Cross-linguistic influence and language co-activation in acquiring L3 words: What empirical evidence do we have so far?

Abstract

Third language (L3) lexical acquisition is still underexplored. In this article I overview theoretical and empirical evidence on L3 lexical acquisition and the role of cross-linguistic influence (CLI) in learning L3 words. I explain the mechanism of CLI as resulting from language co-activation in the multilingual learner’s/user’s mind. Consequently, I aim to ground L3 lexical studies in previous research on second language (L2) word processing and learning, and to emphasize the role of cross-linguistic similarity (cognates and false cognates) in the process. While capitalizing upon similarity predominantly facilitates acquiring L2 and L3 words, the precise mechanisms of L3 lexical acquisition are still obscured. It is unclear whether any overlap of an L3 form with the native or L2 form suffices to boost learning, or whether all previous languages influence L3 lexical acquisition cumulatively. To seek answers to this issue, I review empirical evidence for CLI and cross-linguistic similarity in L3 vocabulary acquisition from three research strands: L3 word processing experiments, L3 cognate guessing tasks, and L3 word learning experiments. Overall, this article aims to bridge the gap between psycholinguistic and applied linguistic research on L3 lexical acquisition, and argues for controlling an array of variables modulating research outcomes.

I Introduction

The mechanisms of learning words in the third language (L3) have not yet been convincingly explained. ¹ Since L3 acquisition depends on an interaction between the ‘new’ and the ‘old’ linguistic knowledge at the learner’s disposal (Angelovska and Hahn, 2017; Puig-Mayenco et al., 2020), the reliance on previous languages makes L3 acquisition more complex than second language (L2) acquisition. The learner might use their knowledge and learning experience from the native language (L1) and the L2, subconsciously or purposefully, as scaffolding for the new L3 knowledge. Since more languages are involved, the patterns of cross-linguistic influence (CLI) in L3 acquisition are essentially more diverse than in the L2 acquisition.

While there is a growing body of empirical evidence on the role of CLI in the acquisition of L3 morpho-syntax (e.g. Angelovska and Hahn, 2017; González Alonso and Rothman, 2017; Puig-Mayenco et al., 2020; Westergaard et al., 2017), L3 word learning and its dependence on CLI is much less explored and explained. The existing overviews concerning L3 acquisition have focused on: (1) morpho-syntactic transfer, in particular on the theoretical proposals attempting to model linguistic transfer (Puig-Mayenco et al., 2020), (2) cognitive and strategic advantages of multilingual children and adults in novel language learning (Festman, 2021; Hirosh and Degani, 2018), and (3) evidence for CLI in L3 during online (real-time) processing of syntax and vocabulary (Festman, 2018; Lago et al., 2021). In contrast, this review uniquely focuses on L3 vocabulary acquisition, thus considerably extending the previous ones. I aim to present quantitative empirical evidence concerning the role of CLI and cross-linguistic similarity (cognateness) in L3 acquisition from three types of studies: (1) psycholinguistic experiments on L3 word processing, (2) L3 cognate guessing tasks in receptive multilingualism studies, and (3) experiments in learning L3 words. Throughout the review, I pinpoint how the acquisition of L3 words is modulated by individual differences between language learners, and argue for attention to an array of variables in future research on L3 acquisition.

II Patterns of CLI and language co-activation in L3 acquisition

Part and parcel of multilingual language knowledge and use is the co-existence of several languages in the user’s mind (or user’s multicompetence, e.g. Cook, 2016; Festman, 2021) and interactions of those languages across various subsystems of language (vocabulary, morpho-syntax, phonology). Such interactions, deemed cross-linguistic influence (CLI; Kellerman and Sharwood Smith, 1986), are not restricted to L1 transfer effects on the L2, or the L3 (Kellerman and Sharwood Smith, 1986; Sharwood Smith, 2021). ² All languages known to any extent co-exist in the user’s/learner’s mind forming a network, and can potentially influence one another. Understood in this way, the basic mechanism of CLI consists in co-activating alternate connections across the language network in the multilingual user’s mind (Sharwood Smith, 2021) leading to interactions between the ‘new’ and the ‘old’ linguistic knowledge (Angelovska and Hahn, 2017; Puig-Mayenco et al., 2020). This reasoning is in line with the cognitive views on the co-activation (or competing activation) of the known languages in the bilingual/multilingual mind (e.g. Kroll et al., 2006, 2015; Rice and Tokowicz, 2020): Whenever one language of a multilingual is active, the other language(s) will also be activated to some degree and involved in the processing mechanisms.

