1. Introduction
When confronted with the question ‘How many animals of each kind did Moses take on the Ark?’ most people tend to answer ‘two’ and do not notice that it was Noah and not Moses who took the animals on the Ark. This phenomenon is known as the Moses-Illusion, a sort of semantic illusion in which readers often fail to notice the anomalous word (Noah) in a sentence that is perfectly aligned with world knowledge (Erickson & Mattson, Reference Erickson and Mattson1981). At first, the sentence seems to be semantically correct, but upon closer examination, a semantic anomaly can be detected. The representational nature and processing of semantic illusions have been extensively studied during first language (L1) comprehension (Gomes & Maia, Reference Gomes and Maia2024), as we will see later on. However, the question regarding the nature of semantic illusions that may arise during second language (L2) processing remains almost unexplored (Dhaene et al., Reference Dhaene, Dirix, Van Marcke and Woumans2022; Vaessen, Reference Vaessen2017). Additionally, whether the age of L2 acquisition and language dominance modulate the effect or not requires further investigation. The present research aimed to examine this issue in the two languages spoken by early and late European Portuguese (EP; L1)-German (GER; L2) bilinguals with different degrees of L2 proficiency, and thus varying language dominance (referring to the degree of balance between the two languages of the bilingual speaker). Two groups of native speakers of EP (EP speakers with no knowledge of GER) and of GER (GER speakers with no knowledge of EP) were also examined as control groups.
1.1. Semantic illusions in first language processing
In order to understand how semantic illusions emerge, it is crucial to discuss existing theories and influencing factors on the occurrence of semantic illusions in an L1.
In one of the first studies on the so-called ‘Moses-Illusion’, Erickson and Mattson (Reference Erickson and Mattson1981) contradicted general belief showing that participants do not process the critical name in a semantic illusion, neither at the phonemic nor at the graphemic levels, and that they respond to the experimental question without acknowledging the illusive name. In their first experiment, participants were asked to read Moses-Illusion-like sentences out loud (e.g., How many animals of each kind did Moses take on the Ark?) and to answer aloud questions as fast as possible to ensure that the incongruous word was encoded at least at the phonemic level. Since semantic illusions were observed, even though sentences were read out loud, the authors argued that the inconsistent name (e.g., ‘Moses’) was indeed encoded during sentence processing. Hence, the authors proposed that participants’ failure to recognize inaccuracies during the processing of Moses-like illusions stems from the focus of the questions (e.g., ‘How many animals’ vs. ‘Who took the animals’). Therefore, the second experiment was conducted to test this hypothesis. They shifted the focus of the questions away from the inconsistent name to examine whether the illusion only arose in a particular form of questions. Target questions used in the first experiment were turned into statements, such as ‘Moses took two animals of each kind on the Ark’. Participants had to read statements and circle ‘true’, ‘false’ or ‘I don’t know’ as response in a booklet. Although focalization attenuated the percentage of semantic illusions, illusions were still observed.
Additionally, Erickson and Mattson (Reference Erickson and Mattson1981) stated that the Moses-Illusion seems to work in specific combinations only, since the anomaly is easily identified by participants whenever the incongruous word is not related to the correct word (e.g., ‘How many animals of each kind did George Washington take on the Ark?’). Frequently, inconsistent names used in an illusion context are based on either phonological or semantic similarity with the correct word, resulting in two possible hypotheses. Thinking about the Moses-Illusion, Moses and Noah contain two syllables with the first syllable being stressed in both words and having an ‘o’ sound at the beginning of the word. In this case, the illusion could be caused by phonological similarity. On the other hand, Moses and Noah share semantic features; both are biblical figures who received messages from God and are linked to the sea. Therefore, the Moses-Illusion could also be due to semantic similarities. The procedure of the third experiment was similar to the second one – different variations of the questions were presented by inserting other names. In the Moses-Illusion, for example, the phonological similarity hypothesis was tested by replacing the illusive name ‘Moses’ with ‘Adam’ or ‘Abraham’. The three names were chosen because they vary in the degree of phonological similarity and because they are similar at the semantic level, representing biblical names that appear in the Old Testament. Likewise, to test the semantic similarity hypothesis, ‘Noah’ was replaced by a nonbiblical figure, namely ‘Nixon’. Results indicated that a semantic illusion is more likely to occur if inconsistent and correct target words have a semantic resemblance. No significant differences between the three levels of phonological similarity (for example, Moses, Adam and Abraham) were found. The authors concluded that the Moses-Illusion seems to occur when a bunch of semantic features are connected to the content of the sentence, confirming the semantic similarity hypothesis. Thus, when a participant reads something about an Ark, a semantic network connected to the concept of the Ark will be activated, including biblical figures, such as Moses, Adam or Abraham.
The seminal study by Erickson and Mattson (Reference Erickson and Mattson1981) led many studies to (re)examine the semantic similarity hypothesis (e.g., Reder & Kusbit, Reference Reder and Kusbit1991; Van Jaarsveld et al., Reference van Jaarsveld, Dijkstra and Hermans1997; Van Oostendorp & de Mul, Reference Van Oostendorp and de Mul1990; Van Oostendorp & Kok, Reference Van Oostendorp and Kok1990), the effect of the position of the incorrect name in occurrences of semantic illusions (e.g., Bredart & Modolo, Reference Bredart and Modolo1988; Van Jaarsveld et al., Reference van Jaarsveld, Dijkstra and Hermans1997), and the relation between syntactic structure complexity and the illusions (Gomes & Maia, Reference Gomes and Maia2024). Overall, most studies support the results found by Erickson and Mattson (Reference Erickson and Mattson1981). However, research on the impact of semantic relatedness in occurrences of semantic illusions also led to the postulate that participants seem to create a complete representation of the information presented by the question but are not able to match the information with the representation available in memory. Thus, the illusion question itself might have been encoded adequately, and all relevant information might have been retrieved, but the mismatch between the retrieved information and the information presented in the question remains unnoticed (e.g., Van Oostendorp & de Mul, Reference Van Oostendorp and de Mul1990; see also partial-match hypothesis or imperfect memory match hypothesis postulated by Reder & Kusbit, Reference Reder and Kusbit1991, for more details).
There are more recent hypotheses that consider a differential, although related, cognitive mechanism underlying the semantic illusion effect, namely the Node Structure Theory (NST) developed by Shafto and MacKay (Reference Shafto and MacKay2000). This model consists of a vast network of interconnected representational units called nodes, organized in semantic and phonological systems (see Figure 1).
Figure 1. The Node Structure Theory (NST) explained for the Moses-Illusion.
Source: Shafto and MacKay (Reference Shafto and MacKay2000, p. 375).
Taking Figure 1 into account, the participant might notice the inconsistent name Moses (i.e., the phonological input) when the Moses-Illusion is presented, but additional information might already be activated due to shared nodes with the correct word. For instance, information such as biblical characters and Old Testament might be primed. Information about the event of building the Ark and about the number of animals of each kind is given by the question (i.e., the semantic input). According to this model, Noah receives more priming than Moses and becomes activated under the most-primed-wins-principle. Consequently, participants do not notice the inconsistent name Moses.
