. However, Orfanidou and Sumner (2005) also examined the impact of orthographic specificity on language switching in a lexical decision task in Greek–English bilinguals. Notably, English and Greek share some letters but not others; this then allowed for the manipulation of whether stimuli contained letters unique to just one language or not. In this design, switch costs were significantly reduced when stimuli included language-specific orthography, suggesting that orthographic overlap can impact lexical decision-making.

There is currently very little published data on switching languages when producing written words; this is true for both handwriting and typing. One recent notable exception that investigated written language switching is a study by Wong and Maurer (2021), where both input (voice or picture) and output (speaking or handwriting) were manipulated. Chinese–English bilinguals were found to be slower when producing words by handwriting than speaking. In addition, switch costs were observed when producing words via handwriting (independently of the input modality) similarly to when producing words via speaking. However, no differences in proportional switch costs between the two output modalities were found, suggesting that similar language control mechanisms may be at play.

Moreover, a study by Schaeffner et al. (2017) investigated how language switching differed when switching between two languages that have a spoken form (German and English; Experiment 2) or switching between a language that has a spoken form and one that has not (English or German and a sign language; Experiment 1). Most relevant to the current study, Schaeffner et al. (2017) found that switch costs for English/German ¹ did not differ in the unimodal (both languages were spoken) and bimodal (one language was spoken, one typed) conditions. Thus, we can conclude that typing a word leads to similar levels of competition than if the word had been spoken, consistent with the phonological mediation hypothesis. Alternatively, the results by Schaeffner et al. (2017) could also indicate that the same level of orthographic activation occurs during typing as during speaking. These two (not mutually exclusive) types of competition have not been fully disentangled in language switching.

The current study

The goal of the study was to investigate how phonological and orthographic overlap influences written language switching. To test this, we investigated cued written language switching in translation-equivalent word pairs that differed in phonological and orthographic overlap. Participants were always German–English bilinguals that were either recruited locally or via a web-based platform. Different word types (e.g., noncognates and cognates) were always presented in separate blocks (though see Experiment 4 for an exception) to investigate the impact of each word type without influences of another. Participants had to name pictures by typing their responses.

Overall, we anticipated to see the general patterns typically observed in a cued spoken language switching experiment with unbalanced bilinguals (Wong & Maurer, 2021). We thus predicted switch costs and potentially L1 slowing, though we will not be able to assess L1 slowing directly, as participants always had to type picture names in both languages within the same block. Based on a recent meta-analysis, we did not predict asymmetrical switch costs (Gade et al., 2021). In addition, we predicted that naming performance and language switching would become easier, the more translation-equivalent word pairs overlapped in phonological and orthographic representations. More precisely, we hypothesised that performance and switch costs for noncognates would be worse than for cognates (Experiment 1), consistent with the results by Declerck et al. (2012) in spoken language switching. Such results would be indicative of increased phonological and/or orthographic co-activation across languages for cognates, thus facilitating switching between similar words even when writing.

In Experiments 2–4, we created two high-overlap cognate sets of German–English translation-equivalent word pairs that either were pronounced very similarly but spelled differently (so-called homophones, for example, SHOE [English] and SCHUH [German]) or that were spelled the same but pronounced differently (so-called homographs, for example, BALL is pronounced /bɔːl/ in English and /bal/ in German). We contrasted these high-overlap cognates with another set of low-overlap cognates (e.g., BOOK/BUCH) that had also been used in Experiment 1. This approach allowed us to maximise the impact of phonological (Experiment 2) and orthographic overlap (Experiments 2–4) on written language switching (see Table 1 for an overview of the experiments). We predicted that if written language switching was highly sensitive to phonological overlap, it should be easier to switch with homophones than low-overlap cognates. In contrast, if language switching benefits were driven by orthographic overlap, we should observe stronger facilitation effects for the homographs in comparison to the low-overlap cognates or homophones.

Table 1.

Overview of word types used across experiments. X indicates that a word type was included in the experiment. For each word type, one example is given in brackets.

Experiment	Word type (English/German)
	Noncognates (DRESS/KLEID)	(Low-overlap) cognates (STONE/STEIN)	Homophones (HOUSE/HAUS)	Homographs (TIGER/TIGER)
1	X	X
2	X	X	X
3		X	X	X
4	X	X		X

Experiment 1

Method

Participants

Participants were either recruited via the participant recruitment platform Prolific (www.prolific.co; N = 20) or students from RWTH Aachen University (N = 17). All were tested online (i.e., on the internet). Participants recruited on Prolific were paid approximately 8€ per hour. Locally recruited students received course credit for participation. Participants gave informed consent according to an internal ethics procedure consistent with the World Medical Association Declaration of Helsinki. All participants reported at least a B1 level of English proficiency ² and were native speakers of German. Proficiency in German and English was measured with the Lexical Test for Advanced Learners of English (LexTALE; Lemhöfer & Broersma, 2012). Both versions were implemented in the Gorilla Experiment Builder (Anwyl-Irvine et al., 2020), but their design was identical to the web-based LexTALE test with the exception that participants were given up to 5 s to respond (instead of unlimited time). All participants also completed a typing speed test ³ in which they typed three sentences in both German and English as quickly and correctly as possible. Sentences included on average 10 words, and participants were given up to 20 s to respond. Sentences were scored by subtracting misspelled words from the sentence length. We then calculated a score for number of correct words typed per minute that was averaged across sentences. We did not collect information on what type of keyboard participants used. English and German keyboard differ in the location of the letters Z and Y (these letters take the place of each other). There were two word pairs which could have been affected by this difference (CITY/STADT and FLASH/BLITZ), and these were only used in Experiment 1.

Three participants were excluded from analysis because their German LexTALE score was below 70%, indicating low German (L1) proficiency (N = 2) or because their picture naming accuracy in the experimental task was two standard deviations below the mean (N = 1), leaving 34 participants for analysis. For the remaining participants, the average score for their L1 was 88% (SD = 5%; range = 75%–98%) and for their L2 was 78% (SD = 11%; range = 58%–98%). Participants were able to type 47 words per minute (SD = 14 words) in German and 47 words per minute (SD = 13 words) in English. More detailed information on participants’ background can be found in the Supplementary Material A1 for all experiments. An a priori power analysis that was conducted with G*Power (Faul et al., 2007) predicted a required sample of N = 34 to detect effects of the size of d = 0.5 at the power of 0.8.

Stimuli

Eight unique noun item pairs were selected per word type (noncognate, cognate; see Table 2). A small number of stimuli were used to minimise spelling errors. For the noncognates, words never overlapped at onset, and they were separated by an orthographic Levenshtein distance (LD) ⁴ of at least 4 (M = 4.5 ± 0.5). For cognates, translation-equivalent words always started with the same first letter and showed a high degree of phonological and orthographic overlap. The latter is reflected in the overall smaller LD (M = 2.3 ± 0.9; range = 1–3) than for noncognates. Words within a word type and across word types were matched in frequency and length where length was estimated by the numbers of letters. Frequency was extracted from the SUBTLEX databases for German (Brysbaert et al., 2011) and British English (van Heuven et al., 2014), respectively. For English, Zipf log values were available as part of the reported values; for German, these were calculated based on the formula given by van Heuven et al. (2014). Frequency for all used stimuli was high with Zipf log values over 4 (range = 4–6). Frequency and length for all stimuli are given as part of the Supplementary Material A1. Pictures were black-and-white and selected from different databases (Bates et al., 2003; Severens et al., 2005; Snodgrass & Vanderwart, 1980). ⁵ Naming language was cued with either a German or British flag at the centre of the screen.