In L3 acquisition, the CLI mechanisms based on language co-activation allow multilinguals to exploit their prior knowledge. In brief, when acquiring an L3, learners have two potential sources of CLI: the L1 and the L2. Thus, CLI in L3 acquisition may result from the influence of one known language (L1 or L2), as the primary or sole source of CLI (one-to-one transfer; De Angelis, 2007), or the combined influence of all languages known (many-to-one transfer; De Angelis, 2007). Several theoretical accounts have been put forward regarding the sources of CLI in L3 acquisition, and the cognitive economy (Puig-Mayenco et al., 2020; Rothman, 2015) in utilizing the prior knowledge in the learner’s mind. However, those models have been mostly tested with regards to morpho-syntax, not vocabulary. Essentially, one-to-one transfer accounts (Lago et al., 2021) postulate that CLI into the L3 has a single source. This is either the best developed and automatized L1 (e.g. Hermas, 2010, 2014) or the L2 (L2 Status Factor; e.g. Bardel and Falk, 2007, 2012; Hammarberg, 2001), because non-native languages have similar cognitive representations underlain by the declarative memory system (e.g. Paradis, 2009). Finally, the Typological Primacy Model (TPM; Rothman, 2011, 2015) distinguishes between CLI and transfer, postulating full transfer of linguistic representation at the initial stages of L3 acquisition. A more similar language is chosen for transfer based on the hierarchy of L3 linguistic cues (lexical, phonological, morphological, and surface-syntactic) and on general cognitive economy, the default path of least mental effort in acquiring knowledge. Similarly, many-to-one transfer models assume cumulative CLI triggered by the linguistic similarity between the L3 and prior languages (Lago et al., 2021), facilitative transfer to L3 from previously acquired languages (Cumulative Enhancement Model, CEM; Berkes and Flynn, 2012; Flynn et al., 2004), or facilitative transfer on property-by-property basis (Linguistic Proximity Model; Westergaard et al., 2017; Scalpel Model; Slabakova, 2017). The acquisition mechanisms are sensitive to property-specific variation in the target L3/Ln input, determining the potential source of facilitative CLI in the case of cross-linguistic similarity (Hopp et al., 2018; Westergaard, 2021).

None of the CLI accounts and models explicitly deals with L3 vocabulary. Still, based on them we may pose questions concerning CLI and similarity in lexical acquisition. Since the TPM and the L2 Status Factor have received much empirical support for morpho-syntax, does one-to-one CLI also prevail when acquiring L3 words? The TPM predicts CLI from the more similar language, the L2 Status Factor predicts CLI from the L2. If one-to-one CLI is indeed the case, which language plays a more important role in lexical transfer: the automatized L1, or the L2, more cognitively similar to the new L3? Alternatively, are the sources of CLI different for the acquisition of L3 morpho-syntax and vocabulary (see Hopp et al., 2018), so that the acquisition of L3 vocabulary depends cumulatively on the similarities across the multilingual lexicon? Then, if the many-to-one CLI is the case in lexical acquisition, what factors modulate the process?

III CLI in L3 vocabulary acquisition

1 Cross-linguistic similarity and language co-activation

Words in the multilingual mental lexicon are stored independently of their language membership. Thus, identification of words by multilinguals is based on language co-activation and enhanced by cross-linguistic similarity (Dijkstra and Van Heuven, 2018). Upon seeing or hearing a novel L3 word its orthographic neighbours across languages known are co-activated leading to CLI. Thus, cognates (words that share similar from and meaning across languages) and false cognates (words that share similar forms across languages) will be co-activated faster than less similar words.

Most knowledge about lexical CLI comes from research on L2 cognates. L2–L1 cognates are more learnable, are faster retrieved from memory, and more resistant to forgetting than other L2 words. They are visually recognized and reacted to faster than non-cognate words by bilinguals, a phenomenon called the ‘cognate facilitation effect’, and explained by the organization of the mental lexicon. When a cognate is visually processed, the L1 and L2 orthographic representations are co-activated, leading to the activation of a common semantic representation, which speeds up the word’s recognition. Seeing a false cognate does not lead to facilitation because the semantic representations for the L1–L2 false cognates are divergent (see Dijkstra and Van Heuven, 2018; Marecka et al., 2021). Possibly, cognates (but not false cognates) may share their orthographic representation. Thus, higher cumulative frequency of the shared orthographic representation could also contribute to the cognate advantage in processing and learning (Lemhöfer et al., 2008).

2 One-to-one and many-to-one CLI in L3 acquisition

The issue of cumulative frequency of the shared orthographic representation are particularly pertinent to cognates across three languages, as explained by the Parasitic Model (PM; Ecke, 2015; Hall and Ecke, 2003) of the multilingual lexicon. The PM, which postulates the levels of orthographic, lemmatic and semantic word representation, proposes that L3–Ln words are initially ‘parasitically’ stored in the mental lexicon: new L3 words are initially connected to the known L1 and L2 words, ‘hosts’. Any cross-linguistic similarity (formal, i.e. orthographic and phonological, and/or semantic) between the ‘host’ (L1 or L2) and a new ‘parasite’ (L3 word) helps to establish a link, because hearing or reading an L3 word should co-activate its ‘host’ words in the other languages. A new L3 cognate with the known words is initially connected with existing representations wherever similarity is detected (Ecke, 2015). Extrapolating, L3 words that are similar across more languages should have more chances to be connected and stored in the mental lexicon (compare the many-to-one transfer; De Angelis, 2007).

Apart from the PM, two further similarity-based mechanisms have been proposed for the acquisition of L3 cognates (Bartolotti and Marian, 2017). Under the scaffolding account, word learning relies on any formal similarity as an anchor to existing knowledge. Thus, one direct link to the existing ‘old’ knowledge and CLI from any previous language should be sufficient for learning benefits. Under the accumulation account, L3 words should benefit more from similarities common to all languages. A higher overall frequency of orthotactic/phonotactic similarity features and orthographic neighbour forms across L1, L2 and L3 should lead to more learning due to co-activating CLI from all languages. Under the accumulation account, in L3 acquisition cognates shared by three languages might have an advantage in learning and processing over those shared by only two languages. In brief, Bartolotti and Marian’s (2017) scaffolding account would predict one-to-one CLI and their accumulation account would predict the many-to-one CLI (De Angelis, 2007).