Shafto and MacKay (Reference Shafto and MacKay2000) propose that not only semantic similarity (as also stated in previous research confirming the semantic similarity hypothesis, proposed by Erickson & Mattson, Reference Erickson and Mattson1981) but also phonological similarity might cause participants to fall for semantic illusions. In the first experiment, participants heard four versions of experimental questions: a valid version, which contained the appropriate name, and three invalid versions containing names that were semantically related, phonologically related or unrelated to the valid name. A partial shadowing task was used, demanding special attention to critical words because participants read a written version of each question containing blank slots and had to shadow whatever auditory word occupied each blank slot. Findings indicated that semantic relatedness between the inconsistent name and the correct word might not be the only explanation for occurrences of semantic illusions, ruling out purely semantic theories (e.g., Erickson & Mattson, Reference Erickson and Mattson1981; Van Oostendorp & Kok, Reference Van Oostendorp and Kok1990). Thus, the phonological similarity of words seems to play an important role in the semantic illusion effect as well. As an example of an illusion caused by phonological similarity, Shafto and MacKay (Reference Shafto and MacKay2000) presented the Armstrong-Illusion. After being exposed to the question, ‘What was the famous line uttered by Louis Armstrong when he first set foot on the moon?’, participants usually understand this question as being valid, even though they know that Louis Armstrong was a jazz musician and that the first man landing on the moon was, in fact, Neil Armstrong. However, the exclusion of purely semantic theories does not account for purely phonological theories of the Armstrong effect. More specifically, three summated factors, one phonological (i.e., the name Armstrong) and two semantic (i.e., information about being an astronaut and the first man on the moon), caused the unrelated name (i.e., Neil Armstrong) to accumulate more priming than the presented name (i.e., Louis Armstrong) and to become activated.
Some parallels can be found between the NST and the so-called ‘Good Enough’ approach (see Blott et al., Reference Blott, Rodd, Ferreira and Warren2021; Huang & Ferreira, Reference Huang and Ferreira2021), according to which readers are not always aware of semantic ambiguities, leading to misinterpretations of the read sentences. This occurs because readers resort to shallow processing to be more efficient. Consider semantic garden-path sentences, like ‘Sally worried that the ball was going to be too crowded for her liking’. Since ball admits two semantic representations (the toy and the dancing event), two interpretations are possible and both are activated even when one of them may be more dominant than the other. The theory predicts that, in such a scenario, readers create a superficial linguistic representation of the meaning behind the presented sentence and consider it ‘good enough’ (Ferreira, Reference Ferreira2003). The good-enough approach could also account for the occurrence of semantic illusions. Since both, the anomalous word and the correct word, share semantic representations, the participant might not be fully aware of the illusion. Thus, parallel to the most-primed-wins-principle, the sentence illustrated above offers information such as too crowded that might lead to more priming of the dancing event than of the toy.
The theories mentioned thus far provide a general understanding of semantic illusions. Overall, they agree with the idea that semantic illusions seem to occur due to some kind of semantic overlapping between the critical words. However, none of them can explain why the semantic illusion effect is modulated by factors such as the location of the erroneous word within the question or statement, formulation of instructions (e.g., Van Jaarsveld et al., Reference van Jaarsveld, Dijkstra and Hermans1997), the font of the written illusion (e.g., Song & Schwarz, Reference Song and Schwarz2008), and individual differences like reading fluency, long-term and working memory capacities (see Hannon, Reference Hannon, Rapp and Braasch2014 for a critical overview). It seems, therefore, that different cognitive mechanisms are involved in the failure to detect semantic anomalies, or as Hannon (Reference Hannon, Rapp and Braasch2014) stated, multiple sources of misinformation lead people to fall into illusions.
1.2. Semantic illusions in second language processing
The picture is even more complex when participants have to deal with more than one language. The question of whether the effect also appears during second language (L2) processing has not yet been fully explored and has guided the present research. This question may seem trivial at first; however, its answer has important implications for models of bilingual processing that assume that age of acquisition (AoA), and the degree of language proficiency can influence the way lexical access occurs, and consequently, how sentences are processed in the L2 (the Multilink model, Dijkstra et al., Reference Dijkstra, Wahl, Buytenhuijs, Van Halem, Al-Jibouri, De Korte and Rekké2019). Besides, studies report differences in types of processing that characterize the L1 and the L2 (a more intuitive processing for the L1 vs. a more controlled processing for the L2; Costa et al., Reference Costa, Foucart, Arnon, Aparici and Apesteguia2014; Keysar et al., Reference Keysar, Hayakawa and An2012) that may influence the semantic illusions effect. To the best of our knowledge, only a handful of studies have examined semantic illusions in bilinguals, and results are inconclusive (i.e., Bautista Martín, Reference Bautista Martín2022, unpublished master’s thesis; Dhaene et al., Reference Dhaene, Dirix, Van Marcke and Woumans2022; Vaessen, Reference Vaessen2017, unpublished master’s thesis; see also Geipel et al., Reference Geipel, Hadjichristidis and Surian2015).
Vaessen (Reference Vaessen2017) examined the effect of cross-linguistic differences (L1 vs. L2) and L2 proficiency on semantic illusions in late sequential Dutch-English bilinguals. The author expected that lower levels of English (L2) proficiency would have a negative effect on occurrences of semantic illusions (i.e., a higher proportion of semantic illusions) due to shallow sentence processing. In the first experiment, sentences were presented in an auditory manner. Four different conditions were tested: semantic illusions in L1 (condition 1), correct sentences in L1 (condition 2), semantic illusions in L2 (condition 3) and correct sentences in L2 (condition 4). Participants had to judge whether aurally presented questions were correct, incorrect or whether they did not know. Findings showed a significantly higher rate of semantic illusions in English (L2) than in Dutch (L1). However, no influence of L2 proficiency on semantic illusions was found in both assessed languages.
The second experiment consisted of a Rapid Serial Visual Presentation (RSVP) task accompanied by a scale for acceptability judgment. In this task, sentences were presented visually word by word instead of aurally, meaning that participants were unable to look back at the critical part of the sentence. No differences were found in the rate of illusion occurrences and in response times between L1 and L2. However, slight differences in response times were observed, indicating that bilinguals were faster to respond to semantic illusions in L1 than in their L2. Again, the level of L2 proficiency did not influence the overall results.
A recent eye-tracking study by Dhaene et al. (Reference Dhaene, Dirix, Van Marcke and Woumans2022) found similar results to those of Vaessen (Reference Vaessen2017). The authors postulated that the cognitive load hypothesis could account for the mechanisms underlying semantic illusions. They examined the effect of semantic illusions in late sequential high proficient Dutch (L1)-English (L2) bilinguals. Anomalous and non-anomalous questions were presented on a screen and participants had to respond to the questions aloud. Results showed a higher proportion of illusions and slower reading times in the L2 than in the L1, sustaining the above-mentioned hypothesis. Accordingly, the cognitive load would cause a lower detection rate of illusions for the L2 than for the L1 (and thus, a high percentage of semantic illusions), probably because of a partial matching of the semantic information (as discussed by Reder & Kusbit, Reference Reder and Kusbit1991). The opposite scenario would be expected if L2 processing was characterized by a more controlled processing, as Costa et al. (Reference Costa, Foucart, Arnon, Aparici and Apesteguia2014) hold.
