The Impact of Lexical Co-activation through Cognates on L2 Rule Learning

The Impact of Lexical Co-activation through Cognates on L2 Rule Learning Submitted by Noèlia Sanahuja Cobacho In partial fulfilment of the requirements for the Master in Theoretical and Experimental Linguistics Universidad del País Vasco/Euskal Herriko Unibertsitatea Vitoria-Gasteiz, June 3rd, 2019 Supervisor: Prof. Kepa Erdozia ACKNOWLEDGEMENTS I would like to express my sincere gratitude to my supervisor, Kepa Erdozia. Your guidance and friendship throughout this research have turned this experience into an unforgettable one. I would also like to thank A. Dieguez, I. San Martín and A. IsasiIsasmendi for kindly agreeing to record the experimental items used in this research. Finally, I have to thank my family for supporting me in my decision to enrol in this project. I owe you everything. i TABLE OF CONTENTS Index of figures and tables ............................................................................................. ii Abstract ........................................................................................................................... 1 1. Introduction ................................................................................................................ 2 1.1. L1 and L2 acquisition of vocabulary and grammar ............................................... 2 1.1.1. Evidence for multiple learning mechanisms ................................................... 2 1.1.2. Proponents of a single statistical learning mechanism .................................... 4 1.2. The interaction between the acquisition of vocabulary and grammar ................... 5 1.2.1. Friederici et al.’s (2002) study ........................................................................ 6 1.2.2. Walker et al.’s (2017) study ............................................................................ 7 1.3. Cognates and lexical access ................................................................................... 9 1.3.1. The cognate facilitation effect (Dijkstra et al., 1999) ..................................... 9 1.3.2. The Inhibitory Control model (Green 1998) and the Language-specific model (Costa et al., 1999) .................................................................................................. 10 1.4. The impact of cognates on syntax learning and processing ................................. 12 1.4.1. Hopp’s (2017) study ...................................................................................... 13 1.4.2. Bastarrika & Davidson’s (2017) study .......................................................... 14 2. The present study ..................................................................................................... 15 2.1. Research question, hypothesis and predictions .................................................... 15 2.2. Some relevant properties of Basque grammar ..................................................... 16 2.3. Method ................................................................................................................. 16 2.3.1. Participants .................................................................................................... 17 2.3.2. Materials ........................................................................................................ 17 2.3.3. Procedure ....................................................................................................... 20 2.3.4. Data analysis ................................................................................................. 27 3. Results ........................................................................................................................ 28 3.1. Performance on LexTALE and the digit span test ............................................... 28 ii 3.2. Performance on pre-test trials .............................................................................. 28 3.3. Vocabulary learning ............................................................................................. 30 3.4. Performance on test trials .................................................................................... 31 3.5. Performance on post-test trials ............................................................................ 32 3.6. Comparison of performance across the pre-test, the test and the post-test .......... 33 4. Discussion .................................................................................................................. 39 4.1. Rule learning through cognates ........................................................................... 39 4.2. Explicit vocabulary learning ................................................................................ 42 4.3. Processing canonical and non-canonical word orders ......................................... 43 5. Open issues for future research ............................................................................... 44 6. Conclusions ............................................................................................................... 47 References...................................................................................................................... 48 Appendices .................................................................................................................... 53 Appendix A ................................................................................................................. 53 Appendix B ................................................................................................................. 54 Appendix C ................................................................................................................. 55 Appendix D ................................................................................................................. 70 INDEX OF FIGURES AND TABLES Figure 1. Board game used for language acquisition (Friederici et al., 2002) ................. 7 Figure 2. Screenshot of the scenes displayed during the learning task (Walker et al., 2017) ......................................................................................................... 8 Figure 3. (A) Switching performance of native speakers of Spanish low-proficient learners of Catalan. (B) Switching performance of Spanish–Catalan highly proficient bilinguals in their two dominant languages (Costa & Santesteban, 2004). .................... 10 Figure 4. Picture for the sentences Aktoreak medikua pintatu du (SOV) and Medikua aktoreak pintatu du (OSV) ............................................................................................. 