Table 2.

Overview of stimuli used (separated by word type). Stimuli in the noncognate and cognate condition that were used both in Experiments 1, 2, and 4 are marked with an asterisk. Please note that higher values for the orthographic and phonological Levenshtein distance indicate a lower similarity, whereas higher values for the median phonological similarity rating indicate higher similarity for translation-equivalent word pairs.

Word type	English (L2)	German (L1)	IPA English (L2)	IPA German (L1)	Orthographic Levenshtein distance	Phonological Levenshtein distance	Median phonological similarity rating
Noncognate	Chair*	Stuhl*	tʃεː	ʃtuːl	5	4	1
	City	Stadt	ˈsɪti	ʃtat	5	4	1
	Dress*	Kleid*	drεs	klait	4	5	1
	Flash	Blitz	flaʃ	blɪts	4	4	1
	Horse*	Pferd*	hɔːs	pfeːrt	5	5	1
	Roof*	Dach*	ruːf/rʊf	dax	4	3.5	1
	Smoke*	Rauch*	sməʊk	raux	5	5	1
	Tree*	Baum*	triː	baum	4	4	1
	M ± SD		–	–	4.5 ± 0.53	4.31 ± 0.59	1 ± 0
(Low-overlap) cognate	Book*	Buch*	bʊk	buːx	3	3	6
	Fire*	Feuer*	ˈfʌɪə	ˈfɔyər	3	3	6
	Fish	Fisch	fɪʃ	fɪʃ	1	0	10
	Ghost*	Geist*	ɡəʊst	ɡaist	2	2	5
	House	Haus	haʊs	haus	2	1	9
	Light*	Licht*	lʌɪt	lɪçt	1	2	6
	Stone*	Stein*	stəʊn	ʃtain	3	3	5
	Sun*	Sonne*	sʌn	ˈzɔnə	3	4	5
	M ± SD		–	–	2.25 ± 0.89	2.25 ± 1.28	6.5 ± 1.93
Homophone	Beer	Bier	bɪə	biːr	1	3	10
	Fish	Fisch	fɪʃ	fɪʃ	1	0	10
	Glass	Glas	ɡlɑːs/ɡlas	glaːs	1	2	8
	House	Haus	haʊs	haus	2	1	9
	Ice	Eis	ʌɪs	ˈeːɪs	3	3	10
	Shoe	Schuh	ʃuː	ʃuː	3	0	9
	M ± SD		–	–	1.77 ± 0.92	1.57 ± 1.20	9.33 ± 0.82
Homograph	Ball	Ball	bɔːl	bal	0	2	8
	Bus	Bus	bʌs	bʊs	0	1	8
	Hand	Hand	hand	hant	0	1	7
	Pilot	Pilot	ˈpʌɪlət	piˈloːt	0	5	8
	Rose	Rose	rəʊz	ˈroːzə	0	4	7
	Tiger	Tiger	ˈtʌɪɡə	ˈtiːgər	0	4	8
	M ± SD		–	–	0 ± 0	2.83 ± 1.72	7.67 ± 0.52

SD: standard deviation.

Note. The word pairs HOUSE/HAUS and FISH/FISCH were used as cognates in Experiment 1 and homophones in Experiment 2; this is why they are listed both in the cognate and homophone sections. Phonological Levenshtein distance was calculated by first transcribing each word phonologically and then calculating the Levenshtein distance between them. For English words, standard British realisations were assumed. IPA pronunciations for English were retrieved from Oxford English Dictionary (Oxford English Dictionary, n.d.); IPA pronunciations for German were retrieved from Langenscheidt (n.d.). For the two English words where there were two phonological realisations, phonological Levenshtein distance was calculated for both and the average of them is reported here. In addition, 13 unbalanced German–English bilinguals with a similar language background as the tested participants that were not familiar with the hypotheses of this research rated the similarity of each translation-equivalent word pair on a scale of 1 (do not sound similar at all) to 10 (sound identical). These values are reported in the last column of Table 2.

Procedure

After completing the consent and data protection forms, participants completed the language background questionnaire, the German LexTALE, the English LexTALE and the typing speed test (always in this order). Instructions were always in German with the exception of the English LexTALE, for which the instructions were in English. Subsequently, participants were asked to name all the pictures in both German and English. If they gave an incorrect response, participants were provided with the correct word for the picture. Participants were then given a short, written explanation of the task in German. During this, both speed and accuracy in their responses were emphasised. Participants were instructed to always type in lower case, even though German nouns are typically capitalised.

Participants completed a cued language switching experiment, where they named the presented pictures. In the beginning of each trial, participants saw a centred fixation point (+) for 400 ms. This was followed by the presentation of the centred language cue for 1,000 ms. The to-be-named picture was then presented in the centre without the language cue until participants completed their typed response by pressing the ENTER button. Participants could see their typed responses below the to-be-named picture. Reaction time (RT) was recorded for each letter press separately. The next trial started immediately after the ENTER button was pressed. Participants first completed a practice block of 16 trials that included both noncognates and cognates. Subsequently, they completed four experimental blocks of 80 trials each (N_total = 320). In each block, only one word type was presented. Participants either first completed two blocks with noncognates or two blocks with cognates; the order was counterbalanced across participants. Each of the eight item pairs was presented 10 times (five times for each language) per block. The order of items was randomised within a set of eight where each item pair was presented once (half in German, half in English) to minimise direct repetition of the same response. The randomisation procedure resulted in slightly more switch than repetition trials (switch trials = 56%). This was also true when considering number of switch/repetition trials separately for each language and word type.

The experiment lasted approximately 35 min. Trials were presented and recorded using the Gorilla Experiment Builder (www.gorilla.sc) (Anwyl-Irvine et al., 2020). The exact size of pictures could not be controlled, as participants completed the experiment on their own home computer or laptop. It was not possible to access the experiment via a phone or tablet.

Design

The goal of Experiment 1 was to examine the impact of phonological and orthographic overlap on participants’ written language switching. As independent variables, we manipulated word type (noncognate, cognate), language (L1 [German], L2 [English]), and trial transition (repetition, switch) within subjects. The dependent variables were always RT and error rates. RT was recorded for each key stroke, but only the first letter press was analysed (similarly to voice onset in the spoken domain; cf. Schaeffner et al., 2017). ⁶ Any spelling mistake (including the use of the “delete,” “backspace” or any other special key) counted as an error.