Other studies have also made important contributions to this research domain. However, their research goals were not to intentionally study the semantic illusion phenomenon. For instance, Geipel et al. (Reference Geipel, Hadjichristidis and Surian2015) aimed to examine the foreign language effect (i.e., modulations in the cognitive processes responsible for judgment and decision-making when thinking in a non-native language vs. native language) in late but highly proficient sequential Italian (L1)-GER (L2) bilinguals. They conducted a series of experiments, in which the third one is of special interest for the present research. In this experiment, the Moses-Illusion was evaluated in both languages to test whether a foreign language promotes analytical or controlled reasoning (contrary to the more intuitive reasoning that characterizes L1 processing), as previous studies suggested (Costa et al., Reference Costa, Foucart, Arnon, Aparici and Apesteguia2014; Keysar et al., Reference Keysar, Hayakawa and An2012). Findings showed a similar rate of semantic illusions regardless of language, and thus the L2 did not improve performance on the Moses-Illusion task. Again, this conclusion is not consistent with the idea that an L2 promotes a switch from intuitive to more controlled processes (e.g., Costa et al., Reference Costa, Foucart, Arnon, Aparici and Apesteguia2014; Keysar et al., Reference Keysar, Hayakawa and An2012), but rather suggests that intuitive processes remain present regardless of the tested language.
To summarize, results from the above-mentioned studies, although not conclusive, seem to suggest a similar processing regardless of the language (L1 or L2). If differences across languages appear, these are explained in terms of the cognitive load hypothesis. Thus, the aim of the present research is to present further evidence in this line of research by examining the effect of semantic illusions in EP (L1)-GER (L2) bilinguals who vary in the age of L2 acquisition and L2 proficiency. A group of EP native speakers (without knowledge of GER) and another of GER native speakers (without knowledge of EP) were also tested to assess the suitability of materials to provoke semantic illusions and, if so, to compare the proportion of semantic illusions across populations. Participants were asked to read sentence-formed illusions instead of illusion questions. This methodological choice was based on recent work done with native speakers of Brazilian Portuguese, showing evidence that sentences with semantic illusions, despite being anomalous, are perceived as well-formed sentences (Gomes & Maia, Reference Gomes and Maia2024).
Having taken into consideration tenets of the above-discussed theories on native speakers, and more specifically, those raised with bilinguals, such as the cognitive load hypothesis (Dhaene et al., Reference Dhaene, Dirix, Van Marcke and Woumans2022), we expected to find no differences in the effect of semantic illusions across languages in early bilinguals with balanced language proficiency (as Bautista Martín, Reference Bautista Martín2022, has recently found). For late bilinguals, differences may, however, be observed between the languages in the form of more semantic illusions in L2 than in L1 (as well as slower reaction times), in line with postulates of the cognitive load hypothesis and consistent with what Dhaene et al. (Reference Dhaene, Dirix, Van Marcke and Woumans2022) and Vaessen (Reference Vaessen2017) observed, especially for participants with lower degrees of L2 proficiency.
Language proficiency may, besides, blur the picture around age of language acquisition (AoA), since the L1 is not always the dominant language of bilingual speakers (Flores et al., Reference Flores, Zhou and Eira2022). This is an important issue to examine since there is evidence that L2 proficiency and age of acquisition contribute independently to bilinguals’ comprehension and production behavior (García-Pentón et al., Reference García-Pentón, García, Costello, Duñabeitia and Carreiras2016; see also Bonfieni et al., Reference Bonfieni, Branigan, Pickering and Sorace2019; Sá-Leite et al., Reference Sá-Leite, Flores, Eira, Haro and Comesaña2023). Therefore, variations in proficiency and their impact on the semantic illusion effect were explored in terms of dominance. Dominance describes the difference in proficiency between two languages spoken by bilinguals and it is closely related to proficiency. We decided to include dominance in the analysis of the semantic illusions effect because recent research showed that dominance might affect bilingual language processing more than language proficiency, analyzed independently for L1 and L2 (Sá-Leite et al., Reference Sá-Leite, Flores, Eira, Haro and Comesaña2023).
2. Experiment 1
In the first experiment, we aimed to assess the appropriateness of our materials and to examine the occurrence rate of semantic illusions in both languages, EP and GER. Therefore, this first experiment was conducted with GER-native speakers and EP-native speakers only, serving as a control study for the second experiment with EP-GER bilinguals.
Occurrences of semantic illusions were expected to be observed in both groups of participants, i.e., slower responses and more errors to sentences with anomalous words, such as ‘After the plane crash, the authorities buried the survivors in the mountains of the Alps’, than to correct sentences, ‘ the dead’ , because of shared semantic nodes between correct (dead ) and anomalous words ( survivors) , as observed in previous studies (e.g., Erickson & Mattson, Reference Erickson and Mattson1981; Gomes & Maia, Reference Gomes and Maia2024; Shafto & MacKay, Reference Shafto and MacKay2000).
3. Method
3.1. Participants
A total of 62 participants were tested – 27 were native speakers of GER (Mage = 31.3; SD = 12.9) of which 16 were females (one did not want to say and one chose ‘other’), and 35 were native speakers of EP (Mage = 20.9; SD = 6) of which 30 were females (two did not want to say). All participants had normal or corrected-to-normal vision, normal hearing and were naïve about the main purpose of the study.
The experiment was carried out on the web-based Platform PCIbex (Zehr & Schwarz, Reference Zehr and Schwarz2018) to promote greater adherence to the study and to diminish the personal contact due to the pandemic situation of Covid-19. Participants were invited to participate through social networks (e.g., Facebook and Instagram) and personal contacts.
For the EP group, the same strategy was used, but EP students could also access an accreditation platform, which allowed them to improve their grades during the current semester through course credits after completing all the tasks of the study.
3.2. Materials
3.2.1. Stimuli
The task comprised 52 experimental sentences and 52 filler sentences in each language. Seventy percent of the sentences were taken from previous studies and translated into GER and EP by proficient EP and GER-speaking researchers (Barton & Sanford, Reference Barton and Sanford1993; Bredart & Modolo, Reference Bredart and Modolo1988; Gomes & Maia, Reference Gomes and Maia2024; Koornneef & Reuland, Reference Koornneef and Reuland2016; Muller et al., Reference Muller, Resnik and Phillips2020; Shafto & MacKay, Reference Shafto and MacKay2000; Vaessen, Reference Vaessen2017), whereas others were newly created following the structure of the former. Seventeen of the experimental sentences were of the type: ‘After the plane crash, the authorities buried the survivors in the mountains of the Alps’ (Gomes & Maia, Reference Gomes and Maia2024), and 35 experimental sentences were based on general knowledge as in: ‘After the great flood, Moses brought the animals in pairs to the Ark’.