Only one study tested the two mechanisms, albeit for an artificial L3. Bartolotti and Marian (2017) taught English–German bilinguals written 4-letter-long pseudowords designed to draw on the similarity to L1-English (e.g. sumb), L2-German (e.g. kenf) or both English and German (e.g. baft). Bartolotti and Marian (2017) found that cross-linguistic similarity to any language improved the speed of L3 word recognition and the accuracy in production. Also, most participants reported a strategy for learning L3-word spelling by anchoring the novel word to a similarly spelled existing word in English or German to make one strong association between the L3 word and an L1 or L2 word. Bartolotti and Marian (2017) concluded that similarity to either language provided the same benefit as the cumulative overlap across both languages. Their results supported the scaffolding learning model, which predicts single-sourced (one-to-one) CLI from either language. This also aligns with the assumptions of the PM model (Ecke, 2015; Hall and Ecke, 2003), whereby new words are integrated into the existing lexical network with the least possible redundancy.

Overall, Bartolotti and Marian’s (2017) accounts are less complex and elaborate than Ecke’s (2015) PM. The PM proposes mechanisms for the levels of formal similarity, lemma (sentence use) and semantics, while Bartolotti and Marian’s (2017) accounts only propose mechanisms at the level of formal similarity. For testing hypotheses at this level it may not be possible to disentangle whether Ecke’s (2015) PM or Bartolotti and Marian’s (2017) accounts make better predictions. Still, for words learned in context (i.e. the levels of lemmatic and semantic similarity), the PM is by far more elaborate and allows for testing more hypotheses (for testing the level of semantics, see, for example, Suhonen, 2020). Further, Bartolotti and Marian’s (2017) study testing the accounts employed highly artificial L3 words. Whether the scaffolding or the accumulation accounts hold for natural cognates remains unresolved. Thus, evidence from three research strands – L3 word processing, L3 cognate guessing, and learning L3 words – will serve us to extrapolate conclusions concerning the single-sourced or cumulative nature of CLI in acquiring L3 words.

IV Empirical evidence on formal CLI in L3 acquisition

1 Processing of words cognate with L3

There are surprisingly few studies on L3 word processing, but many point to cumulative rather than single-sourced CLI effects. In an early trilingual study, Lemhöfer et al. (2004) ran a lexical decision task (deciding quickly whether the string of letters on the screen is or is not a word in a given language) on Dutch–English–German trilinguals in their L3 German. She used double L1–L3 cognates, triple L1–L2–L3 cognates, and non-cognates. The results showed a cognate facilitation effect for both types of cognates. However, the reaction times for the triple L1–L2–L3 cognates were shorter than for the double L1–L3 cognates. Importantly, the cumulative facilitation effect did not depend on whether the L2-English was activated before the study. Along similar lines, Szubko-Sitarek (2011) investigated Polish–English–German trilinguals using the lexical decision task. The results in the weakest L3-German showed that L1–L2–L3 cognates (e.g. plan, echo) were recognized faster and more accurately than the matched L1–L3 cognates (e.g. dach, pech). Next, Szubko-Sitarek (2015) investigated whether triple L1–L2–L3 cognate advantage over double L1–L3 cognates would hold when participants performed in their strongest L1. Using the same stimuli and participants she corroborated the previous finding. Interestingly, she did not find a difference between the double L1–L3 and L2–L3 cognates. Further, Poarch and Van Hell (2014) examined picture naming by trilinguals using languages with the same or different scripts. First, German–English–Dutch trilinguals (same script) immersed in their L3-Dutch or L1-German, performed a picture-naming task in L2-English and L3-Dutch. For L2-English, triple cognates (L1–L2–L3) were always named faster and more accurately than double cognates (both L1–L2 and L2–L3) and noncognates. For L3-Dutch, triple cognates were named faster than non-cognates, and all cognates surpassed non-cognates in accuracy. Next, Russian–English–German trilinguals (different script) immersed in L3-German performed the same L2 and L3 tasks. Similar to the same-script languages, triple cognates were named faster and more accurately than other words. Altogether, the results by Lemhöfer et al. (2004), Szubko-Sitarek (2011, 2015) and Poarch and Van Hell (2014) evidence processing advantages when similarity across three languages is co-activated.

Still, the facilitation effects in multilingual visual word recognition are modulated by learner individual differences. Van Hell and Dijkstra (2002) ran an L1 lexical decision on two groups of Dutch–English–French trilinguals with a low and high L3-French proficiency. They used L1 words which were L1–L2 cognates, L1–L3 cognates, or non-cognates. L1–L2 cognates, always showed cognate facilitation effects, but L1–L3 cognates only for those participants whose L3-French proficiency was high enough. Also Zhu and Mok (2020) found proficiency effects in L3 lexical decision experiments testing cognates and false cognates in L2 and L3 on a group of Cantonese–English speakers learning L3-German (beginning to intermediate). In L2-English, clear facilitation effects were obtained for L2–L3 cognates, but in L3-German no cognate effect was found. Finally, when L2–L3 cognates and false cognates were used, the cognates were responded to faster and more accurately than non-cognates and false cognates in L3-German, which might have resulted from practice effects.