Results

Analyses were always implemented in R (Version 4.0.2; R Core Team, 2020). The first trial of each experimental block, all error trials, all trials following an error trial, and trials with RTs below 100 ms were excluded from RT analyses. Moreover, we excluded the rare case in which the randomisation procedure had resulted in a trial where the same response had to be given as the preceding one or when the same picture was shown twice in a row. Finally, RTs of all trials were z-transformed for each subject, and trials with a z-score of −3/+3 were excluded as outliers. For error rate analyses, the same criteria were used with the exception that error trials were now included. Overall, 28% of the RT data and 17% of the error rate data were excluded. The high exclusion rate was the result of a high number of typing errors (see the Supplementary Material A2 for means). Data were analysed with a 2 × 2 × 2 within-subject, repeated measures analysis of variance (ANOVA) (language × trial transition × word type). To test the robustness of our results, we also conducted post hoc all main analyses with a less strict error definition for all experiments, where trials were only excluded as errors if they did not start with the correct letter and were closer in LD to the other language’s spelling. The results are reported as part of the Supplementary Material B1. The results of the supplementary analyses were very similar to the ones reported here.

RT

RT data from Experiment 1 are depicted in Figure 1. There was a significant effect of language, F(1, 33) = 62.48, p < .001, generalised η² = 0.0428, indicating that participants’ naming was slower in their L1 than in their L2 (M_L1 = 1,116 ms; M_L2 = 1,062 ms). Moreover, we found a significant effect of trial transition with responses being faster in repetition than switch trials (M_repetition = 1,069 ms; M_switch = 1,104 ms), F(1, 33) = 38.95, p < .001, generalised η² = 0.0190, as well as a main effect of word type with shorter RTs for the cognates than the noncognates (M_noncognate = 1,174 ms; M_cognate = 1,004 ms), F(1, 33) = 309.50, p < .001, generalised η² = 0.2919. Importantly, there were also significant interactions between language and word type, F(1, 33) = 52.09, p < .001, generalised η² = 0.0205, as well as trial transition and word type,F(1, 33) = 5.55, p = .025, generalised η² = 0.0027, respectively. No other interaction reached significance (all Fs < 1).

Figure 1.

RT (in ms) across word types for languages (L1, L2) and trial transition (switch, repetition) in Experiment 1.

Error bars indicate ± one standard error.

To investigate the significant interactions, we split the RT data by word type. For all post hoc tests, the Bonferroni correction was used to adjust the level of α to evaluate significance. For noncognates, we found a significant effect of language, F(1, 33) = 71.60, p < .001, generalised η² = 0.1093, and trial transition, F(1, 33) = 25.66, p < .001, generalised η² = 0.0340, in the expected directions. For cognates, there were also significant effects of language, F(1, 33) = 9.61, p = .004, generalised η² = 0.0046, and trial transition, F(1, 33) = 14.08, p < .001, generalised η² = 0.0078. Switch costs amounted to 50 ms for noncognates and to 21 ms for cognates. For both word types, RTs were longer for L1 than L2, but the effect was numerically smaller for cognates (18 ms) than for noncognates (94 ms).

Error rate

For error rate, the only effects that reached significance were the ones of language (M_L1 = 0.1452; M_L2 = 0.1150), F(1, 33) = 5.18, p = .030, generalised η² = 0.0265, and trial transition (M_repetition = 0.108; M_switch = 0.148), F(1, 33) = 19.65, p < .001, generalised η² = 0.0343. Participants made more errors when they had to name a picture in their L1 than in their L2 and when they had to switch the language than when they did not. In addition, the interaction of language and trial transition reached significance, F(1, 33) = 5.34, p = .027, generalised η² = 0.0065. Finally, there was a nonpredicted significant three-way interaction of language, trial transition and word type, F(1, 33) = 4.48, p = .041, generalised η² = 0.0047. All other main effects and interactions did not reach significance (Fs < 1.25). Follow-up ANOVAs on the split data by language found that the interaction of language and trial transition was driven by a significant effect of trial transition for L1 trials, F(1, 33) = 24.98, p < .001, generalised η² = 0.0530, but not for L2 ones (F = 3.67), suggesting overall smaller switch costs for L2 than for L1. The effect of word type was nonsignificant for both L1 (F < 0.01) and L2 (F = 0.25).

Discussion

Although we observed robust switch costs in both dependent variables, we did not find evidence for asymmetrical switch costs in L1 in the RT analyses (although the effect was found in error rates). ⁷ In addition, there were longer RTs and higher errors rate in L1 than L2; this pattern is compatible with L1 slowing, though this conclusion can only be drawn with confidence if a nonmixed language block were included for comparison. Importantly, consistent with previous studies in spoken language switching, cognates were typed more quickly than noncognates and it was easier to switch languages when only naming cognates (e.g., Declerck et al., 2012; Li & Gollan, 2018). Cognate cross-language co-activation facilitated selection of the correct word in our nonmixed design. Cognate facilitation effects were only evident in the RT data of the first key stroke but not error rates.

Our findings suggest that control processes during language switching may be similar when speaking and writing (Wong & Maurer, 2021). However, it is unclear to what extent the here observed cognate facilitation effect is driven by phonological co-activation and/or orthographic one. Thus, Experiment 2’s goal was to further investigate the impact of phonological overlap and co-activation during written language switching. For this purpose, we included an additional word type of higher-overlap cognates, which we termed homophones, in the experiment: Here, homophones are a subtype of cognates that are pronounced very similarly (sometimes identically) but are spelled differently. By including this new word type, we asked how difficult written picture naming and language switching is when overlap between translation-equivalent words is more isolated to phonology, but not orthography.

Experiment 2

Method

Participants

We again recruited German–English bilinguals with at least B1 fluency in English via the web-based platform Prolific (www.prolific.co; N = 40) as well as from a local student pool (N = 22). Participants recruited on Prolific were paid approximately 8€ per hour for their participation and locally recruited students were given course credit for participation. We recruited more participants for Experiment 2 than all other experiments with the goal to explore whether locally recruited and online recruited participants differed in their response patterns. However, as described in the section “Results,” a preliminary analysis did not reveal any differences between the participant groups or interactions with any of the factors of interest. Participants gave consent according to an internal ethics procedure. Proficiency in German (L1) and English (L2) was again measured with the LexTALE (Lemhöfer & Broersma, 2012). All participants also completed the typing speed test that was introduced in Experiment 1. Two participants were excluded because of low German language proficiency (<70% on German LexTALE); three participants were excluded because of low picture naming accuracy (two standard deviations below mean). This left 57 participants for analysis. For the remaining participants, the average score for German was 87% (SD = 6%; range = 71%–97%) and for English was 78% (SD = 14%; range = 48%–100%). They were able to type 49 words per minute (SD = 15 words) in German and 53 words per minute (SD = 16 words) in English.