In this type of sentences, the post-verbal element could be an anomalous element, such as ‘the survivors’ or a correct element ‘the dead’. Sentences were all grammatical, however, conceptual combination between verb and anomalous complement was not plausible. All sentences were manipulated with the goal that the first clause had a specific context, such as after the plane crash that should prime a specific verb-complement relationship, such as bury-dead. The control condition was composed of sentences in which the context primed the critical word. The semantic illusion condition was composed of control sentences, but with the verbal complement replaced by an anomalous complement, such as survivors. All sentences were divided into six segments and segments were controlled for length – one to eight words. Additionally, we tried to control for sentence structure, as best as possible. For EP, the sentences were composed of the following segments: sentence context – subject – verb – critical segment – rest of the sentence. GER sentences followed EP sentence structure, but the subject and the verb had to be switched due to the German verb second property. Thus, GER sentences had the following structure: sentence context – verb – subject – critical segment – rest of the sentence.
Additionally, corresponding target words in the control and illusion conditions were controlled for their frequency (logarithmic frequency, ps > .69) and length (all ps > .74) in each language. However, it is important to note that item length and sentence length showed differences across languages (ps < .01), due to the fact that GER words and phrases were generally longer than those in EP. Frequency values were obtained from the SUBTLEX-DE database (Brysbaert et al., Reference Brysbaert, Buchmeier, Conrad, Jacobs, Bölte and Böhl2011) and the SUBTLEX-PT database (Soares et al., Reference Soares, Machado, Costa, Iriarte, Simões, de Almeida, Comesaña and Perea2015). Words were also controlled for their phonological degree of similarity (NLD). The NLD between the anomalous and correct EP words was equal to 0.69 and the NLD between GER words was 0.71 (p = .109). Values for phonological NLD were taken from the PHOR-in-One database (Costa et al., Reference Costa, Comesaña and Soares2022).
A total of 104 sentences (52 experimental + 52 distractor) were created in both languages, EP and GER, in a 2 × 2 Latin square experimental design (two conditions per two languages), as in Table 1.
Table 1. Latin square: Condition vs. item
Across conditions, materials following the critical segment (the survivors vs. the dead) were kept the same: an adverbial phrase, such as in the Alps. This manipulation affected the material after the critical word and does not directly affect responses to this task on the processing of the critical segment. The use of this element is also intended to compensate for spill-over effects widely described in literature in self-paced reading tasks (see Jegerski, Reference Jegerski, Jegerski and VanPatten2014).
We employed a self-paced Rapid Serial Visual Presentation (RSVP) task with a moving window, wherein the reading segment disappears as soon as the next segment is revealed. Consequently, there is no possibility of returning to the previous segment. Thus, participants were forced to read one segment at a time and could not read the word before or after the critical word. However, participants could continue to process a segment as they read subsequent segments. This phenomenon is described in the literature as spillover effect: the difficulty induced by a particular segment could slow down the reading/processing time of one or more subsequent segments (Koornneef & van Berkum, Reference Koornneef and Van Berkum2006; Mitchell, Reference Mitchell, Kieras and Just1984; Smith & Levy, Reference Smith and Levy2013).
From these materials, four lists were created: two lists of stimuli in each language (one list with control sentences in EP and one in GER) and two lists of illusion sentences (one in EP and one in GER). Each participant read only one sentence of each type, half of the experimental phrases contained the illusion term, while the other half contained the correct term. Thus, a total of 104 sentences (52 experimental + 52 fillers) comprised each list in both languages. In addition, each list containing experimental sentences (control and illusion) and fillers was pseudo-randomized, respecting the criterion of including at least one distractor item among experimental items. Four lists were randomly presented to participants. This design allowed for testing the same materials in both control and illusion paradigms (see Table 1) in a within-subjects paradigm.
3.2.2. General knowledge test
Additionally, eight lists were created for the general knowledge test on Google Forms (Google, 2022) for each condition and each language. The Google Form contained semantic illusions that had been previously shown to participants and required general knowledge. The test was designed as a multiple-choice test in which participants had to indicate whether they knew the answer or not. For example, when presented with a general knowledge sentence such as ‘After the great flood, Moses brought the animals in pairs onto the Ark’, a question focusing on the target followed: ‘Did Moses bring the animals onto the Ark?’. Participants were given three possible choices: ‘yes’, ‘no’ or ‘don’t know’. General knowledge was tested as a prerequisite for a semantic illusion to occur; participants must have some world knowledge about the information presented. If the complete semantic concept is unfamiliar, participants would not fall for the illusion (Van Oostendorp & Kok, Reference Van Oostendorp and Kok1990).
3.2.3. Language experience and proficiency assessment
A sociodemographic questionnaire and the standardized DIALANG Vocabulary Size Placement Test (see Alderson, Reference Alderson2005) were used to assess participants’ language experience and proficiency. The questionnaire included questions about participants’ age, sex, education level, and knowledge of any second language. The standardized DIALANG Vocabulary Size Placement Test (see Alderson, Reference Alderson2005) is a lexical decision task in which participants must decide if a certain sequence of letters exists as a word in the given language. Monolingual controls completed the task in their native language – EP or GER. Lexical competence has been argued to be a valid predictor of language proficiency (Laufer & Nation, Reference Laufer and Nation1999; Treffers-Daller & Korybski, Reference Treffers-Daller, Korybski, Silva-Corvalán and Daller2016), because speakers’ lexical inventory grows as proficiency increases. The DIALANG Vocabulary Size Placement Test assesses lexical knowledge through a list of 75 items, including 50 real words and 25 non-words. Following Alderson’s (Reference Alderson2005) proposal, participant’s scores are computed based on the total number of words correctly identified as either real words or non-words, with 1 point awarded per correct assessment (see also Flores et al., Reference Flores, Zhou and Eira2022).
3.3. Procedure
Before starting the experiment, all participants were asked to sign the online informed consent form, which aimed to explain the main objectives of this study and the confidentiality of participants’ data. Upon accepting to participate in the study, participants were directed to perform the self-paced RSVP. One of the advantages of the self-monitoring reading paradigm is its compatibility with web-based platforms. This study was hosted on the web-based platform PCIbex – PennController for Internet-Based Experiments (Zehr & Schwarz, Reference Zehr and Schwarz2018). Participants were tested in a single online session lasting approximately 15 minutes (inclusive of the informed consent and instructions).
Prior to the experimental session, participants needed to complete a training session in their native language with 10 sentences each. As in a self-paced reading experiment, participants determined the speed at which sentence segments were presented on the screen by pressing the space button, allowing for the measurement and evaluation of reading times (RTs) for each segment. Each segment was presented in a moving-window paradigm (SPR-moving-window) at the center of the screen.