Further proficiency effects come from picture-naming. González Alonso (2012) asked Polish–English–Russian trilinguals to name pictures in their L3-Russian. Some L1 and L2 stimuli were L1–L3 or L2–L3 false cognates (e.g. L1-Polish głaz – rock vs. L3-Russian глаз – eye), but not translation equivalents. ³ Before naming a picture, participants heard an L1 or L2 word. González Alonso (2012) found CLI effects of form-related L1 and L2 primes on L3 word naming, reflected in enhanced retrieval speeds, especially in less proficient participants. Also Poarch and Van Hell (2012) found some proficiency effects when testing picture naming in primary school children (L1-Germans learning L2-English, balanced German–English bilinguals, and trilinguals). For L2-English learners, cognate facilitation was limited to the L2. For the balanced German–English bilinguals and trilinguals, the cognate facilitation also was obtained in the dominant L1-German. The findings suggest that the degree of language co-activation and cognate facilitation in the multilingual lexicon depends on the L2 and L3 proficiency of the participants.

Finally, a series of studies by Lijewska and colleagues demonstrated factors modulating cognate facilitation in sentences. Lijewska and Chmiel (2015) made proficient Polish–German–English trilinguals verbally produce L1 and L2 translations of L3-English target words (L2–L3 cognates and non-cognates) presented in the context of high-constraining (highly predictive) and low-constraining (neutral) sentences. Significant cognate facilitation effect was found in L3–L1 translation, although the study employed only L2–L3 cognates and no L1–L3 cognates. Still, the L3–L2 translation resulted in lower accuracy rates. The authors interpreted their findings as resulting from acquiring L3-English via the L1, rather than L2-German: At the formal level, L3 words were strongly connected to their L1 equivalents, but only weakly to the L2 ones. Lijewska and Błaszkowska (2021) extended the previous study and used two groups: Polish–German–English trilinguals (who learned L3-English and L2-German via Polish) and German–English bilinguals (who learnt L2-English via German). Contrary to studies based on identical cognates, they used non-identical L2–L3 and L1–L3 cognates embedded in sentences. The study did not corroborate results concerning learning experience, but revealed proficiency effects in L3 processing. Significant cognate effects were recorded in translation speed obtained from advanced bilinguals, but not from trilinguals whose L2-German and L3-English were intermediate. Finally, Lijewska (2022) tested the translation of non-identical double cognates (taken from Lijewska and Błaszkowska, 2021) versus triple cognates embedded in sentences, high- or low-constraining. The study combined behavioural measures with eye-tracking, but found no cognate facilitation effects for any measures and kinds of cognates (triple and double) during natural reading. The author believes it was due to insufficient cross-language overlap, or task specificity.

Concluding, this line of research shows nonselective access and activation of words in the trilingual memory system when processing the L3, and demonstrates stronger facilitation effects for the triple than double cognates (Lemhöfer et al., 2004; Poarch and Van Hell, 2014; Szubko-Sitarek, 2011, 2015). This indicates cumulative similarity effects in L3 acquisition (despite contradictory results by Lijewska, 2022). A threshold level of proficiency is required before weaker language effects become noticeable (Lijewska and Błaszkowska, 2021; Poarch and Van Hell, 2012; Van Hell and Dijkstra, 2002; Zhu and Mok, 2020). It remains unknown what L2 and L3 proficiency threshold levels are needed, especially that many studies reported above employed self-ratings (apart from Lijewska, 2022; Lijewska and Błaszkowska, 2021; Lijewska and Chmiel, 2015; Poarch and Van Hell, 2014).

2 Receptive multilingualism and cognate guessing tasks

Much evidence for the role of proficiency in lexical CLI comes from receptive multilingualism research, based on the observation that related languages can be mutually intelligible. Receptive multilingualism involves languages and dialects which are typologically close, but can also take place in situations where an L3 is understood thanks to the knowledge of an L2 (e.g. speakers of L1-Estonian trying to understand L3-Ukrainian thanks to the knowledge of L2-Russian; Gooskens, 2019). Receptive multilingualism research investigates how much can be translated from an unknown L3–Ln thanks to cross-linguistic similarities. For instance, Mieszkowska and Otwinowska (2015) asked L1-Polish participants proficient in L2-English with various L3–Ln languages to orally translate into L2-English a text in Ln-Danish, an unknown language (think-aloud procedure). As demonstrated, both typological and participant-related variables were at play when inferencing. Success in translating cognates and making sense of the text depended on the similarity between Danish and the known languages (typology), but also correlated with participants’ proficiency in the L3–Ln languages. This points to the role of cumulative language learning experience and stronger co-activation of the better-known languages beyond the L2 in the inferencing processes.

Apart from think-aloud, receptive multilingualism research typically utilizes word translation lists (Gooskens, 2019) called cognate guessing tasks (Berthele, 2011; Vanhove and Berthele, 2015a, 2015b). Multilingual participants translate into their L1 unknown L3 words which are cognate and vary in their orthographic overlap with the L1 and L2 forms. Because stimuli in such tasks are devoid of any contextual cues, participants must engage in cross-linguistic inferencing and draw on their previous L1–L2 linguistic and meta-linguistic knowledge when performing the task. Thus, cognate guessing affords insights into factors modulating the success in inferencing (Vanhove and Berthele, 2015a), especially individual learner-related factors influencing facilitative CLI: cumulative language learning experience, proficiency, and fluid intelligence.