20 iii Figure 5. (A) Picture displayed for the noun margolari ‘painter’. (B) Picture displayed for the verb aukeratu ‘to choose’. .................................................................................. 22 Figure 6. Picture displayed in one of the practice trials preceding the test. ................... 24 Figure 7. Example of a picture displayed in the test. ..................................................... 25 Figure 8. Mean correct responses of the cognate and the non-cognate group across the three experimental blocks ............................................................................................... 34 Figure 9. Mean reaction times (ms) of the cognate and the non-cognate group across the three experimental blocks ............................................................................................... 38 Table 1. Example experimental materials. All four sentences mean ‘The actor has painted the doctor’.. ..................................................................................................................... 19 Table 2. Descriptive statistics for the accuracy of the cognate and the non-cognate group across the experimental blocks.. ..................................................................................... 34 Table 3. Descriptive statistics for the reaction times of the cognate and the non-cognate group across the experimental blocks ............................................................................. 37 Table 4. Vocabulary used in the two versions of the artificial language. Cognates first row, non-cognates second row ....................................................................................... 53 Table 5. Means, Standard Deviations (in brackets) and t-tests for length (number of letters), orthographic and auditory overlap (Levenshtein’s Distance) and frequency per million for the lexical items of the test ........................................................................... 54 Table 6. Means, Standard Deviations (in brackets) and t-tests for length (number of letters), orthographic and auditory overlap (Levenshtein’s Distance) and frequency per million for the lexical items of the post-test ................................................................... 54 1 Abstract The cognate facilitation effect (Dijkstra et al., 1999) makes cognates easier to process than non-cognates, and this advantage has been found to ease lexical and syntactic processing (e.g. Costa et al., 2000; Hopp, 2017). To my knowledge, there is not currently any study which assesses whether cognates ease rule learning. I developed an artificial language which allowed me to test the hypothesis that cross-linguistic activation of the lexicon through cognates facilitates the learning of an L2 grammatical rule. I created two versions of the artificial language, the difference between the two being the cognate (Spanish-Basque) or non-cognate status of their lexical items (n= 30). In this language, sentences responded to either SOV or OSV word orders. The grammatical rule to be learnt described the subject-object assignment pattern: subjects were marked with an -ak morpheme and objects with an -a. Forty native speakers of Spanish (20 for each version of the language) were explicitly taught the cognate/non-cognate vocabulary and the grammatical rule. Later, learning was assessed by means of a sentence-picture matching paradigm. In a post-test, all participants listened to sentences made up of previously unheard cognates to test the hypothesis that the cognate facilitation effect would improve non-cognate learners’ performance. Reaction times and accuracy rates were measured. I found that rule learning was attained when the vocabulary of the language was both cognate and non-cognate with the L1, even if the magnitude of the learning was greater for the cognate group of participants. This finding could be explained by claiming that retrieving non-cognates was very costly, while the co-activation of the L1 through cognates made the retrieval of these words virtually cost-free. As a result, cognate learners disposed of plenty of resources to learn the grammatical rule, but non-cognate learners did not. This result was further corroborated in the post-test, when participants of the noncognate version of the language significantly improved in their learning of the rule. Keywords: cognates, rule learning, artificial language, Basque The impact of lexical co-activation through cognates on L2 rule learning 2 1. Introduction Language acquisition is a complex process which requires infants to acquire the phonemes of the language, parse the speech stream into words, associate lexical items with meaning and extract and generalize grammatical rules (de Diego-Balaguer & LopezBarroso, 2010). Yet, babies are known to acquire their native language in a natural and apparently effortless way. Learning a second language in adulthood, on the other hand, appears to be way more complicated. When learning a new language, adults must face similar challenges. They must learn new words and how to use different syntactic structures. Just as babies, speakers must associate a large set of lexical items with their appropriate semantic content, and they must learn to pronounce them according to the phonemes of the second language (L2), which, on many occasions, differ from those of the first language (L1). Additionally, learners must usually assimilate syntactic structures of the target language which are not present in their native language and which are, therefore, more costly to learn and process (Weber-Fox & Neville, 1996). In this study I examined how second language learning could be eased and, more precisely, whether the learning of a grammatical rule could be facilitated by the crosslinguistic activation of the L1 lexicon through cognates. To do this, two versions of an artificial language drawing on Basque were designed. Each version counted on just one grammatical rule and a set of thirty lexical items, either cognates or non-cognates with Spanish. In the first subsection of the introduction, I will review some of the most relevant studies tackling word and rule acquisition and I will outline how these relate to the debate around single or multiple learning mechanisms for vocabulary and grammar. Then, some considerations will be done on the special status of cognates and their relation to lexical access. Finally, the potential impact of cognates on grammar learning will be addressed. 