Each experiment consisted of the following events: A sentence trace is revealed, segment by segment, in the center of the screen. After the participant pressed the space button, the next segment is revealed until they read the entire sentence. The time a segment is visible is used as a dependent measure and reflects reading time (RT). After reading the entire sentence, a question appears on the screen: Does it make sense? Participants must then decide whether the read sentence makes sense in EP or GER. They were instructed to answer, ‘Yes’, if the sentence was semantically coherent and grammatically well-formed, and ‘No’ otherwise. Participants answered the questions by pressing one of two buttons, which were counterbalanced (left and right) between participants. Each participant was randomly assigned to a counterbalanced list of stimuli and was instructed to take the test on a computer, while reading as normally as possible, trying to understand the sentences, and answering the questions as accurately as possible. An 85% accuracy threshold on globally plausible items in the behavioral task was set to ensure that participants were paying attention.
Finally, participants were asked to respond to the sociodemographic questionnaire, the general knowledge test and the standardized DIALANG Vocabulary Size Placement Test. Links to the general knowledge test (corresponding to the list they were assigned to for the self-paced RSVP) and the standardized DIALANG Vocabulary Size Placement Test were sent to each of the participants after they completed the sociolinguistic questionnaire.
4. Results
4.1. Results for ‘yes’ responses
Error rates (%E) and reading times (RT) in ms obtained for the first segment of the sentences, in which the anomalous word or correct word appeared (RT1), and for the segment following the first one (RT2) were analyzed using linear mixed-effect models (e.g., Baayen, Reference Baayen2008; Baayen et al., Reference Baayen, Davidson and Bates2008). The analysis of the second segment was included in order to examine spillover effects. It is important to note that, in the RT analyses, we included correct responses to control items (‘yes’ responses) and errors to semantic illusion items (‘yes’ responses). To this end, the lme4 package of R was used (Bates et al., Reference Bates, Mächler, Bolker and Walker2015). A series of fixed structure models were created to examine the hypotheses of the study, with the %E or RTs as dependent variables. As fixed effects, models included Condition (control or semantic illusion) and Length (number of letters of the item) for native speaker analyses. Length was centered and transformed into Z-scores. To control for variability due to participants and items (including the type of illusion) involved in the experiment, we introduced participants and items as random effects in the models. As items could appear in two different conditions (control or semantic illusion), we also included a slope of Condition in the random effect of items. We were unable to increase the complexity of the random structure of the models as this would not allow them to converge. The formula of the models was the following: (%E or RT) ~ condition + length + (1 | participant) + (condition | item).
The significance of effects and interactions was determined using log-likelihood ratio tests (R anova function). The contribution of each effect and interactions were assessed by comparing a model that included them with another model in which they were not included. Also, results of t-test analyses for coefficient estimates of fixed effects are reported. To this end, Satterthwaite’s approximations to the degrees of freedom of the denominator were used (p-values were estimated by the lmerTest package; Kuznetsova et al., Reference Kuznetsova, Brockhoff and Christensen2017).
4.1.1. EP group
Prior to the analyses, some data filtering was conducted to remove observations with RTs below 300 ms or above 3000 ms, as well as observations corresponding to general knowledge items for which the participant did not know the answer. This resulted in the removal of 8.22% of total observations. No participant or item was excluded from the analyses. See Table 2 for mean RT and %E of the experiment.
Table 2. Mean RT and %E of each experimental condition for the EP group
Note: RT1 shows the mean RT for the critical segment, whereas RT2 shows the mean RT for the segment following the critical one. Standard errors are included in parentheses.
Results revealed no significant effect (p > .05) of item (between straight relation verb-complement and contextual semantic illusion) and no significant effect of Condition on %E (estimate = 0.43, SE = 0.28, 95% CI [−0.14, 1.03], z = 1.52, p = .129, χ 2(1) = 2.26, p = .133). However, the number of semantic illusions was slightly higher than the errors for control sentences (i.e., difference of 4.8%). In terms of RT1s, there was also no significant effect of Condition on the critical segment (estimate = 38.90, SE = 31.83, 95% CI [−26.49, 100.86], t = 1.22, p = .227, χ 2(1) = 1.50, p = .220). RT1s for control sentences were found to be only slightly lower than for semantic illusions (i.e., difference of 28 ms). In contrast, when considering RT2s (i.e., RTs for a segment following the critical one), results showed a significant effect of Condition (estimate = 94.09, SE = 28.78, 95% CI [35.55, 150.40], t = 3.27, p = .002, χ 2(1) = 10.18, p = .001), indicating significant slower RTs in semantic illusions compared to control sentences (i.e., difference of 59 ms).
4.1.2. GER group
The same data filtering procedure as the one used for EP native speakers was applied, resulting in the removal of 7.26% of total observations. No participant or item was excluded from the analyses. See Table 3 for mean RT and %E of the experiment.
Table 3. Mean RT and %E of each experimental condition for the GER group
Note: RT1 shows the mean RT for the critical segment, whereas RT2 shows the mean RT for the segment following the critical one. Standard errors are included in parentheses.
In the GER group, results indicated a significant effect of Condition on %E. Participants committed more errors in the illusion condition than in the control condition (estimate = 1.15, SE = 0.36, 95% CI [0.41, 1.89], z = 3.20, p = .001, χ 2(1) = 9.60, p = .002), showing a difference of 15.4%. Regarding RT1s, the effect of Condition on the critical segment was not significant (estimate = 21.95, SE = 36.74, t = 0.60, 95% CI [−50.21, 98.42], p = .553, χ 2(1) = 0.37, p = .545), suggesting that control sentences and semantic illusions yielded close RT1s. Still, participants responded slightly slower to the illusion sentences than to the control sentences (i.e., difference of 40 ms). Additionally, the effect of Condition on the segment following the critical segment was also non-significant (estimate = −8.23, SE = 28.46, 95% CI [−71.88, 42.10], t = 0.29, p = .774, χ 2(1) = 0.08, p = .773), further suggesting that there were no differences in RT2s between control sentences and semantic illusions in this segment as well. Yet, again, participants responded slightly slower to the illusion sentences than to the control sentences (i.e., difference of 11 ms).
4.2. Results for correct responses
We also ran the analyses on RTs for correct responses to control items (‘yes’ response) and correct responses to semantic illusion items (‘no’ response). See Tables 4 and 5.
Table 4. Mean RT and %E of each experimental condition for the EP group
Note: RT1 shows the mean RT for the critical segment, whereas RT2 shows the mean RT for the segment following the critical one. Standard errors are included in parentheses.
Table 5. Mean RT and %E of each experimental condition for the GER group
Note: RT1 shows the mean RT for the critical segment, whereas RT2 shows the mean RT for the segment following the critical one. Standard errors are included in parentheses.
4.2.1. EP group
In terms of RT1s, there was a significant effect of Condition on the critical segment (estimate = 57.82, SE = 22.99, 95% CI [12.01, 104.29], t = 2.52, p = .016, χ 2(1) = 5.95, p = .015). RT1s for semantic illusions were slower in comparison to control sentences (i.e., difference of 55 ms). When considering RT2s (i.e., RTs for a segment following the critical one), results also showed a significant effect of Condition (estimate = 105.03, SE = 25.06, 95% CI [50.29, 144.15], t = 4.19, p < .001, χ 2(1) = 15.01, p < .001), indicating slower RT2s in semantic illusions compared to control sentences (i.e., difference of 109 ms).