For example, Berthele (2011) examined the understanding of Dutch and Swedish written words by speakers of L1-German and German dialects, and Romansch and Romanian written words by L1-French speakers of Spanish and Italian. Participants performed guessing tasks with the words presented in context and out of context. Berthele (2011) demonstrated that the more multilingual a participant was, the more likely he/she was to correctly infer the meaning of cognates and non-cognates in an unknown language. Additionally, multilinguals highly proficient in at least two languages did better in the guessing tasks than multilinguals of lower proficiency. Also Vanhove and Berthele (2015b) tested participants with various language combinations. They found that learners with larger linguistic repertoires were better at understanding words in an unknown L3. Vanhove and Berthele (2015a, 2015b) tested a large number of L1-German participants aged 10–86 years who were asked to translate written and spoken L3-Swedish words into German. Some words had translation-equivalent cognates (orthographic neighbours) in L1-German and/or the L2 languages (French and English), common to most participants. All participants did a battery of cognitive and linguistic tests. Crucially, Vanhove and Berthele (2015a) demonstrated that when guessing written L3 cognates, participants with more linguistic experience performed better on L3–L1 words with lesser orthographic overlap. Also, higher fluid intelligence helped when coping with the less overlapping L3–L1 cognates. When re-examining the same data, Vanhove and Berthele (2015b) demonstrated clear age-related effects in cognate guessing. For written words, inferencing improved throughout adulthood, but for spoken stimuli, it started to decrease from the age of 50 years. Overall, both cross-linguistic similarity and participant-related variables such as proficiency, fluid intelligence, and age can modulate L3 word guessing (Vanhove and Berthele, 2017, 2015a, 2015b).

Receptive multilingualism research shows that multilinguals can use their previous languages when guessing, often in a cumulative manner (Mieszkowska and Otwinowska, 2015; Vanhove and Berthele, 2015a, 2015b). It makes a strong point about using item-related and participant-related characteristics in studying CLI in L3 acquisition. Its main drawbacks are focusing on typologically close languages, where cross-linguistic similarities are expected, and inferencing rather than lexical gains.

3 Experimental studies

Few experimental studies have investigated the acquisition of L3 lexis, and most did not address questions concerning directions and amount of CLI (see Hirosh and Degani, 2018). If they did, the focus was not on cognateness in CLI (Suhonen, 2020), the main area of interest in this review. Thus, conclusions on the role and direction of CLI in L3 vocabulary learning can only be indirect. The studies reviewed below are divided into those using pseudowords, and those using natural L3 words.

a Psycholinguistic experiments with L3 pseudowords

All studies discussed below concerned grapheme–phoneme or grapheme–grapheme mappings in L3 acquisition. For example, Kaushanskaya and Marian (2009) examined the adults’ ability to resolve cross-linguistic inconsistencies in orthography-to-phonology mappings when learning new L3 words. English monolinguals and English–Spanish bilinguals learned L3 pseudowords that overlapped with English orthographically, but not phonologically. Participants heard the L3 word and saw its written form and English translation on the screen, so L1 orthographic information interfered with L3 word encoding. The bilinguals outperformed English monolinguals on the L3 word-learning task in both immediate and delayed testing. The authors claim that L2 knowledge reduced interference effects associated with L1 letter-to-phoneme mappings, ‘protecting’ bilinguals from negative CLI. Further, Escudero et al. (2013) examined how L2-English and L3-Dutch proficiency affects learning L3 Dutch-like pseudowords. L1-Spanish speakers with different proficiencies in L2-English and L3-Dutch were trained and tested on matching L3 pseudowords to pictures, some orthotactically problematic for L1-Spanish learners. Spanish learners’ proficiency in L2-English predicted their accuracy on the L3 learning task. Thus, learning an L2 with a larger vowel inventory than the L1 proved beneficial for L3 word learning, possibly due to L2–L3 CLI effects. Also two experiments by Bartolotti and Marian (2019, 2017; see Section III.2 above) have pointed to benefits of prior language knowledge in L3 word learning. Bartolotti and Marian (2019) checked how learners manage interference between the L1, L2 and L3. They taught Spanish–English bilinguals L3 pseudowords, designed to conflict with English and Spanish letter–sound mappings. The interference from existing languages was higher for L3 words similar to L1 or L2 words. Learners experienced competition from their L1 and L2 because hearing spoken L3 words activated orthography in all three languages, even in the absence of phonological overlap.

Overall, these studies showed L2 proficiency effects in L3 word learning due to CLI from the L2 and L1–L2 co-activation when learning L3 words. They were not designed to show the direction of CLI in L3 learning, but demonstrated that the novel L3 word knowledge is anchored in both L1 and L2 knowledge. However, the experiments used L3 pseudowords, so their ecological validity may be limited.

b Experiments with natural L3 words

The research below focused on natural languages, which sheds more light on CLI. Two studies hint that cross-linguistic similarity from L2 and L1 differently affect L3 word learning. Mulík et al. (2019) used paired-associate tasks to investigate whether Spanish–English bilinguals activated lexical knowledge from their L1-Spanish and L2-English relative to L2 proficiency when learning L3-Slovak words. L1-Spanish participants with high or low proficiency in L2-English learned L3-Slovak words presented auditorily with their written L1-Spanish translations. Some L3 words overlapped phonologically with either L2-English or L1-Spanish (homophones). Interestingly, the effects of L3–L1 and L3–L2 phonological similarity were comparable, although no L2-English was used in the task. However, more proficient L2 participants showed higher advantages resulting from L2 similarity. This suggests that the degree of L2 activation is relative to learners’ L2 proficiency, which corroborates earlier results for pseudowords (Escudero et al., 2013). Next, utilizing the same stimuli as Mulík et al. (2019) and event-related potentials (ERPs), Mulík and Carrasco-Ortiz (2023) investigated the neurocognitive mechanisms in using phonological information from L1-Spanish and L2-English in L3-Slovak lexical learning in Spanish–English bilinguals. Behavioural results did not reveal differences between bilinguals’ L1 and L2. But after only three days of training, the ERPs evidenced opposite N400 effects in response to L1-Spanish and L2-English interlingual homophones (respectively: more inhibited, vs. less inhibited language interference in word recognition). The authors attributed the differences to divergent neurocognitive mechanisms in bilinguals’ learning via the L1 and the L2.