Stimuli

Six unique item pairs were selected per word type (noncognate, low-overlap cognate, homophone; see Table 2). Translation-equivalent noncognates never overlapped at onset and were separated by an LD of at least 4 (M = 4.5 ± 0.6). Translation-equivalent low-overlap cognates always started with the same first letter and generally showed a high degree of orthographic overlap, resulting in the smaller LD (M = 2.5 ± 0.8). Translation-equivalent homophones sounded very similar or even identical but were spelled differently. The orthographic overlap between these words was nevertheless relatively high, as seen in their LD of (M = 1.8 ± 1.0). To quantify phonological similarity across translation-equivalent words, stimuli were first transcribed into International Phonological Alphabet (IPA) referring to British English and German transcriptions as published in common dictionaries (Langenscheidt, n.d.; Oxford English Dictionary, n.d.). Subsequently, LD was calculated for the phonological transcriptions. This revealed a phonological LD of (M = 4.4 ± 0.7, M = 2.8 ± 0.8), and (M = 1.5 ± 1.4) for noncognates, low-overlap cognates, and homophones, respectively. Phonological LD did not always capture psychological reality of what “sounded very similar.” For example, the translation-equivalent pair STONE/STEIN had the same phonological LD as the pair BEER/BIER. Thus, we also asked a number of unbalanced German–English bilinguals with a similar language background as the tested participants that were not familiar with this research to rate the similarity of each translation-equivalent word pair (N_raters = 13). The ratings are also reported as part of Table 2. As before, words within a set and across sets were matched in length. For Experiment 2, picture representations were coloured and selected from the database MultiPic (Duñabeitia et al., 2018). Naming language was again cued with the German or British flag at the centre of the screen.

Procedure

The trial procedure was identical to Experiment 1. As before, after naming all the pictures used in both languages and receiving feedback, participants completed 16 practice trials of the experimental task that contained all possible word types. The experiment was separated into six blocks of 60 trials (N_total = 360 trials). Two subsequent blocks included words from the same stimulus set (e.g., Blocks 1–2 always included only noncognates). The order of the blocks was counterbalanced across participants. Within a block, each picture was presented 10 times (five times for each language). The order of items was randomised as before, resulting in again a slightly higher number of switches (58%) than repetition trials. This was also true when considering distribution of trial transitions across for languages or word types separately. The experiment lasted approximately 40 min. As before, the experiment was hosted by Gorilla Experiment Builder (www.gorilla.sc).

Design

The goal of Experiment 2 was to investigate the impact of phonological overlap of translation-equivalent words on language switching. To do so, we replicated the comparison between the noncognates and low-overlap cognates (Contrast 1). In addition, we contrasted performance between the low-overlap cognates and the homophones (Contrast 2). Both contrasts were preplanned. The independent variables were word type (noncognate, low-overlap cognate, homophone), language (L1 [German], L2 [English]), and trial transition (repetition, switch). Independent variables were manipulated within-subject. The dependent variables were the RT of the first key stroke and error rate.

Results

The same exclusion criteria were used as for Experiment 1, resulting in the exclusion of 27% of the RT and 17% of the error data. A preliminary analysis revealed no significant differences between the two recruitment groups (Prolific, local students) and interactions with any of the variables of interest (all Fs < 1); therefore, we analysed the two groups together to simplify the design. RT data from all three word types are presented in Figure 2. For brevity, the results of the main 2 × 2 × 3 within-subject ANOVA (language × trial transition × word type) with all word types are reported in the Supplementary Material B2. Contrasts were tested with a 2 × 2 × 2 within-subject ANOVA, where only two of the three word types were included. A supplementary analysis with a less strict error definition is reported in the Supplementary Material B1.

Figure 2.

RT (in ms) of first key stroke across word types for languages (L1, L2) and trial transition (switch, repetition) in Experiment 2.

Error bars indicate ± one standard error.

Contrast 1

When comparing noncognates and cognates, we found a significant effect of language, F(1, 56) = 44.37, p < .001, generalised η² = 0.0273, with picture naming RTs being longer for L1 than L2 (M_L1 = 1,126 ms; M_L2 = 1,077 ms). Moreover, there was a significant effect of trial transition consistent with switch costs (M_repetition = 1,079 ms; M_switch = 1,118 ms), F(1, 56) = 87.96, p < .001, generalised η² = 0.0221, as well as a main effect of word type (M_noncognate = 1,182 ms; M_{low-overlap cognate} = 1,022 ms), F(1, 56) = 312.90, p < .001, generalised η² = 0.2541, in line with a cognate facilitation effect. The interaction between language and word type reached significance, too, F(1, 56) = 14.95, p = .003, generalised η² = 0.0085. Similarly, the interaction between trial transition and word type reached significance as well, F(1, 56) = 6.17, p = .016, generalised η² = 0.0014. None of the other interactions were close to significance (all Fs < 1).

To further investigate the significant interactions, we split the data by word type and ran a 2 × 2 within-subject, repeated measures ANOVA (language × trial transition). For noncognates, we found significantly slower performance in the L1 than in the L2, F(1, 56) = 35.17, p < .001, generalised η² = 0.0551, and in switch than in repetition trials, F(1, 56) = 52.77, p < .001, generalised η² = 0.0295. Similarly, for low-overlap cognates, there were significant effects of language, F(1, 56) = 11.65, p = .001, generalised η² = 0.0067, and trial transition, F(1, 56) = 42.66, p < .001, generalised η² = 0.0149, in the same directions as for the noncognates. However, as before, effect sizes generally indicated smaller differences for low-overlap cognates than for the noncognates: Switch costs amounted to 31 ms for low-overlap cognates and to 53 ms for noncognates. The difference between L1 and L2 performance was smaller for cognates (22 ms) than for noncognates (75 ms).

These analyses were repeated with error rate as the dependent measure. The main effects that reached significance were the one of language (M_L1 = 0.132; M_L2 = 0. 0.103), F(1, 56) = 15.32, p < .001, generalised η² = 0.0257, and the one of trial transition (M_repetition = 0.096; M_switch = 0.133), F(1, 56) = 25.00, p < .001, generalised η² = 0.0391, as participants made more errors when naming pictures in their L1 than their L2 and in switch than repetition trials (all other Fs < 2.2).

Contrast 2

When comparing low-overlap cognates and homophones, there was a significant effect of language with RTs again being longer for the L1 than the L2 (M_L1 = 1,037 ms; M_L2 = 1,018 ms), F(1, 56) = 14.44, p < .001, generalised η² = 0.0047. In addition, RTs for picture naming were significantly shorter in repetition than switch trials (M_repetition = 1,013 ms; M_switch = 1,038 ms), F(1, 56) = 42.94, p < .001, generalised η² = 0.0103. Importantly, there was not a significant effect of word type (p = .165), indicating similar picture naming performance for both word types. However, there was a significant interaction of word type and trial transition, F(1, 56) = 4.39, p = .041, generalised η² = 0.0007 (all other interactions’ Fs < 2). We investigated the one significant interaction by splitting the data by word type. Here, we found an effect of trial transition for low-overlap cognates, F(1, 56) = 42.66, p < .001, generalised η² = 0.0149, as well as for homophones, F(1, 56) = 14.28, p < .001, generalised η² = 0.0062, with the effect being more pronounced for low-overlap cognates (switch costs = 31 ms) than homophones (switch costs = 19 ms).