1.1. L1 and L2 acquisition of vocabulary and grammar 1.1.1. Evidence for multiple learning mechanisms Since this study examines the relationship between the learning of vocabulary and grammar, I found pertinent to start the literature review by commenting on some of the several studies which have focused on word and rule acquisition in a second language. Given the complexity of natural languages, many researchers have opted to study language acquisition by means of artificial languages which can be learned in the The impact of lexical co-activation through cognates on L2 rule learning 3 laboratory. Both adults and infants have been shown to be capable of calculating the transitional probabilities between phonemes, adjacent and nonadjacent syllables for the acquisition of phonotactics, word segmentation and morphosyntactic rules, respectively (de Diego-Balaguer & Lopez-Barroso, 2010). One of the key studies supporting the claim that infants extract words from fluent speech on the basis of transitional probabilities was the one carried out by Saffran et al. in 1996. In their work, the authors tested 24 eight-month-old babies on their ability to segment an artificial speech stream into words. The artificial language was made up of trisyllabic nonsense items repeated in a random order (e.g. bidakupadotigolabubidaku…), and the only cues to word segmentation were the transitional probabilities between syllables, which were 1.0. within words but 0.33 between words. The procedure consisted of a 2minute familiarization phase followed by the presentation of two different stimuli: 1) words contained in the speech stream and 2) nonwords that were similar to words in the artificial language but did not belong to it. While the stimuli were presented, babies’ visual fixations on a blinking light were measured. The duration of these fixations was found to be significantly larger for nonwords than for words, reflecting that babies were able to discriminate between familiar and novel sounds. In other words, eight-month old babies were capable of segmenting words from fluent speech based on the transitional probabilities between speech sounds. A post-test further showed that they learned the words to a high degree of specificity, since they attended less to “words” than to “partwords”, i.e., items which consisted of the final syllable of a word plus the first two syllables of another word. Following these findings, the question arose as to whether this statistical computation could also be responsible for grammar learning or if, by contrast, words and rules were learned through different mechanisms. Against this debate, the study by Peña et al. (2002) was crucial in tipping the balance in favour of distinct mechanisms for simultaneous word and rule-learning. Peña and colleagues’ research aimed to determine whether adult speakers could segment a speech stream and extract its underlying grammatical rules by computing nonadjacent transition probabilities. To do this, fourteen adult French speakers were tested on a trisyllabic artificial language called AXC, since for every item, A predicted C (e.g. [puliki], [puRaki], [pufoki]). After undergoing a 10-minute long familiarization phase, participants were presented with both words (AiXCi) and “part The impact of lexical co-activation through cognates on L2 rule learning 4 words” (CAX or XCA) and were asked to judge which item resembled more a word within the sequence of sounds previously heard. Participants chose substantially more the words than the part words, indicating that adults were also capable of segmenting the speech stream into words by computing transitional probabilities. A second experiment was carried out to determine whether, apart from identifying the words in the speech stream, participants were also aware of the structural generalization underlying this language. Interestingly, when asked to decide whether a “part word” or a “rule word” —an item not present in the speech stream but congruent with the generalisation— resembled more a word of the artificial language, participants did not prefer rule words to part words. Yet, when in a third experiment segmentation cues —subliminal gaps of 25ms— were added after each word, speakers were able to capture the AXC generalisation. That is, they judged “rule words” to be more similar to the items previously heard than “part words”. According to the authors, these silences would suppose the introduction of a minimal prosody into the speech stream which, in turn, would ease rule learning. Thus, even if a continuous speech stream triggered statistical computations, it was the insertion of minor pauses that allowed participants to extract grammatical generalizations. This evidence in support of specialized mechanisms for extracting words and grammatical regularities was formalized by Endress and Bonatti (2007) as the More than One Mechanisms (MOM) hypothesis. 1.1.2. Proponents of a single statistical learning mechanism In spite of the evidence for multiple learning mechanisms, several alternatives to the MOM hypothesis have been proposed. Amongst these, I will shortly report two of the most influential ones: 1) Perruchet et al.’s (2004) PARSER model and 2) Aslin & Newport’s (2012) account based on saliency. In 2004, Perruchet and colleagues proposed a model which they claimed could better account for Peña et al.’s results. In 1998, Perruchet had already argued, with Vinter, that the chunking of a speech stream into words was the result of the interaction of two principles organizing the cognitive system. The first principle states that those units perceived within one attentional focus (i.e., the chunks) constitute a new unit, which can either vanish or be strengthened depending on whether the association of the very same units further reoccurs in the input. The second principle claims that the perception of chunks evolves as the knowledge about them increases through experience. So, to sum The impact of lexical co-activation through cognates on L2 rule learning 11 This cost has been taken to reflect competition between the first and the second language. That is, when communicating in the L2, speakers avoid interference from the L1 by partly reducing the activation of this language, and the opposite happens when communicating in the L1. This inhibition must be overcome when speakers want to switch from one language to the other. Crucially, in low proficient bilinguals, the inhibition applied to the L1 is larger than the one applied to the L2. Therefore, switching from the L2 to the L1 is more costly