4.2.2. GER group
Regarding RT1s, the effect of Condition on the critical segment was significant (estimate = 102.52, SE = 29.76, 95% CI [45.37, 164.26], t = 3.45, p = .001, χ 2(1) = 10.61, p = .001), showing slower RT1s for semantic illusions than for control sentences (i.e., difference of 96 ms). In contrast, the effect of Condition on the segment following the critical segment was non-significant (estimate = 53.42, SE = 32.96, 95% CI [−11.45, 120.39], t = 1.62, p = .112, χ 2(1) = 2.59, p = .107). However, participants responded slower to the semantic illusions when compared to the control sentences (i.e., difference of 50 ms).
5. Discussion
The primary goal of the first experiment was to examine the suitability of the materials used in eliciting the effect of semantic illusions on %E and RTs in EP and GER native speakers, and if so, to determine if semantic illusions occur in both languages. For the group of EP speakers, it was observed that participants did not make significantly more errors in the illusion condition than in the control condition, contrary to initial expectations. However, it is important to note that %E was numerically higher for the illusion condition compared to the control condition (4.8% of semantic illusions, see Tables 2 and 4). Yet, it is essential to emphasize once again that a condition effect was observed for RT2. The effect was more evident for GER speakers – participants committed significantly more errors in the illusion condition when compared to the control condition (13.4% of semantic illusions). In contrast to EP, however, no significant differences in RTs (neither RT1 nor RT2) between both conditions were shown, although RTs were slightly slower for the illusion condition. As for correct responses to control items (‘yes’ response) and correct responses to semantic illusion items (‘no’ response), it is important to note that RTs showed a condition effect for EP and GER, suggesting that participants needed more time to respond accurately to illusion sentences than to control sentences.
As a possible explanation for why the effect was nearly three times larger in GER speakers than in EP speakers, we propose that the robust semantic illusion effect is influenced by participants’ reading speed, as GER participants were slower than EP participants (see Tables 2 and 3, and Tables 4 and 5). Consequently, more room may exist for the effect of semantic illusions to manifest. However, an analysis of the correlation between the mean of RT and %E showed a non-significant relationship (r = −.19; p = .136). Hence, there seems to be another explanation behind these results that cannot be solely attributed to participants’ reading speed. This issue will be further addressed in the General Discussion.
6. Experiment 2
The second experiment aimed at examining the occurrence of semantic illusions in each of the two languages (EP and GER) spoken by the early and late EP-GER bilinguals. In particular, %E and RTs to each semantic illusion phrase were assessed.
As mentioned earlier, no differences between the two languages were expected for early bilinguals (see also Bautista Martín, Reference Bautista Martín2022), while a higher occurrence rate of semantic illusions in the L2 than in the L1 was predicted for late bilinguals (as found by Dhaene et al., Reference Dhaene, Dirix, Van Marcke and Woumans2022; Vaessen, Reference Vaessen2017). Furthermore, it was expected that no impact of dominance would be observed, given previous work showing that semantic illusions occur independently of the level of proficiency (see Vaessen, Reference Vaessen2017).
7. Method
7.1. Participants
Data of 40 bilingual speakers of EP and GER with a mean age of 36.9 years (SD = 11.2), 33 females, were collected. As inclusion criterion, participants had to have normal or corrected-to-normal vision, normal hearing and had to be naïve about the objective of the study. In addition, they had to be able to maintain a fluent conversation in EP and GER.
All 40 bilinguals were L1 speakers of EP. Twenty-one acquired GER in childhood in a naturalistic context, before the age of 10, being early sequential bilinguals. Nineteen were late learners, having acquired GER as teenagers in a classroom context. More information on the speakers’ sociolinguistic background was collected in a sociodemographic questionnaire and the standardized DIALANG Vocabulary Size Placement Test, and is available in Table 6 (see Section 7.2).
Table 6. Summary of the bilinguals’ background variables and DIALANG results
Note: The total score of self-evaluation is the sum of the speaking and the writing scores (maximal score: 14). The difference between the score obtained in the GER and in the EP test (negative values indicate dominance in EP; positive values indicate dominance in GER; 0 equal dominance for EP and GER) is abbreviated as DIALANG Dif.
7.2. Materials
7.2.1. Stimuli
Stimuli remained the same as those used in the first experiment. Since the second experiment involved testing bilingual participants, it was necessary to create eight experimental lists: the four lists used in the first experiment, but now in both EP and GER.
7.2.2. General knowledge test
The same general knowledge tests were used. Bilingual participants completed the tests in their L1.
7.2.3. Language experience and proficiency assessment
In addition to the same questions asked in Experiment 1, information on the age of onset of bilingualism and place of residence was collected. Participants were also asked to self-rate their proficiency in reading, writing and speaking in each of the languages on a scale from 1 to 7 (Likert Scale) and to indicate the percentage of time they spend per day using each of the languages in a week. This allowed us to examine the age of L2 acquisition, participants’ subjective proficiency, and degree of use for both languages. Finally, they were asked whether they have ever lived abroad and if so, for how long, so that some insight about their linguistic environment throughout their lives could be obtained.
Furthermore, the same standardized DIALANG Vocabulary Size Placement Test (see Alderson, Reference Alderson2005) was administered twice – once in GER and a second time in EP, or vice versa. Table 6 summarizes the main background variables.
As can be seen in Table 6, early and late bilinguals presented similar rates of current contact with GER (early: 30.2 [19.6]; late: 30.4 [21.9]). A t-test showed no group differences regarding current contact with GER (p = .98). As expected, current contact with EP was higher compared to GER (early: 58.3 [22.4]; late: 50.7 [23.3]), but again no statistically significant group differences were observed (p = .37). This indicates that, at time of testing, both groups had similar profiles regarding the amount of contact with the tested languages. As for the difference in self-evaluation in GER, early bilinguals assessed their proficiency higher than late learners (early: 12.3 [1.6]; 9.9 [2.8]), reaching statistical significance (p = .008). The opposite was observed for self-evaluation in EP, with late bilinguals assessing their proficiency in EP higher than early bilinguals (early: 12.7[11]; late: 13.5[0.9]). This difference was also significant (p = .038). When considering the DIALANG values, the difference between early and late learners was significant for GER (p < .001), but not for EP (p = .58), as expected.
7.3. Procedure
The procedure for all tasks was similar to the one used in Experiment 1. The only difference was that bilinguals were solicited to conclude the self-paced RSVP task in both languages. Therefore, after signing the informed consent form, each participant did half of the self-paced RSVP in EP and half of the experiment in GER in a single session, in two blocks separated by a short break and a short message to make participants aware of the language changing. The order of the languages was counterbalanced – half of the participants started the experiment in EP and the other half started it in GER. Half of the sentences within each block contained the correct verbal complement (control), and the other half the incorrect one (illusion), following the same procedure as for native speakers. In addition, the experimental sentences (control and illusion) and fillers in each of the eight lists were pseudo-randomized. Additionally, the eight lists with fillers were randomly presented. Finally, participants were requested to respond to the sociodemographic questionnaire, the General Knowledge Test and the DIALANG Vocabulary Size Placement Test.