We repeated the analyses with error rate as the dependent measure. The main effects that reached significance were language (M_L1 = 0.132; M_L2 = 0.108), F(1, 56) = 7.93, p = .007, generalised η² = 0.0148, indicating higher error rates for participants’ L1 than their L2, and trial transition (M_repetition = 0. 110; M_switch = 0.127), F(1, 56) = 6.75, p = .012, generalised η² = 0.0092. In addition, the two-way interaction of language and word type, F(1, 56) = 5.75, p = .020, generalised η² = 0.0058, as well as the unpredicted three-way interaction of language, trial transition and word type was significant, F(1, 56) = 5.17, p = .027, generalised η² = 0.0035. All other Fs were below 3.52. To investigate the interactions, we again split the data by word type. We found that for low-overlap cognates, there was a main effect of language, F(1, 56) = 17.88, p < .001, generalised η² = 0.0456, and trial transition, F(1, 56) = 10.23, p = .002, generalised η² = 0.0267. In contrast, for homophones, neither of the main effects reached significance (Fs < 1)—but their two-way interaction did, F(1, 56) = 5.85, p = .019, generalised η² = 0.0124: Switch costs existed for L1 but not L2, where average error rate was even slightly higher for repetition (M = 0.129) than switch (M = 0.113) trials.

Discussion

Written picture naming was quicker for cognates than noncognates, even when the cognate set was limited to what we termed “low-overlap cognates.” In addition, switch costs were again lower in the cognate than the noncognate stimulus set. These results further support a cognate facilitation effect in the written domain (cf., Declerck et al., 2012; Li & Gollan, 2018). Moreover, picture naming performance was comparable for low-overlap cognates and homophones, as there were no differences in overall naming performance. However, there was a significant interaction between trial transition and word type due to smaller switch costs in the homophones than the low-overlap cognates. ⁸ This indicates that the additional phonological overlap of the homophones in comparison to the low-overlap cognates may have facilitated switching, possibly due to parallel activation. It is unclear then why parallel co-activation may have facilitated switching, but not naming.

In Experiment 3, we no longer included noncognates, but instead included homographs, where translation-equivalent words were pronounced differently but spelled identically. By doing so, we were able to examine more closely the impact of orthographic overlap on written language switching. In addition, Experiment 3 replicated the comparison of low-overlap cognates and homophones, thus clarifying the facilitation pattern of phonological overlap.

Experiment 3

Method

Participants

40 German–English bilinguals recruited via Prolific (www.prolific.co) were paid approximately 8€ per hour for their participation. They all identified as native German speakers with a self-reported B1 level of English proficiency and gave consent according to an internal ethics procedure. One participant was excluded because of low naming accuracy. For the remaining participants, the average LexTALE score for German was 87% (SD = 7%; range = 73–98%) correct and for English was 81% (SD = 11%; range = 57–96%) correct. Their typing speed was 54 words per minute (SD = 16 words) for German and 60 words per minute (SD = 17 words) for English.

Stimuli

The stimuli used for the low-overlap cognate and homophone set were identical to the ones used in Experiment 2. For homographs, translation-equivalent words were selected that were spelled identically but not pronounced the same in German and in English (Table 2). As a result, the set’s LD was 0 (SD = 0).

Procedure

The procedure of Experiment 3 was identical to the one of Experiment 2. There again was a slightly higher percentage of switch trials (58%); this was also true when looking at the trial transition distribution per word type and language.

Design

The goal of Experiment 3 was to further examine the impact of phonological and orthographic overlap on language switching when writing. We compared low-overlap cognates and homophones (Contrast 1), low-overlap cognates and homographs (Contrast 2) and performance for homophones and homographs (Contrast 3). The independent and dependent variables were identical to Experiment 2 with the exception that the variable word type now included low-overlap cognates, homophones, and homographs.

Results

The same exclusion criteria were used as for the study’s previously reported experiments, resulting in 26% of trials being discarded for RT analyses and 17% of trials being discarded for error rate analyses. RT data from all three word types are presented in Figure 3 (see the Supplementary Material A2 for means). The results of the main 2 × 2 × 3 within-subject ANOVA can be found in the Supplementary Material B2. The different contrasts were investigated with a 2 × 2 × 2 within-subject ANOVA with only two of the three possible word types.

Figure 3.

RT (in ms) of first key stroke across word types for languages (L1, L2) and trial transition (switch, repetition) in Experiment 3.

Error bars indicate ± one standard error.

Contrast 1

When comparing low-overlap cognates and homophones, there was again a significant effect of language with slower performance for L1 than L2 (M_L1 = 990 ms; M_L2 = 974 ms), F(1, 38) = 8.72, p = .005, generalised η² = 0.0049, and a significant effect of trial transition indicating switch costs (M_repetition = 967 ms; M_switch = 993 ms), F(1, 38) = 30.32, p < .001, generalised η² = 0.0111. In contrast to Experiment 2, there was a significant main effect of word type, F(1, 38) = 6.50, p = .015, generalised η² = 0.0101: Picture naming was quicker for low-overlap cognates (M_{low-overlap cognate} = 970 ms) than for homophones (M_homophone = 995 ms). None of the interactions reached significance, suggesting language switching was performed similarly across word types (all Fs < 2.3).

When entering error rates as the dependent measure, the only effects that reached significance were the ones of language, (M_L1 = 0.154; M_L2 = 0.123), F(1, 38) = 19.92, p < .001, generalised η² = 0.0303, and trial transition (M_repetition = 0.112; M_switch = 0.158), F(1, 38) = 41.09, p < .001, generalised η² = 0.0585, in the expected directions. None of the other main effects or interactions were close to reaching significance (all other Fs < 1).

Contrast 2

When comparing low-overlap cognates and homographs, performance was slower for L1 than L2 (M_L1 = 931 ms; M_L2 = 925 ms), F(1, 38) = 4.11, p = .0496, generalised η² = 0.0012, as well as for switch than repetition trials, (M_repetition = 919 ms; M_switch = 935 ms), F(1, 38) = 29.04, p < .001, generalised η² = 0.0059. Moreover, the main effect of word type reached significance with slower responses for low-overlap cognates (970 ms) than for homographs (894 ms), F(1, 38) = 44.06, p < .001, generalised η² = 0.1042, in line with a strong homograph facilitation effect. In addition, the interactions between word type with language, F(1, 38) = 10.00, p = .003, generalised η² = 0.0032, and trial transition, F(1, 38) = 9.10, p = .005, generalised η² = 0.0035, were significant as well. None of the other interactions reached significance (all Fs < 1).

To examine the two significant interactions, the data were split by word type. For the low-overlap cognates, there was a significant effect of language, F(1, 38) = 9.41, p = .004, generalised η² = 0.0080, and trial transition, F(1, 38) = 27.66, p = .0045, generalised η² = 0.0178, with slower performance for L1 than L2 and switch than repetition trials. In contrast, for the homographs, none of the main effects or interactions reached significance (all Fs < 1.5), suggesting very similar picture naming performances across all conditions.