than going from the L1 to the L2 (Costa & Santesteban, 2004). Alternatively, the Language-specific model (Costa et al., 1999) argues that there is no cross-linguistic competition and that bilingual speakers are only sensitive to the activation of the lexical items of the to-be-produced language. As such, they ignore the activation of the words of the non-target language without needing to suppress their activation. Costa and colleagues took as evidence for this claim the fact that in a picture-naming task carried out by English-Spanish bilinguals, speakers were faster at naming the picture of a dog in Spanish (“perro”) when the distractor word was “dog” that when it was “chair”. The cross-language identity effect found was taken as evidence for the fact that there is not competition across languages, since if this were the case, the target’s translation should slow down picture naming rather than facilitate it. Along these lines, the cognate facilitation effect observed in bilingual picture naming (Costa et al., 2000) further supports the claim that the target and the non-target languages do not compete for Figure 3. (A) Switching performance of native speakers of Spanish low-proficient learners of Catalan. (B) Switching performance of Spanish–Catalan highly proficient bilinguals in their two dominant languages (Costa & Santesteban, 2004). The impact of lexical co-activation through cognates on L2 rule learning 12 selection, since if competition existed, cognates should also hamper naming rather than enhance it. To date, the question around lexical competition remains unanswered. However, irrespective of the selection mechanism assumed, it has been established that while lexical retrieval is facilitated by cognates, accessing non-cognate words is more costly (Santesteban and Schwieter, in press). This finding is key for the present study, since it predicts that participants learning cognates will process the vocabulary of the artificial language faster and with less effort than those learning non-cognates. 1.4. The impact of cognates on syntax learning and processing In the previous section it has been mentioned that the cognate facilitation effect makes cognates easier to process than non-cognates. Despite most studies have focused on examining cognates in isolation (e.g. Dijkstra et al., 1999, 2010), the cognate facilitation effect has also been observed in sentence contexts. Previous research has analysed whether the facilitation attested in single words persists when cognates appear in full sentences (e.g. van Hell & de Groot, 2008). Nevertheless, these studies have usually focused on lexical processing and have neglected the impact cognates can have on syntactic processing. When two languages share a syntactic structure, the production of such a structure in one of the languages facilitates the utterance of a sentence with the same structure in the other language. This phenomenon is known as bilingual syntactic priming (Loebell & Bock, 2003), and provides evidence that the syntax of a bilinguals’ two languages gets activated during sentence processing (Soares et al., 2018). In the section above, it has also been mentioned that both of bilinguals’ lexicons are activated during language processing. Taking this into account, Schoonbaert et al. (2007) argued that the use of cognate words in a structure shared in two languages could augment the magnitude of the bilingual syntactic priming effect. That is, since cognates are activated faster and more strongly than non-cognates, the boost in lexical co-activation following the use of these words could further facilitate bilinguals’ utterance of a target structure. In the following lines, I will review two studies which illustrate the impact of cognates on grammatical rule learning and processing. The impact of lexical co-activation through cognates on L2 rule learning 13 1.4.1. Hopp’s (2017) study Hopp (2017) examined the differences in German-English bilingual processing of sentences containing L1-L2 cognate and non-cognate verbs. The structures tested were English reduced relative clauses (cf. 1), whose word order partially overlaps with German embedded clauses (SOV). As a result, when reading the first type of structures, L1 German learners of English temporarily activate the canonical word order of German embedded clauses. By introducing cognate verbs in the sentences, Hopp tested between two possible interactions of lexical and syntactic co-activation. First, it could be that the increased lexical co-activation of the L1 through cognates led to a stronger syntactic coactivation of the L1. If this were the case, processing sentences with cognates could be more costly than with non-cognates, since the activation of the embedded clauses in the L1 would interfere with a target-like processing of the reduced relative clauses in the L2. Second, since accessing cognates requires less resources than accessing non-cognates, processing L2 sentences with the first type of words could leave more resources available to inhibit the syntax of the L1. As a result, processing should be easier, since the L1 syntax would not interfere with the target L2 syntactic structure. In order to test these hypotheses, Hopp had 39 German learners of English read 32 sentences and 138 fillers in English. As participants did this, their eye-movements were recorded. The reduced relative clauses had the following structure (2017:105): (1) When the doctor Sarah ignored tried to leave the room the nurse came in all of a sudden. The results of the study indicated that bilinguals’ reading of sentences with cognates was faster than with control verbs, suggesting that syntactic processing was eased by the first type of words. That is, processing cognate verbs was less costly than processing noncognates. Consequently, in the first case more resources could be devoted to inhibiting the L1 syntax and this, in turn, eased target-like L2 processing. By contrast, when the structure contained non-cognates, the high cost associated with lexical processing left very few resources available to inhibit the L1 syntax, which interfered with a fast processing of L2 reduced relative clauses. The impact of lexical co-activation through cognates on L2 rule learning 14 1.4.2. Bastarrika & Davidson’s (2017) study Turning to language learning, Bastarrika & Davidson (2017) explored Spanish adults’ capacity to learn a grammatical rule for number marking in Basque. The authors also assessed brain responses to morphosyntactic violations using magnetoencephalography (MEG), but only behavioural results