Experiments carried out over 20 years have always used participants’ L1 as the language of L3 instruction (see Hirosh and Degani, 2018). Two studies have recently explored how the bilingual advantage in L3 lexical learning depended on different languages of instruction (L1 or L2) and regulatory skills in the languages. Bogulski et al. (2019) tested learning L3-Dutch words (cognates, false cognates and non-cognates) via English translations, where English, a language similar to Dutch, was either participants’ L1 or L2. They examined English monolinguals and bilinguals with various language configurations: English–Spanish, Spanish–English and Chinese–English bilinguals. The results showed that English–Spanish bilinguals were more accurate than English monolinguals when learning L3-Dutch words through their L1-English, replicating the multilingual advantage from previous studies. The advantage in learning L3-Dutch words disappeared when Spanish–English and Chinese–English bilinguals were compared to the same monolinguals. Possibly, results could differ for more balanced bilinguals. Still, Bogulski et al. (2019) concluded that the multilingual advantage in L3 word learning occurs only when the L3 is acquired via the native L1 or the dominant language, and when bilinguals engage in language regulation processes, best developed for the L1.

Hirosh and Degani (2021) also examined whether the success in learning L3 words depended on the language of instruction, explicitly focusing on written cross-linguistic similarity and CLI. They presented adult L1-Hebrew learners of L2-English with words in L3 German: L2–L3 cognates; L2–L3 false cognates, and L3 non-cognate controls. Half of the participants learned the L3 words through their L1-Hebrew translations, and half through their less fluent L2-English translations. As expected, L2–L3 cognates were learned better than other words in both languages of instruction, and L2–L3 false cognates were learned better than controls when participants learned through their L2-English, which points to language co-activation and cross-linguistic similarity at play. Again, better L3 learning through the less similar L1 was evidenced, especially for the non-cognate control items. This corroborates the results by Bogulski et al. (2019) and also resonates with the results by Lijewska and Chmiel’s (2015) processing study, whereby the L2 was also the less activated language. Hirosh and Degani (2021) explain that learning L3 vocabulary through L1 translations might have unlocked more cognitive resources and more experience in language regulation, relative to learning through L2 translations. Together, the results revealed the modulating effects of the language of instruction.

Finally, Salomé et al. (2022) evidenced some age effects in L3 word learning. They tested schoolchildren in a longitudinal experiment involving two paired-associate learning sessions and four testing sessions. French monolinguals and French–German bilinguals learned L3-English, half of which were L2–L3 non-identical cognates (e.g. L2-German Schwan – swan). Their L3 knowledge was tested immediately after the session and one week later with an orthographic judgement task, and visual and auditory recognition tests. The French–German bilinguals outperformed the monolingual group only in the orthographic judgement tasks for L2–L3 cognates. Salomé et al. (2022) found no other evidence of the L2–L3 cross-linguistic similarities in the L3 lexical learning. They attributed the lack of the general advantage for L3 cognates to learners’ insufficient L2 proficiency, or their low cognate awareness due to age.

The role of cognate awareness, i.e. conscious cross-linguistic strategies in L3 lexical acquisition, was explored in several studies. Molnár (2010) tested the L2 Status Factor (see Section II) and the role of awareness of cross-linguistic similarity in learning L3-English words on teenage L1-Hungarian and L2-Romanian students. Molnár (2010) randomly divided the students into experimental and control groups and measured their L3-English vocabulary knowledge with a test where half the words were L2–L3 cognates. Before testing, the experimental group was instructed about orthographic and morphological correspondences between L2-Romanian and L3-English words. This group outperformed the control group on the entire test, and on the cognates. This points to the importance of conscious strategic reliance on L2–L3 similarities.

Learners’ similarity-based strategies were indirectly investigated in two other experiments. Hall et al. (2009) longitudinally tested L1-Spanish L2-English speakers learning either L3-German or L3-French to check whether participants assumed syntactic frames for L3 cognate verbs in agreement with their L1 and L2 frames. Participants learned L3-German verbs (cognate with L1-Spanish, L2-English, or non-cognate) or L3-French verbs (cognate with L1-Spanish, L2-English, or non-cognate) and were exposed to cross-linguistic similarity across three languages. L3-French verbs yielded stronger effects for the L1-Spanish grammatical frames than the L3-German verbs and their L2-English frames. In delayed testing, the effect was maintained for French. Learners assumed that L3 verbs shared syntactic frames with cognates in the typologically closer languages (Spanish–French), although they did not report any conscious cognate-based strategies. Hall et al. (2009) concluded that form similarity and typological proximity jointly affect learners’ assumptions about the grammatical properties of L3 words, which shows that L3 learner’s judgements might influence the CLI effects.