For error rate, the main effects of language, (M_L1 = 0.103; M_L2 = 0.083), F(1, 38) = 12.39, p = .001, generalised η² = 0.0222, trial transition (M_repetition = 0.082; M_switch = 0.101), F(1, 38) = 8.25, p = .007, generalised η² = 0.0170, and word type (M_{low-overlap cognate} = 0.134; M_homograph = 0.056), F(1, 38) = 48.15, p < .001, generalised η² = 0.2026, reached significance: Participants committed more errors in their L1 than their L2, in switch than repetition trials as well as for low-overlap cognates than homographs. Moreover, the interaction between language and word type, F(1, 38) = 5.54, p = .024, generalised η² = 0.0125, in addition to the interaction of trial transition and word type, F(1, 38) = 22.79, p < .001, generalised η² = 0.0427, were significant. None of the other interactions reached significance (Fs < 2.9). To investigate the source of the significant interactions, we split the error data by word type. For low-overlap cognates, significant main effects of language, F(1, 38) = 12.79, p = .001, generalised η² = 0.0429, and trial transition existed, F(1, 38) = 23.07, p < .001, generalised η² = 0.0707, while this was not true for the homographs (all Fs < 2.2).

Contrast 3

When comparing homophones and homographs, we found a significant effect of trial transition (M_repetition = 934 ms; M_switch = 943 ms), F(1, 38) = 8.78, p = .005, generalised η² = 0.0025, and word type (M_homophone = 995 ms; M_homograph = 894 ms), F(1, 38) = 59.64, p < .001, generalised η² = 0.1577, as well as a significant interaction of language and word type, F(1, 38) = 4.56, p = .039, generalised η² = 0.0014. There was no significant interaction of trial transition and word type (F = 3.62; all other Fs < 2). For homophones only, there was no significant effect of language (F = 3.63). Moreover, as reported for Contrast 2, none of the main effects or interactions reached significance for homographs. Closer inspection of the data suggested that the significant interaction of language and word type was driven by a performance pattern consistent with an effect of language in the homophones (even though it was not significant; M_L1 = 1,001 ms; M_L2 = 989 ms) but not homographs.

For error rates, we found a significant effect of language (M_L1 = 0.104; M_L2 = 0.090), F(1, 38) = 4.26, p = .046, generalised η² = 0.0112, trial transition (M_repetition = 0.0898; M_switch = 0.1025), F(1, 38) = 4.13, p = .049, generalised η² = 0.0092, and word type (M_homophone = 0.143; M_homograph = 0.056), F(1, 38) = 75.89, p < .001, generalised η² = 0.2513. Moreover, the interaction of trial transition and word type reached significance, F(1, 38) = 12.44, p = .001, generalised η² = 0.0306. Although there were switch costs (M = 0.039) for homophones, F(1, 38) = 11.67, p = .001, generalised η² = 0.0469, there were none for homographs.

Discussion

We again found differences between low-overlap cognates and homophones. However, in contrast to Experiment 2, we observed that picture naming was quicker for low-overlap cognates than homophones although switch costs did not differ across word types. When we repeated the analyses with the combined data of Experiments 2 and 3 (see Supplementary Material B3), we found that low-overlap cognates were named more quickly than homophones and that switch costs were smaller for homophones than low-overlap cognates. These results suggest that each experiment by itself was underpowered to detect the subtle differences between low-overlap cognates and homophones consistently. The quicker naming of low-overlap cognates than homophones suggests that participants’ performance benefitted from cross-language co-activation but that this led to worse switching. In contrast, blocks with homophones may have resulted in a relatively lower co-activation than low-overlap cognates (potentially to reduce interference from phonology) which then helped switching between languages. Here, low-overlap cognates may represent some kind of sweet spot for co-activation: They are similar enough across languages to help typing them quickly but not quite so similar to lead to confusions.

Picture naming of homographs was generally easier (i.e., quicker, less errors) than naming low-overlap cognates or homophones and not affected by language switching. This suggests that written performance benefits greatly from orthographic co-activation of translation-equivalent words (or, similarly, is not affected by a phonological mismatch). One critical caveat is that participants could have solved picture naming in the homograph blocks by adopting a strategy where they ignored the language cue on each trial and only paid attention to the actual picture. This strategy was possible, as correct responses were identical for each picture and language combination. At best, these data provide first tentative evidence for the importance of orthographic co-activation and subsequent facilitation during written language switching. However, as we cannot exclude the possibility that language cues were simply ignored, we conducted a fourth follow-up experiment where different word types were intermixed. If typing and switching between homographs is still facilitated, we can conclude that orthographic overlap highly impacts written language switching.

Moreover, we also investigated in Experiment 4 if language switching between homographs was impacted by the context in which they were presented in (similarly to what has been described with low-overlap cognates), where context was defined as the word types they were mixed with. We hypothesised that mixing homographs with noncognates may result in larger switch costs for homographs than when mixing with low-overlap cognates, as the latter block make-up allows for the more consistent co-activation of both languages which should benefit homographs.

Experiment 4

Method

Participants

Participants were 40 German–English bilinguals recruited via the local student pool and received course credit for participation. Thirty-nine identified as native German speakers and as fluent in English (at least self-reported B1 level); one participant was not a native speaker of German and was therefore excluded from all analyses. Participants’ average LexTALE score for German was 86% (SD = 6%; range = 75–98%) correct and for English was 71% (SD = 10%; range = 52–97%) correct. Their typing speed for German was 42 words per minute (SD = 12 words) and 42 words for English (SD = 12 words).

Stimuli

There were three word types: noncognates, low-overlap cognates and homographs. Noncognates and low-overlap cognates were those used in Experiment 2. Homographs were the same as in Experiment 3.

Procedure

The overall procedure of Experiment 4 was identical to the one of Experiment 3 with the exception that word types were always intermixed within the same block. Half of mixed blocks included noncognates and homographs; the other half included low-overlap cognates and homographs. The order of mixing conditions was counterbalanced across participants. There were overall 10 blocks of 48 trials, resulting in a total of 480 trials. Each of the 12 item pairs (six per word type) was presented four times (two times for each language) per block. The order of items was randomised within a set of 12 where each item pair was presented once (half in German, half in English) to minimise direct repetition of the same response. As before, this randomisation procedure resulted in slightly more switch than repetition trials (switch trials = 57%; comparable for different word types and languages).

Design

The goal of Experiment 4 was to examine the impact of orthographic overlap on written language switching in three preplanned contrasts: Contrast 1 compared performance between noncognates and homographs. Contrast 2 compared performance between low-overlap cognates and homographs. Finally, Contrast 3 compared homographs in the two types of mixed blocks. The dependent variables were again RT of first key stroke and error rate.

Results

The same exclusion criteria were used as for the previously reported experiments. This approach resulted in 21% of trials being discarded for RT analyses and 13% of trials discarded for error rate analyses. RT data from all three word types are presented in Figure 4 and averages for all sub-conditions can be found in the Supplementary Material A2. Contrasts were tested with a 2 × 2 × 2 within-subject, repeated measures ANOVA.

Figure 4.

RT (in ms) of first key stroke across word types for languages (L1, L2) and trial transition (switch, repetition) in Experiment 4. (a) Data from blocks mixing noncognates and homographs. (b) Data from blocks mixing cognates and homographs.

Error bars indicate ± one standard error.