are reported here since just these go in line with our interests. In order to make sure that the learning effects were indicative of grammatical rule learning, Bastarrika & Davidson created a miniaturized version of Basque (MiniBasque) where all vocabulary was cognate with Spanish. As such, learning was mainly reduced to syntax. Participants of the study were seventeen Spanish speakers with no previous knowledge of Basque. The rule to be learnt could be formulated as: Grammatical number is marked in both Spanish and Basque, but while the former marks it in all the elements of the phrase, the latter just does so on the last element of the structure: (2) a. El dado verde the.Masc.Sg dice.Masc.Sg green.Sg a’. Dado berde-a dice green-the.Sg ‘The green die’ b. Los dado-s verde-s the.Masc.Pl dice.Masc-Pl green-Pl b’. Dado berde-a-k dice green-the-Pl ‘The green dice’ The study consisted of two experimental sessions carried out on two consecutive days. In the first session, participants were explicitly taught the grammatical rule and further trained on their sentence comprehension and production in Mini-Basque via judgement tasks and picture description tasks. That is, on the one hand, they were auditory presented with grammatical sentences and with sentences containing a number marking violation and they were asked to judge which one was good or correct. On the other hand, they were confronted with a picture with either one or two drawings presented in one colour (e.g. a single green die or a pair or green dice) and they subsequently had to produce either The impact of lexical co-activation through cognates on L2 rule learning 15 a singular or a plural noun phrase that described the picture (cf. 2). Following this training, in the second session participants were tested on their knowledge by repeating the very same tasks but with different stimuli. The aim of this test was to make sure that participants could apply the rule to new words, and that they did not just memorize the correct or incorrect noun phrases. In addition, they were also tested on their comprehension and production in Spanish in order to have a baseline for further comparison with Basque. The results of this study found that participants learned the grammatical rule to the point that they achieved a high proficiency in comprehension and production of Mini-Basque (≥ 95%). All in all, Hopp’s (2017) study provides valuable evidence for the fact that cognates have an impact on syntactic processing. On another note, Bastarrika & Davidson’s (2017) study used an artificial language made up of cognate vocabulary to explore rule learning, but the aim of their work was not to test whether these lexical items eased this process. This is reflected by the fact that the authors did not include non-cognates as part of their language’s vocabulary, which would have allowed for an analysis of the effect that the different types of items have on syntax acquisition. Hence, the role of cognates on rule learning remains unexplored. 2. The present study This study examined Spanish speakers’ learning of the grammatical rule of an artificial language drawing on Basque through a cognate and a non-cognate vocabulary. The reason why I decided to base the artificial language on Basque is that Spanish and Basque share a relatively large part of their vocabularies and phonological inventories, but substantially differ in their grammatical systems. I decided to exploit this characteristic to explore rule learning when different degrees of vocabulary acquisition are required. In what follows, I will present the hypotheses and the predictions made and I will detail the method and procedure that were used to test them. 2.1. Research question, hypotheses and predictions The present study tackled the broad question of how second language learning can be eased. More precisely, it asked whether the use of Spanish-Basque cognates facilitates the learning of an artificial language’s grammatical rule. The main hypothesis stated that The impact of lexical co-activation through cognates on L2 rule learning 16 cognates do ease grammatical learning. Thus, it predicted that native speakers of Spanish would learn the grammatical rule way faster and more accurately when the lexical items of the artificial language were cognates with their L1 than when they were not. The language contained both SOV and OSV sentences. This word order alternation was introduced to be able to test for participants learning of the rule, which relied on subjectobject marking (see section 2.3.3. below). That is, if the language was only made up of SOV sentences, then having a rule marking the subject and the object differently would make no sense, since the first item mentioned would always be the subject. Previous studies have found that subject-first sentences are processed faster than OSV ones (Erdocia et al., 2009, 2014). Furthermore, OSV sentences have been reported to involve a higher processing cost due to their syntactic complexity (Matzke et al., 2002). Hence, I hypothesized that SOV sentences would be processed with greater ease than OSV ones. This hypothesis predicted that participants would respond to SOV stimuli faster and more accurately. 2.2. Some relevant properties of Basque grammar In this section I make a brief excursus to characterize some relevant properties of Basque grammar (see A brief Grammar of Euskara by Laka (1996) for more information). This is important since the rule of the artificial language was similar to an existing one in Basque but, crucially, differed from that of the natural language in more than one aspect. Basque is a free word order language; nearly all constituent combinations yield a grammatical sentence. Yet, Basque grammar has been argued to be SOV (de Rijk, 1969; Greenberg, 1963) because it has most properties of this type: it has postpositions instead of prepositions, determiners follow the noun and inflected auxiliaries follow the verb. Basque is also an ergative language. That is, subjects of intransitive clauses and objects of transitive clauses are morphologically identical and bear no overt case ending. By contrast, agentive subjects of transitive clauses carry an ergative case marker (-k). In section 2.3.3. below, we will see how the artificial language’s rule detached from these characteristics. 