8. Results
The same data filtering procedure was applied as for the EP native speakers and GER native speakers, resulting in the exclusion of 5.05% of total observations from the analyses. No participant or item was excluded. For bilingual analyses, fixed effects were the triple interaction and second-order interactions between Condition (control or semantic illusion), Target Language (GER or EP) and Age of Acquisition (AoA; early or late). Language DominanceFootnote 1 and Length (number of letters of the item) were introduced as covariables (~ condition × target_lang × AoA + lang_dominance + length (1 | subject) + (condition | item)). Language Dominance and Length were centered and transformed into Z-scores. In addition, dichotomous variables were coded using sum contrast coding (−0.5 for the first level and +0.5 for the second level of each factor); Condition: control (−0.5), illusion (+0.5); Target Language: GER (−0.5), EP (+0.5); AoA: early (−0.5), late (+0.5). As in Experiment 1, we introduced participants and items as random effects in the models, and we also included a slope of Condition in the random effect of items. In the RT analyses, we included correct responses to control items (‘yes’ response) and error responses to semantic illusion items (‘yes’ response). See Table 7 for mean RT and %E of the experiment.
Table 7. Mean RT and %E of each experimental condition for the bilingual group
Note: RT1 shows the mean RT for the critical segment, whereas RT2 shows the mean RT for the segment following the critical one. Standard errors are included in parentheses.
Analyses showed a significant interaction on %E between Target Language and Condition, χ 2(1) = 7.54, p = .006. As can be seen in Figure 2, although the %E in the illusion condition was significantly higher than in the control condition in GER (estimate = 0.61, SE = 0.23, 95% CI [0.16, 1.06], z = 2.67, p = .008) and EP sentences (estimate = 1.27, SE = 0.23, 95% CI [0.83, 1.72], z = 5.61, p < .001), the effect was larger in EP sentences than in GER sentences. No other significant interactions were found in the analyses of %E. Note that RTs to GER stimuli were again slower than to their EP counterparts, although this finding was mediated by the interaction between target language and AoA (see Figures 3 and 4).
Figure 2. Interaction between target language (GER vs. Portuguese, PT) and condition (control vs. illusion) on %E.
Figure 3. Interaction between target language (GER vs. Portuguese, PT) and AoA group (early vs. late) on RTs of the critical segment.
Figure 4. Interaction between target language (GER vs. Portuguese, PT) and AoA group (early vs. late) on RTs of the segment following the critical segment.
Data analysis on RTs to the critical segment (see Figure 3) showed an interaction between AoA and Target Language, χ 2(1) = 9.00, p = .003, due to faster RT1s in late bilinguals compared to early bilinguals in EP (estimate = −270.2, SE = 111.0, 95% CI [−495, −45.1], t = 2.44, p = .020), but not in GER (estimate = −80.8, SE = 112.0, 95% CI [−308, 146], t = 0.72, p = .474). This interaction was also observed in RTs of the segment following the critical segment (see Figure 4), χ 2(1) = 12.34, p < .001, again with faster RT2s in late bilinguals compared to early bilinguals in EP (estimate = −277.9, SE = 105.0, 95% CI [−491, −64.6], t = 2.65, p = .012), but not in GER (estimate = −66.4, SE = 116.0, 95% CI [−282, 149.2], t = 0.62, p = .536). In addition, there was a significant Condition effect in RT2s (estimate = 95.65, SE = 37.98, 95% CI [21.11, 170.19], t = 2.52, p = .015), showing that RT2s for semantic illusions were slower than for control sentences.
As in Experiment 1, we conducted analyses for correct responses to control items (‘yes’ response) and correct responses to semantic illusion items (‘no’ response). See Table 8.
Table 8. Mean RT and %E of each experimental condition for the bilingual group
Note: RT1 shows the mean RT for the critical segment, whereas RT2 shows the mean RT for the segment following the critical one. Standard errors are included in parentheses.
Data analysis on RTs to the critical segment showed an effect of Condition (estimate = 124.07, SE = 36.35, 95% CI [52.74, 195.39], t = 3.41, p = .001), suggesting that RT1s for semantic illusions were slower than for control sentences. Analyses also showed a significant triple interaction between Group, Target Language and Condition on RT2s, χ 2(1) = 3.88, p = .049. As can be seen in Figure 5, this interaction reflects an effect of semantic illusions which was restricted to EP sentences in the group of late bilinguals (see Figure 5), late bilinguals showed slower RT2s in the illusion condition than in the control condition (estimate = 144.9, SE = 60.2, 95% CI [−10.5, 300.3], t = 2.41, p = .017).Footnote 2
Figure 5. Interaction between condition, target language (GER vs. Portuguese, PT), and AoA group (early vs. late) on RTs of the segment following the critical segment.
9. Discussion
The aim of Experiment 2 was to examine the effect of semantic illusions in EP-GER bilinguals who vary in age of L2 acquisition and L2 proficiency. It was hypothesized that no differences would be observed for early learners of GER, but for late learners more semantic illusions were expected to occur in the L2 than in the L1 (as observed by Dhaene et al., Reference Dhaene, Dirix, Van Marcke and Woumans2022; Vaessen, Reference Vaessen2017). Results were clear-cut: we observed an overall effect of SI in the %E in both languages (although larger in EP) regardless of group (early and late). When looking at RTs from ‘yes’ responses to correct and incorrect conditions (especially RT2s), results also showed that participants needed more time to respond to semantic illusions than to control sentences. Late bilinguals exhibited faster RTs in EP compared to early bilinguals, who acquired EP as L1 in a GER-dominant setting (as heritage language). This is explained by the fact that early bilinguals are less proficient in EP than late bilinguals, who were raised monolingual in an EP-dominant context. This difference in proficiency was evident from the self-evaluation and the DIALANG test results (see Section 7.2). Crucially, when RTs from correct responses (i.e., ‘yes’ responses to correct sentences and ‘no’ responses to incorrect ones) were considered (especially RT2s), the effect of SI was restricted to late bilinguals in their L1 (EP). No effects were observed in GER.
10. Overall discussion
The present study had the main goal to explore the effect of semantic illusions in early and late EP-GER bilinguals who vary in their age of onset of acquisition of GER and in their degree of language proficiency. Two groups of native speakers – GER native speakers and EP native speakers – worked as controls and were also tested in an attempt to assess the efficacy of materials to produce semantic illusions. Results showed a significant effect of semantic illusions on %E which was restricted to native speakers of GER. The pattern of results was similar for native speakers of EP but it was only significant for RT2. It is important to note that the percentage of semantic illusions observed in these two groups was similar to the results observed in Bautista Martín’s (Reference Bautista Martín2022) study using the same procedure, namely around 15%. Likewise, differences were observed for the native groups between the semantic illusion condition and the correct condition when RTs were analyzed for correct responses to control items (‘yes’ response) and error responses to semantic illusion items (‘yes’ response), suggesting that participants took longer to process sentences with anomalous words than correct sentences.