Contrast 1

Comparing noncognates and homographs, there was a significant main effect of language, F(1, 38) = 49.53, p < .001, generalised η² = 0.0469, trial transition, F(1, 38) = 20.82, p < .001, generalised η² = 0.0073, and word type, F(1, 38) = 642.40, p < .001, generalised η² = 0.4320: RTs were longer for L1 (M_L1 = 1,097 ms) than L2 (M_L2 = 1,045 ms), switch (M_switch = 1,078 ms) than repetition (M_repetition = 1,062 ms) and noncognates (M_noncognate = 1,201 ms) than homographs (M_homograph = 959 ms). In addition, both the interactions of language and word type, F(1, 38) = 65.46, p < .001, generalised η² = 0.0620, as well as the one of trial transition and word type, F(1, 38) = 14.41, p = .001, generalised η² = 0.0032, reached significance (all other interactions F < 1.8). To follow-up the significant interactions, data were split by word type. For noncognates, the effects of language, F(1, 38) = 60.81, p < .001, generalised η² = 0.1488, and trial transition, F(1, 38) = 26.28, p < .001, generalised η² = 0.0153, were significant; for homographs, only the effect of language reached significance, F(1, 38) = 5.33, p = .027, generalised η² = 0.0018, while trial transition was not close to being significant (F = 2.12), indicating no switch costs for homographs.

For error rate, only the main effect of word type reached significance, F(1, 38) = 52.75, p < .001, generalised η² = 0.1737, indicating more errors for noncognates (M_noncognate = 0.139) than homographs (M_homograph = 0.059). In addition, the interactions of language and trial transition, F(1, 38) = 6.13, p = .018, generalised η² = 0.0075, as well as the interaction of trial transition and word type were significant, F(1, 38) = 8.12, p = .007, generalised η² = 0.0097; all other Fs < 3.1. To investigate the interaction of language and trial transition, we split the data by language. For L1 and L2 trials, there was no significant effect of trial transition after adjusting α to evaluate significance using the Bonferroni correction (all Fs < 4.4). To investigate the interaction of trial transition with word type, we split the data by word type. For noncognates as well as homographs, none of the main effects or interactions reached significance after adjusting α to evaluate significance using the Bonferroni correction (all Fs < 4.6).

Contrast 2

When comparing low-overlap cognates and homographs, only the main effects of trial transition, F(1, 38) = 8.56, p = .006, generalised η² = 0.0025, and word type, F(1, 38) = 51.73, p < .001, generalised η² = 0.1087, reached significance, with shorter RTs in repetition (M_repetition = 1,001 ms) than switch (M_switch = 1,010 ms) trials and longer RTs in low-overlap cognates (M_{low-overlap cognate} = 1,055 ms) than homographs (M_homograph = 962 ms). In addition, the interactions of language and word type, F(1, 38) = 11.23, p = .002, generalised η² = 0.0027, and the three-way interaction of language, trial transition and word type, F(1, 38) = 4.78, p = .035, generalised η² = 0.0012, were significant (all other effects or interactions F < 2.4). To investigate the interactions, we again split the data by word type. For low-overlap cognates, the main effects of language, F(1, 38) = 7.06, p = .011, generalised η² = 0.0041, and trial type, F(1, 38) = 6.77, p = .013, generalised η² = 0.0047, were significant in the expected directions. The interaction of language and trial transition was not significant (F = 3.54). In contrast, for homographs, there were no significant effects at all; the effect of language was the one closest to significance; F(1, 38) = 3.00, p = .092, generalised η² = 0.0013 (all other Fs < 1.3).

For error rate, the main effect of language, F(1, 38) = 11.23, p = .002, generalised η² = 0.0257, word type, F(1, 38) = 39.82, p < .001, generalised η² = 0.1798, as well as their interaction, F(1, 38) = 4.98, p = .032, generalised η² = 0.0109, reached significance: Subjects made more errors in their L1 (M_L1 = 0.107) than their L2 (M_L2 = 0.081) and for low-overlap cognates (M_{low-overlap cognate} = 0.135) than for homographs (M_homograph = 0.053). Moreover, splitting the data by word type revealed that the interaction of the two was driven by a significant effect of language for low-overlap cognates, F(1, 38) = 9.19, p = .004, generalised η² = 0.0399, but not for homographs (F = 2.45): For low-overlap cognates, participants conducted more errors in their L1 than their L2.

Contrast 3

Finally, we compared homographs that were mixed with noncognates and ones that were mixed with low-overlap cognates. The only effect that reached significance was the one of language, F(1, 38) = 7.40, p = .010, generalised η² = 0.0015; all other Fs < 3: in contrast to previous analyses, RTs were shorter for L1 (M_L1 = 956 ms) than L2 (M_L2 = 965 ms). Whether homographs were mixed with noncognates or low-overlap cognates did not impact performance (F < 0.1). When the same analysis was repeated with error rate as the dependent variable, none of the main effects or interactions reached significance (all Fs < 3.5).

Discussion

Experiment 4 investigated whether homographs elicit faster responses and smaller switch costs in an intermixed design. We found that homographs were consistently easier to name (shorter RTs and less errors) than noncognates and low-overlap cognates with less orthographic overlap. Moreover, language switching was easier for homographs with no evidence for switch costs across analyses.

The absence of switch costs for homographs suggests that orthographic co-activation across languages is particularly important in written language switching. In addition, it could also indicate that access for homographs is easier because of their de facto increased frequency (both spoken and written), as occurrence in each language counts towards occurrence in both languages due to a potentially shared lexical representation (Higby et al., 2020; Titone et al., 2011). These effects—while they potentially could also play out with phonological overlap and speaking (Higby et al., 2020)—may be maximised here, as the letters on the keyboard are exactly the same, independently of language.

We also investigated whether mixing homographs with noncognates or low-overlap cognates impacted ease of naming homographs. We did not find evidence for such context modulation with very similar performance across mixing conditions. This suggests that participants did not change their switching strategy for homographs depending on what other word types they were presented with (cf., Li & Gollan, 2018). Furthermore, the similar patterns observed in Experiments 3 and 4 suggest that there are no large performance differences between when homographs are named mixed with other word types or not.

General discussion

We will first discuss general effects that were observed during written language switching before moving to a more detailed discussion of the impact of phonological and orthographic overlap on written language switching as well as the study’s limitations.

Written language switching

The language switching paradigm has been used widely to investigate language control processes of bilinguals (Declerck & Philipp, 2015a). Despite its popularity, there are surprisingly little data on written language switching. Wong and Maurer (2021) found similar switch costs for speaking and handwriting. They also found performance consistent with a reversed dominance effect, which was more pronounced when writing than speaking. Similarly, we were able to consistently observe switch costs and performance consistent with reversed dominance in our data. These results are consistent with unbalanced bilinguals inhibiting their more dominant language (L1) to prepare for switches. Moreover, responses may be slower right after a switch because the competing language that was just active has to be suppressed (e.g., Green, 1998). We did not find asymmetrical switch costs (i.e., higher L1 than L2 switch costs), which is consistent with previous studies that did not observe both asymmetrical switch costs and reversed dominance in a single data set (e.g., Christoffels et al., 2007; Mosca & Clahsen, 2016; Mosca & de Bot, 2017; Peeters & Dijkstra, 2018). Together, these results support the notion that language switching while writing is governed by similar control processes as language switching while speaking (see Wong and Maurer, 2021, for similar conclusions based on a direct comparison of speaking and writing).