2.3. Method As stated above, the main objective of this study was to find out whether native speakers of Spanish learned a grammatical rule faster when the vocabulary of the artificial language was cognate with the one in the L1. Using a sentence-picture matching The impact of lexical co-activation through cognates on L2 rule learning 17 paradigm, reaction times and accuracy rates for cognate and non-cognate sentences were studied. According to the hypothesis that learning would be facilitated by cognates, longer reaction times and lower accuracy scores were expected for non-cognate sentences than for cognate ones. 2.3.1. Participants Forty Spanish speakers (36 women and 4 men; age: 18-30) took part in the experiment. All of them reported having no previous knowledge of Basque in a linguistic background questionnaire filled in prior to the experiment. Participants were students at the University of the Basque Country (UPV/EHU) but came from other regions of Spain. Thirty-nine of them were native speakers of Spanish, and one was a native speaker of Catalan who had started learning Spanish at the age of 3 1 . All subjects reported having normal or corrected to normal vision and hearing, and they were paid for their time. Before the experiment began, they all read and signed an informed consent. 2.3.2. Materials The artificial language Two versions of the artificial language were created, each of them counting on thirty lexical items (twenty animate nouns and ten verbs). In one of the versions, the lexical items were full cognates with Spanish words (e.g. Basque bonbero vs. Spanish bombero (‘firefighter’)). On some occasions, the lexical items were slightly modified so that they fulfilled this requirement (e.g. salutatu (Basque) > saludatu (artificial language) for Spanish saludar ‘to greet’). On the other version, vocabulary was made up of noncognates (the Spanish word bombero can also be referred to as suhiltzaile in Basque). In addition, thirty extra cognate words, i.e. twenty animate nouns and ten verbs, conformed the transitive sentences to be listened to in a post-test (see appendix A for the complete list of vocabulary). 1 Catalan is a Romance language which shares a large number of vocabulary items and syntactic features with Spanish. The girl was from the Balear Islands, a bilingual territory in which Catalan and Spanish are in permanent contact. As such, the fact that she was not a native speaker of Spanish was not judged to be an impediment for her to participate in the study. The impact of lexical co-activation through cognates on L2 rule learning 18 Vocabulary items The cognate and non-cognate nouns and verbs to be used in the test were matched in length, since short words are processed more easily than long words (Soares et al., 2018). In addition, their level of orthographic and phonological overlap with their Spanish translations was measured using Levenshtein distance 2 . This was done because during the test participants both read and listened to these words, so the greater the overlap with Spanish, the easier their processing would be (Soares et al., 2018). Levenshtein distance was calculated on bare nouns and verbs (e.g. bonbero for bonberoa ‘firefighter’, salutafor salutatu ‘to greet’). Lexical frequency per million was also calculated for the Spanish nouns and verbs which cognates overlapped with, since I judged that, when confronted with cognates, participants would attribute to these words the frequency the items have in their L1. The dictionary used for that purpose was SUBTLEX-Esp (Cuetos et al., 2011). The mean word length for the nouns used in the present experiment was 7.95 letters. For verbs, it was 8.4. No significant differences were encountered when comparing word length between cognates and non-cognates (p > .05). The Levenshtein distance for the orthographic form of cognate nouns and verbs in respect to their Spanish counterparts was 0.8 and 0.2 respectively. For the phonological forms, it was 0.35 for nouns and 0.1 for verbs. Finally, the mean frequency per million of Spanish nouns and verbs was 16.3 and 19.5, respectively. As for the items to be listened to in the post-test, the mean word length was 8.2 for nouns and 9 for verbs. Nouns’ orthographic and auditory Levenshtein distance with Spanish was 0.95 and 0.6 respectively. By contrast, verb’s orthographic distance with Spanish was 0.7, and auditorily, there existed no differences between these Basque and Spanish words. Finally, the Spanish translations of Basque nouns and verbs had a mean frequency per million of 38.65 and 11.26 respectively (see appendix B for a more detailed description of the lexical items). Experimental sentences Stimuli for each version of the language consisted of 40 SOV sentences plus a set of 40 OSV ones derived from the SOV items (Table 1). For each version of the language, two 2 The minimum number of single-character edits (i.e. insertions, deletions, or substitutions) required to change one word into another. The impact of lexical co-activation through cognates on L2 rule learning 19 lists were generated to prevent participants from listening to both versions of the same sentence. In other words, a participant listening to an SOV sentence did not listen to its OSV counterpart. Hence, the materials were counterbalanced across participants. During the test, every participant listened to 20 SOV sentences and 20 OSV ones which were presented in a randomized order. In addition, all participants listened to 20 sentences (10 SOV and 10 OSV) in a post-test carried out immediately after the test. As mentioned at the beginning of this section, post-test sentences were made up of previously unheard lexical items, all of them cognate with Spanish. Table 1. Example experimental materials. All four sentences mean ‘The actor has painted the doctor’. Audio files The audio files corresponding to the experimental items were recorded by three female native speakers of Basque aged 22-24. These speakers recorded the sentences to be listened to in the pre-test, the test and the post-test (see 2.3.3. below). In addition, the vocabulary items to be used in the learning phase were recorded as well. Verbs were recorded in their citation form (e.g. saludatu ‘to greet’). As for nouns, they were recorded in their monomorphemic form, i.e. the definite article -a, attached to the noun stem in Basque, was removed (e.g. bonberoa > bonbero ‘firefighter’). Recordings took place in a soundproof booth at the Psycholinguistics Laboratory at Micaela Portilla Research Centre, in Vitoria-Gasteiz. Both a Tascam voice recorder (DR-100MKII model, frequency sampling of 44100Hz) and Audacity (version 2.3.0) were used so as to subsequently choose the files having the best sound quality. Speakers were asked to read the items at a normal pace and with natural intonation. Version