Conscious strategic knowledge in L3 word learning was examined by Vanhove (2016), who tested whether exposure to cross-linguistic similarity was enough to learn some L3–L1 graphemic correspondence rules. L1-German L2-English speakers were presented with cognates in L3-Dutch. Some participants were first trained on translating L3-Dutch words bearing systematic orthographic similarities to their L1-German counterparts, presented with one of two L3–L1 systematic correspondence rules and given explicit feedback on the correctness of their translations. Next, all participants had to translate L3-Dutch words, part of which contained both L3–L1 grapheme correspondences to German. Those participants who had been given feedback in the training were particularly successful in translation and benefitted from L3–L1 form similarity. Interestingly, less than 10% of participants could verbalize the correspondence rules they relied on. Overall, this study suggests that learners start benefiting from shared conceptual representation as soon as they notice lexical similarity between two languages, and that explicit training of cross-linguistic regularities facilitates learning. Both studies (Hall et al., 2009; Vanhove, 2016) show a limited role of explicit metalinguistic or strategic knowledge at the early stages of L3 word learning.

V Discussion

This overview focused on the quantitative empirical evidence concerning CLI and cross-linguistic similarity in L3 lexical acquisition. Both psycholinguistic and applied linguistic research was reviewed to bridge the gap between the fields, which overlap in scope, but often provide disparate lines of research. Since the mechanism of CLI (Sharwood Smith, 2021) was explained as resulting from co-activating previous L1 and L2 knowledge and its interaction with L3 knowledge in the multilingual user’s mind, throughout the article I sought evidence for the amount of previous language activation. I concentrated on how much CLI is present in L3 lexical learning and whether cross-linguistic similarity cumulated across languages brings about differences in learning and processing L3 words. To this end, I first discussed some accounts of L3 acquisition dividing them into one-to-one and many-to-one transfer models (De Angelis, 2007; Lago et al., 2021) based on whether they assumed single-source or cumulative CLI from previously known languages. Although these accounts make strong claims about utilizing L1–L2 morpho-syntactic knowledge, most of them do not concern L3 vocabulary acquisition (Puig-Mayenco et al., 2020). Still, based on them I posed questions concerning CLI in L3 lexical acquisition. Because two established models supported by much empirical evidence (L2 Status Factor, Bardel and Falk, 2007, 2012; TPM, Rothman, 2011, 2015) postulate single-sourced one-to-one CLI, I asked whether evidence for such single-sourced CLI prevailed for L3 lexical acquisition. I also asked which language would play a more important role in lexical CLI: the automatized (native) L1, or the (non-native) L2, more cognitively similar to the new L3.

1 Does one-to-one CLI prevail in L3 lexical acquisition?

I discussed three accounts of novel L3 words learning. According to the Parasitic Model (PM; Ecke, 2015), any formal similarity (orthographic, phonological) to any previous language should be enough to anchor the new L3 knowledge in the existing L1 and/or L2 representations. Based on the PM we could assume that triple cognates (similar across three languages) should be easiest to learn as anchored in both L1 and L2 formal representations (many-to-one CLI; De Angelis, 2007). Two more learning accounts were proposed by Bartolotti and Marian (2017). Under their scaffolding account, any formal similarity (L1 or L2) to an existing word would suffice to facilitate learning (one-to-one CLI); under their accumulation account, learning L3 words should be facilitated by similarities common to all languages (many-to-one CLI; De Angelis, 2007). Bartolotti and Marian (2017) found that single-sourced similarity of the L3 form to any L1 or L2 word provided a strong-enough anchor for L3 word learning, which supported their single-sourced, scaffolding account.

However, the empirical results for the role of CLI in L3 word processing, L3 cognate guessing, and L3 words learning do not corroborate single-sourced CLI. Several lexical decision experiments evidenced a processing advantage in visual word recognition when similarity across three languages was co-activated (Lemhöfer et al., 2004; Szubko-Sitarek, 2011, 2015). In fact, all word processing experiments discussed indicated co-activation across three languages, although modulated by the proficiency level in L2 and L3 (Lijewska and Błaszkowska, 2021; Poarch and Van Hell, 2012; Van Hell and Dijkstra, 2002; Zhu and Mok, 2020). Also, receptive multilingualism research and cognate guessing tasks (Berthele, 2011; Mieszkowska and Otwinowska, 2015; Vanhove and Berthele, 2015a, 2015b, 2017) showed evidence for cumulative effects of prior language knowledge. However, they also demonstrated that higher proficiency in the languages known improved learners’ abilities to infer L3 word meanings based on similarity.

Taken together, this research on natural words indicates that CLI from previous languages may operate in a cumulative, many-to-one manner, although moderated by language proficiency. This stands in contrast to Bartolotti and Marian’s (2017) results for pseudowords, but also to the prominent TPM model (Rothman, 2011, 2015), postulating one-to-one transfer based on typological similarity. Possibly, as suggested by Hopp et al. (2018), the sources of CLI can be different for the acquisition of L3 morpho-syntax and vocabulary, but more research focused on this problem is necessary.

2 Do the L1 and L2 have a special status in L3 lexical CLI?

The reviewed research evidences co-activation of all languages during experiments: Even in tasks that did not require L2 use it was active. However, the lexical decision tasks evidenced no differences in processing double L1–L3 and L2–L3 cognates (Szubko-Sitarek, 2011, 2015), which could challenge the claim that language status (native L1 vs. non-native L2) plays a role in CLI to L3. Still, in learning experiments utilizing conflicting L1–L3 grapheme-to-grapheme and grapheme-to-phoneme mappings, L2 knowledge blocked the negative CLI from the L1 (Kaushanskaya and Marian, 2009), or from the L1 and L2 (Bartolotti and Marian, 2019). Also, the L2 effects in the experiments using wordlike pseudowords were modulated by the L2 proficiency of the learners (Escudero et al., 2013), similarly to experiments where real L3 words were used (Mulík et al., 2019; Salomé et al., 2022). Probably, higher L2 proficiency ensured a stronger co-activation of the L2, parallel to L1 activation, which the prerequisite for an increased L2 impact on word processing. Further, participants advanced in their L2 benefitted from conscious L2–L3 strategy training (Molnár, 2010), even though the role of explicit metalinguistic knowledge in L3 word acquisition might be lesser than expected (Hall et al., 2009).