The role of phonological and orthographic overlap in written language switching

Previous research on cognates shows that not only is it easier to name them quickly but it is also easier to switch between them in two languages, at least when they are presented in a nonmixed format (i.e., not mixed with noncognates; Declerck et al., 2012; Li & Gollan, 2018). Our data consistently revealed a cognate facilitation effect in that bilinguals were always quicker at typing cognates than noncognates where translation-equivalent words did not overlap phonologically and orthographically. In addition, switch costs were lower for cognates than for noncognates (even in Experiment 2, where overlap was more limited for cognates). Together, these data suggest that phonological/orthographic overlap in cognates facilitates not just spoken but also written language switching.

The role of phonological overlap in written language switching was investigated as part of Experiments 2 and 3. In Experiment 2, the comparison of the homophones (i.e., words that are pronounced the same but spelled differently) and the low-overlap cognates did not reveal any differences in the naming performance. However, switch costs were smaller for homophones than low-overlap cognates. However, in Experiment 3, participants typed low-overlap cognates more quickly than homophones, though this time there were no differences in switch costs between the two word types. When combining the data from both experiments (see Supplementary Material B3), both effects reached significance; nevertheless, this lack of robustness suggests that these differences in naming performance and switch costs are small. Thus, the maximisation of phonological overlap in the homophones offered a small advantage over low-overlap cognates when language switching. It is surprising that, at the same time, low-overlap cognates were typed more quickly than homophones, even with the higher phonological overlap in the homophones. Together, these results suggest a trade-off between facilitation and interference, where additional phonological overlap can hinder typing an object’s name. More specifically, a picture may activate a word’s spoken form even if a response is given by writing; due to the shared phonological representation of translation-equivalent homophones in bilinguals, activation then also flows to the orthographic representations of the word in both languages, leading to interference. At the same time, this existing co-activation may also make it easier to type the correct response in the previously inhibited language on a switch trial. Whether such activation of a word’s phonology while writing is optional or essential (as suggested by the phonological mediation hypothesis; Geschwind, 1969; Luria, 1970) is not clear from our homophone data.

The role of orthographic overlap in written language switching was investigated in Experiments 3 and 4. Here, we showed that pictures corresponding to homographs (i.e., words that are spelled the same but pronounced differently) in English and German are typed more quickly than noncognates, low-overlap cognates or homophones. In addition, participants experienced no significant switch costs when language switching between homographs. Although this data pattern could be explained by participants’ ignoring language cues in the nonmixed stimulus design of Experiment 3, no such explanation is possible for Experiment 4. Moreover, our results from Experiment 4 suggest that the observed homograph facilitation effect is not limited to nonmixed presentations as observed for cognates when speaking (Li & Gollan, 2018). Even with repeated presentations, homograph co-activation remains facilitatory in written language switching. In spoken language switching, naming cognates repeatedly in mixed presentation is thought to lead to feedback to the lexical level, where it can increase competition for selection (Li & Gollan, 2018). In written language switching, such feedback to the lexical level does not appear to lead to any interference for homographs. This suggests that activation does not flow to the phonological representations of the words, where there would be a mismatch. Overall, our homograph results provide support for the idea that a word’s orthographic form can be accessed independently from its phonological one, as put forward by the orthographic autonomy hypothesis (Rapp et al., 1997). It is also possible that the high orthographic overlap across language translation-equivalent word pairs—in combination with the ability to shield from the words’ phonological mismatch—allowed participants to engage in proactive language control even in mixed language contexts.

In addition, access for homographs may generally be easier because of their increased written frequency: As they “look” the same in each language, each occurrence may count towards the frequency in both languages (Titone et al., 2011). As a result, homographs may have a higher baseline activation than other cognates, and are therefore named more quickly than other word types. This explanation also applies to all cognates (and theoretically even noncognates; Higby et al., 2020) and when speaking, but may be more pronounced for homographs when writing, as produced letters across translation-equivalent pairs are identical independently of language. Thus, when studying language switching, orthographic overlap—independently of phonological one—should be considered (e.g., when investigating cognates more generally). At this point, it is unclear whether this pattern—small facilitation in language switching from phonological overlap but great facilitation in switching when orthographic overlap is maximised—extends to spoken language switching. One may speculate that the pattern is reversed when speaking, with phonological overlap being more important than orthographic overlap in determining the ease at which one can switch between languages.

Limitations

First, speed of typing as a measure of language production may introduce some unique challenges that do not exist in spoken language switching. For example, participants were instructed to always respond in lower case, even though German nouns are typically capitalised. This methodological choice was made, as capitalisation would have required pressing an extra key (SHIFT), which also happens to be relatively far from the letter keys, thus slowing down participants’ responses. In addition, typing in upper and lower case would require participants to both switch between languages (English/German) and typing manner (lower case only/mixed), introducing an additional confound. However, this methodological choice leaves open the possibility that German (L1) responses were slower than English ones because participants had to inhibit capitalising the first letter in German.

Second, a high number of trials needed to be excluded because of participants’ typing errors. This may not be surprising given that participants were spelling in their L1 and L2, and that this skill is acquired later in life than speaking. This aspect of the written language switching paradigm should be considered in future adaptions (e.g., by increasing the number of trials or changing what counts as an error), though supplementary analyses with a less strict error definition and thus less trial exclusions replicated our main findings (see Supplementary Material B1).

Third, the number of word pairs used was relatively small due to an attempt to reduce overall error rate and to closely match them in phonological and orthographic overlap as well as other characteristics (frequency, length, distribution of starting letters on keyboard). Our results may not be generalisable to other words and may be impacted by practice effects, as participants had to type the same words repeatedly. Likewise, our results cannot speak to language combinations other than English and German that differ in orthographic transparency or other potentially relevant variables.

Fourth, there is evidence that even if cognate facilitation can be observed, this does not exclude the possibility of inhibitory effects in other stages of production (e.g., Jacobs et al., 2016; Muscalu & Smiley, 2019). Thus, it is possible that if we had considered other dependent variables beyond RT of the first key press and error rate, increased phonological and orthographic overlap across translation-equivalent word pairs would have led to worse performance. Further research will be needed to clarify under which circumstances overlap may lead to inhibition in written language switching.

Finally, small differences in pronunciations exist (depending on the speaker and stimulus pair) in our homophone set. Unfortunately, such subtle differences in pronunciations are speaker/community-specific, are hard to quantify and can even exist for words that originated in the other language. These small mismatches in pronunciations for our homophone set might have reduced the parallel activation due to phonological overlap between translation-equivalent words and added to the small difference with the low-overlap cognates.

Conclusions

Together, our results suggest that the control mechanisms used during language switching while typing behave similarly as the ones observed when speaking. In addition, we found that orthographic overlap greatly facilitates picture naming and language switching when writing in line with the orthographic autonomy hypothesis. Moreover, our data indicate that the written language switching paradigm can be used to investigate language control phenomena; this may open some fruitful avenues for future research on language switching, for example, by combining the written language switching paradigm with online recruitment.