A (cognate vocabulary) Canonical SOV 1a. Aktore-ak mediku-a pintatu du actor-S doctor-O painted has (V) Non-canonical OSV 1b. Mediku-a aktore-ak pintatu du doctor-O actor-S painted has (V) Version B (non-cognate vocabulary) Canonical SOV 2a. Antzezle-ak sendagile-a margotu du actor-S doctor-O painted has (V) Non-canonical OSV 2b. Sendagile-a antzezle-ak margotu du doctor-O actor-S painted has (V) The impact of lexical co-activation through cognates on L2 rule learning 20 Once the audios were recorded, they were matched in intensity (dB) using Praat (Boersma and Weenik, 1990, 5326 version). The length of the files was also measured. No significant differences were found 1) between the recordings of cognate and noncognate nouns (p= 0.62) and verbs (p= 0.097), 2) between the test SOV and OSV sentences within the cognate (p= 0.08) and the non-cognate (p= 0.21) versions of the language, nor 3) between the SOV and OSV post-test sentences (p= 0.08). The difference in length between the total amount of cognate and non-cognate experimental items was also not statistically significant (p= 0.17). Pictures 120 pictures were created (see appendix C). Pictures of animate nouns were gathered through the Google search engine and further edited using Adobe Photoshop CS5 so that they depicted the transitive actions needed. Each picture corresponded to a pair of SOVOSV sentences, i.e. the same picture corresponded to the sentences a) Antzezleak sendagilea margotu du (SOV) and b) Sendagilea antzezleak margotu du (OSV) ‘The actor has painted the doctor’, as well as to their cognate counterparts c) Aktoreak medikua pintatu du (SOV) and d) Medikua aktoreak pintatu du (OSV). The order in which the subject and the object appeared in the pictures was counterbalanced, so that in half of the pictures the subject appeared to the right and in the other half, to the left (Figure 4). 2.3.3. Procedure The experiment consisted of four parts: 1) a pre-test, 2) a learning phase, 3) the test and 4) a post-test. In all these parts but the learning phase, participants had to carry out a Figure 4. Picture for the sentences Aktoreak medikua pintatu du (SOV) and Medikua aktoreak pintatu du (OSV). Order of characters in the picture: Object (L)-Subject (R). The impact of lexical co-activation through cognates on L2 rule learning 27 administered the test in pen and paper format. First, they provided some information about their gender, the number of years they had been learning Spanish at school and their selfrated proficiency in the language (from 1 ‘nearly non-existent’ to 10 ‘perfect’). Then, participants were informed that they had been given ninety sequences of letters that looked like Spanish, but that only some of them were real words. They were asked to indicate which words they knew by ticking a box next to each word. To compute the final score, the number of words correctly identified, as well as the number of false positives (the nonwords that were recognized as existing words) were taken into account. The completion of the test took from three to five minutes. 2.3.4. Data analysis Data files for each participant were automatically created by E-Prime and saved as E-DataAid 2.0. files. For each of these files, information corresponding to the vocabulary type (cognate or non-cognate), the experimental block (pre-test, learning, test or post-test) and word order condition (SOV or OSV) was collected, as well as the reaction times and the accuracy scores in the four parts of the experiment. All these data were transferred to an Excel file and organised for further analysis. First, each participant’s mean reaction times (and their Standard Deviations (SD)) in the pre-test, the learning phase, the test and the post-test were calculated. Then, those values which were 2.5 times either larger or smaller than the SDs were removed from the sample, since they were considered not to be representative of overall performance. The values removed accounted for 2.97%, 3.48%, 3.25% and 2.88% of all trials in the pre-test, the learning phase, the test and the post-test, respectively. Once these trials had been removed, dynamic tables were created so as to ease the descriptive interpretation of the data. Tables for accuracy were set so that they detailed the sum of correct trials per word order condition. Importantly, no trials were removed from the analysis when it came to accuracy in the pre-test, the test and the post-test. As for the learning phase, the accuracy score of those trials which had been removed on the basis of reaction times (3.48%) was also discarded. This was done because during the task some equipment malfunction was reported. Hence, removing these trials was the best way to make sure that participants’ responses reflected learning, and were not based on chance. Tables arranging reaction times had the same format as the accuracy ones. The impact of lexical co-activation through cognates on L2 rule learning 28 Notwithstanding, instead of the sum, they displayed each participants’ mean reaction times for the SOV and OSV sentences. 3. Results This section analyses the performance of both groups of participants in the Spanish vocabulary size and the WM tests, as well as in the four experimental blocks. Each section starts with a brief analysis of the overall performance in terms of accuracy and reaction times. This allows us to reflect on whether learning was attained irrespective of differences in word order processing. Then, the effect of the two word order conditions on these results is discussed. In the last subsection, a comparison is drawn between both groups’ performance in the pre-test, the test and the post-test. The statistical analysis of the data used paired t-tests and ANOVAs for comparisons within and between parts of the experiment. The whole analysis was carried out in SPSS. 3.1. Performance on LexTALE and the digit span test The results of the LexTALE test determined that participants of both groups were high proficient users of the Spanish language (mean correct responses: cognate group 97,4% and non-cognate group 96,8%). The difference between both groups was not statistically significant (t(31)= 1.28, p= 0.21). In addition, subjects’ self-rated linguistic competence in Spanish was 8.75 for both groups of participants. As for the digit span test, it indicated that participants had a similar working memory capacity (t(38)= -0.85, p= 0.40). 