Overall, results observed for native speakers suggest that semantic illusions influenced the %E, at least in GER, due to similarities between the anomalous and the correct word, supporting the semantic/phonological similarity hypothesis proposed by Shafto and MacKay (Reference Shafto and MacKay2000). Although the effect was not significant in EP, participants tended to commit more errors when confronted with sentences from the semantic illusion condition than when confronted with sentences from the control condition. Given that in the control group the percentage of illusions showed to be relatively small (first experiment with native speakers of EP), we suggest that the effect should also be explored with questions (as also proposed by Bautista Martín, Reference Bautista Martín2022). The percentage of semantic illusions observed in previous studies using questions was substantially higher than the one found in the present study (ranging from 43% in the study by Erickson & Mattson, Reference Erickson and Mattson1981, to 60% in the study by Dhaene et al., Reference Dhaene, Dirix, Van Marcke and Woumans2022).
Still, the reason why the effect was higher in the GER control group than in the EP control group is unclear. We tentatively hypothesize that participants’ reading speed when reading in GER could be explaining the results because GER and EP have distinct language characteristics. However, correlation analyses between RTs and %E debunked this hypothesis, as above mentioned. Besides, results from the bilinguals show clear illusion effects in both languages, EP and GER.
In fact, a major finding of the present study is that early as well as late bilinguals show illusion effects in terms of %E in both their languages, independently of AoA, even though the effect is larger in EP than in GER. Furthermore, language dominance also did not influence the occurrence of illusions. When considering the analyses for correct responses to control sentences and incorrect responses to illusion sentences, early and late bilinguals only differ in their RT2s in EP, but not in GER. Additionally, when looking at correct responses to control sentences and correct responses to illusion sentences, a clear semantic illusion effect is shown for RT1s – participants need more time to respond to illusion sentences. This result in RT1s was independent of language and AoA. Yet, for RT2s, a triple interaction suggests that RT2s are slower for the illusion condition in EP, but only for late learners.
Thus, semantic illusions occur equally in L1 and L2, similarly to what was observed in the second experiment of Vaessen’s (Reference Vaessen2017) study using a Rapid Serial Visual Presentation (RSVP) and Geipel et al.’s (Reference Geipel, Hadjichristidis and Surian2015) study. Geipel et al. (Reference Geipel, Hadjichristidis and Surian2015) observed no difference in occurrence rate between L1 and L2 and concluded that intuitive processes remain active in L1 and L2 during semantic illusion processing, opposed to the cognitive load hypothesis that supports the idea that an L2 promotes a switch from intuitive to more controlled processes (e.g., Costa et al., Reference Costa, Foucart, Arnon, Aparici and Apesteguia2014; Keysar et al., Reference Keysar, Hayakawa and An2012). Although Bautista Martín (Reference Bautista Martín2022) studied solely balanced bilinguals that acquired Spanish and Catalan early in their life, and thus a direct comparison of our results with those of Bautista Martín (Reference Bautista Martín2022) is not possible, it is worth to mention that Bautista Martín (Reference Bautista Martín2022) observed semantic illusions in both languages, as we did in the present study.
If we take into consideration the percentage of semantic illusions observed in previous studies, similar percentages were obtained for the sentences in the present research. For instance, Van Oostendorp and Kok (Reference Van Oostendorp and Kok1990) observed a percentage of semantic illusions of approximately 32%. Similarly, Vaessen (Reference Vaessen2017), who collected data of Dutch-English bilinguals, showed a semantic illusions rate of roughly 30%. These percentages are very close to those observed in the present study (33% for monolingual native speakers, 33.1% and 43.1% for early and late bilinguals in GER respectively).
As we observed semantic illusions in L1 and L2, we confirm the postulates of the NST and related accounts to the semantic similarity hypothesis. Illusions occurred in both languages due to semantic similarities between critical words. As a consequence, theories seem to be more complementary than mutually exclusive. Indeed, results could also be described by the partial-match hypothesis or imperfect memory match hypothesis (Reder & Kusbit, Reference Reder and Kusbit1991; see also Van Oostendorp & de Mul, Reference Van Oostendorp and de Mul1990), given that the semantic overlapping between critical words may promote a partial match of the semantic information to stored mental representations (Dhaene et al., Reference Dhaene, Dirix, Van Marcke and Woumans2022, see also Reder & Kusbit, Reference Reder and Kusbit1991), leading to semantic illusions. We recognize, however, that more research is needed in order to gain a more complete picture of the mechanisms underlying the effect of semantic illusions in other bilingual populations, speaking more or less closely related languages because linguistic transfer seems to be higher across languages with similar features (e.g., Spanish-Catalan bilinguals; Bautista Martín, Reference Bautista Martín2022).
11. Conclusion
The main purpose of the present study was to examine the mechanisms leading to semantic illusions in bilingual processing by comparing early and late bilinguals. The above discussed findings for EP-GER bilinguals lead to the conclusion that no differences occur in the detection rate of semantic illusions in the L1 and L2, independently of AoA. These findings suggest that similar mechanisms underlie the occurrence of semantic illusions in monolingual and bilingual speakers. This may happen because intuitive processing leads to a partial match between the semantic information and the stored mental representations in the L1 and the L2 as complementary hypotheses hold, namely the partial-match hypothesis or the imperfect memory match hypothesis (Reder & Kusbit, Reference Reder and Kusbit1991; see also Van Oostendorp & de Mul, Reference Van Oostendorp and de Mul1990). Moreover, the results showing a similar percentage of semantic illusion in the L1 and L2 replicate those of previous studies (i.e., Geipel et al., Reference Geipel, Hadjichristidis and Surian2015; Vaessen, Reference Vaessen2017).
Acknowledgments
This study was conducted at the Psychology Research Centre (CIPsi/UM), School of Psychology and the Center of Humanities (CEHUM), School of Arts and Humanities, University of Minho, supported by the Foundation for Science and Technology (FCT) through the Portuguese State Budget (UIDP/01662/2020 and UID/00305/2020) and the Center of Linguistics (CLUP), the Department of Portuguese and Romance Studies (DEPER), University do Porto. The current study has also won the ‘Prémio UMinho de Iniciação na Investigação Científica – 2021’ for which we would like to thank the University of Minho.
Data availability statement
The code and data used in the experiments are available at: https://url.avanan.click/v2/___https://osf.io/a8vu6/?view_only=f69a25bf18674ab795bf7dfc22a4b4ea___.YXAxZTpjYW1icmlkZ2Vvcmc6YTpvOjMzYWM3MTFhMzRlYjFlNDczNzkyZTc0ODM1MjJhMmYwOjY6NjY2NjoyZjhjYWM0MDk1OTI4NzI4MTA3YzE0MzRiYTgzZTRlN2UzZTJjNTMyZGM3ZGY1Y2Q4ODhlZDVkZTc1MjY3MGRmOnQ6VDpG.
Open access