Also, learning through the L1 brings better results than through the less activated L2. In research by Bogulski et al. (2019) and Hirosh and Degani (2021), participants who learned the L3 via the L2 were disadvantaged relative to those learning via their L1, even when learning L2–L3 cognates. These results resonate with Lijewska and Chmiel’s (2015) translation study, where the L2 was the less activated language, although the stimuli bore L2–L3 similarity.

Overall, these studies show is that even when the task does not directly involve one of the languages (L1 or L2), all participant’s languages are active, and provide potential sources of CLI, modulated by learners’ proficiency. Higher L2 proficiency enhances cognate effects, but lower L2 proficiency (and, what follows, lower activation), may hinder learning despite L2–L3 cross-linguistic similarity. Thus, it seems that none of the languages is a single source for CLI in L3 lexical acquisition, although the L1 effects are more prominent. The exact impact of the languages is yet to be explored, especially that differential neurocognitive mechanisms have recently been proposed for L3 lexical learning via the L1 and L2 (Mulík and Carrasco-Ortiz, 2023).

VI Conclusions and future research directions

Research on L3 word acquisition is still limited. This review brought together applied linguistic and psycholinguistic perspectives on CLI in L3 vocabulary acquisition, all very diverse in aims, scope, ecological validity and variable control. I have demonstrated that triple L1–L2–L3 similarity leads to co-activating all languages, provided that the L2 and L3 proficiency is high enough. Thus, it can be extrapolated that (natural) cognates shared by three languages should have an advantage in learning over double cognates, which is yet to be tested.

However, for future research we must learn some lessons from the studies reviewed above. First, many psycholinguistic experiments were limited to meticulously controlling for language-related (typological) or item-related (stimuli) variables. They assumed that CLI depended on typological similarity and the degree of cognate overlap, and that learners were relatively homogenous in taking advantage of similarity, which is not the case. Participants considerably differ in their abilities to detect cognateness, as demonstrated by applied research into cognate guessing (e.g. Berthele, 2011; Vanhove and Berthele, 2015a, 2015b, 2017), as well as those experiments which controlled for individual differences within and across groups (Bogulski et al., 2019; Escudero et al., 2013; González Alonso, 2012; Mulík et al., 2019; Poarch and Van Hell, 2012, 2014; Salomé et al., 2022). They showed that cross-linguistic similarity at the word level interacts with learner characteristics, and made a strong point for taking into account both variable types in further L3 acquisition research.

Since learners differ (in intelligence, learning aptitude, learning strategies, abilities to make associations, and metalinguistic awareness of cognateness), participant-related variables should be included in future study designs, and their interactions with item-related features should be tested. To measure the amount of CLI in L3 acquisition one must at least measure proficiency in the different languages, and possibly, activate those languages in the learning tasks (developing new tools may be necessary). Overall, the research presented above and recent theoretical stances concerning multilingualism research (e.g. Dijkstra and Van Heuven, 2018; Ecke, 2015; Festman, 2018, 2021) point to variables to be considered in future L3 vocabulary studies. These can be grouped into: (1) language-related (e.g. typological distance, degree of contact between the languages), (2) item-related (e.g. degree of L1–L2–L3 similarity, frequency, abstractness vs. concreteness of the item), and (3) participant-related factors (proficiency in each language, degree of exposure to input, metalinguistic awareness, language aptitude and intelligence, and executive control including working memory).

Further, a good question is whether guessing involved in cognate learning should be controlled for (e.g. by confidence ratings) and excluded from statistical models, or maybe taken into account as part and parcel of the learning process. As shown by receptive multilingualism research, inferencing about cognateness is an important individual ability (Berthele, 2011; Vanhove and Berthele, 2015b, 2017). Thus, variables related to guessing (e.g. learner’s propensity to guess, number of items guessed) could also be used as covariates in modelling L3 word knowledge.

Finally, future L3 vocabulary projects should not be restricted to one context, most preferably laboratories with strictly controlled conditions, because accounting for learners’ affective factors seems necessary. In highly artificial tasks (e.g. Bartolotti and Marian, 2017, 2019; Kaushanskaya and Marian, 2009) students’ motivation to learn may be different from real classrooms, which may limit the results’ generalizability. It is thus necessary to combine one-off laboratory experiments with less controlled, but more ecologically valid learning paradigms (see, for example, the doctoral project by Suhonen, 2020) and longitudinal designs (e.g. Salomé et al., 2022). More comprehensive projects will provide answers to some interesting questions. For instance, are triple cognates learned better than double cognates? Are they known better by learners at different L3 proficiency levels? Future projects can also combine tasks measuring reactions to stimuli (e.g. lexical decision, picture naming, priming) with traditional tests taken at a slower pace and allowing for the use of declarative and metalinguistic knowledge. Data from complex paradigms, diverse populations and different language combinations will serve to paint a more fine-grained picture of the L3 word learning, the amount of CLI from previous languages, and factors modulating the process.