3.2. Performance on pre-test trials As mentioned in the procedure (section 2.3.3), the pre-test was the first task to be carried out. Its aim was to corroborate that, as reported, participants had no previous knowledge of Basque. For this reason, results of all forty subjects are analysed first as a single group. Then, a division is made into the cognate and the non-cognate group. Participants listened to sixteen non-cognate sentences (eight SOV and eight OSV ones) and had to match them to one of two pictures. In order to be fit for the experiment, their accuracy rate had to reveal that they had responded randomly. The impact of lexical co-activation through cognates on L2 rule learning 29 Accuracy Overall, participants correctly matched the picture to the auditory stimuli a mean of 9.1 times (SD= 2.05). This accounted for 56.875% of the trials and suggested, as such, that they had responded by chance. In terms of group division, a paired t-test revealed that the accuracy rate of those subjects who were to learn cognates (M= 9, SD= 1.90; 56.25%) compared to that of those who were to work with non-cognates (M= 9.2, SD= 2.30; 57.5%) was not statistically significant (t(38)= -1.52, p= 0.14). In order to determine the effect of word order on participants’ accuracy scores, I performed an analysis of variance (ANOVA) with word order (SOV and OSV) as withinsubjects factor and group (cognate or non-cognate) as between-subjects factor. The ANOVA revealed that there was not a significant interaction between word order and group, using the Huynh-Feldt (HF) correction, F(1,38)= 1.567, p= 0.218. This indicates that both groups of participants processed the distinction between SOV and OSV sentences equally. More precisely, the accuracy rate of those subjects who were to learn cognates was 56.25% for both SOV (M= 4.5, SD= 1.28) and OSV (M= 4.5, SD= 1.66) sentences. Similarly, the accuracy rate of participants who were to learn non-cognates did not substantially vary from SOV (M= 5, SD= 0.89; 62.5%) to OSV (M= 4.2, SD= 1.69; 52.5%) items. I therefore pooled the data from the two groups for the remaining analysis. In addition, there was no main effect of word order (F(1,38)= 1.57, p= 0.22). This showed that the difference in the accuracy rate for SOV and OSV sentences within all participants was not statistically significant. Reaction times Turning to reaction times, overall participants spent an average of 4.3 seconds listening to the sentence and matching it to the corresponding picture. A paired t-test was performed to determine if the difference between the response times of the cognate group (M= 4.37 sec, SD= 0.95) and the non-cognate group (M= 4.25 sec, SD= 0.93) was significant. Both groups of participants performed the task similarly, t(19)= 0.364, p= 0.720. To ascertain if there was an effect of word order in the reaction times reported, an ANOVA was conducted. The test revealed that, just as for accuracy scores, there was not a significant interaction between word order and group, F(1,38)= 1.180, p(HF)= 0.284. The impact of lexical co-activation through cognates on L2 rule learning 30 Hence, both groups of participants reacted to the distinction between SOV and OSV items equally. Additionally, no significant main effect of word order was found, F(1,38)= 2.914, p(HF)= 0.096. This indicates that, overall, participants’ response times to SOV sentences compared to OSV ones did not differ from chance. 3.3. Vocabulary learning Participants of both versions of the artificial language were taught thirty words —twenty nouns and ten verbs— as part of their artificial language learning (see section 2.3.3. above). Then, they performed a picture-matching task which tested for their knowledge. Out of the forty participants, just four had to carry out the task twice, all of them belonging to the group learning non-cognate vocabulary. In what follows, I report participants’ accuracy rate and response times in their first and second attempts to the picture-matching task. First attempt Overall, in the first attempt both groups of participants succeeded in learning the lexical items of their language variety. The accuracy score was 99.47% (M= 28.65, SD= 0.57) for the cognate group and 88.88% (M= 25.90, SD= 2.81) for the non-cognate group, and the difference between both groups was not statistically significant (t(19)= 0.99, p= 0.34). As for reaction times, the cognate group performed the task significantly faster than the non-cognate group (M= 1.28 sec, SD= 0.218 vs. M= 2.26 sec, SD= 0.429; t(19)= -7.90, p < 0.001). Second attempt As mentioned, four participants of the non-cognate group had to repeat the picturematching task because their error rate exceeded the limit permitted (i.e., they made more than five errors). In their second attempt to the task, participants’ performance improved, and the accuracy of the non-cognate group rose to 90.95% (M= 26.55, SD= 1.75). If we compare the accuracy rate of the cognate group in their first and only attempt to the task with that of the non-cognate group including participants’ second attempt, the difference between both groups becomes even less statistically significant (t(19)= 5.103, p= 0.99). By contrast, non-cognate participants’ reaction times did not vary from the first attempt (M= 2.26 sec, SD= 0.429) to the second one (M= 2.23 sec, SD= 0.45; t(19)= -1.19, The impact of lexical co-activation through cognates on L2 rule learning 31 p=0.25). As such, the comparison between the measures of the cognate and the noncognate group remained significant (t(19)= -10.2, p < 0.001). 3.4. Performance on test trials Once participants had familiarized themselves with the vocabulary items of their version of the artificial language, they carried out the test. During this part of the experiment, a total amount of forty sentences (20 SOV and 20 OSV) were presented. As expected, participants who listened to sentences made up of cognates were significantly faster and more accurate at applying the grammatical rule of the artificial language than participants who listened to non-cognate items. Accuracy The average number of sentences that were correctly matched to their picture was 34.25 (SD= 3.75) for the cognate group and 25.85 (SD= 4.14) for the non-cognate group. These accounted for 85.625% and 64.625% of the trials, respectively. By means of a paired t-test, it was determined that the difference between both groups of participants was statistically significant (t(19)= 5.751, p < 0.001). In order to determine the effect of word order on the accuracy scores of both groups of participants, a repeated measures ANOVA was performed. The test yielded a main effect of word order (F(1,38)= 4.946, p(HF)= 0.032) but no interaction for word order and group (F(1,38)=3.479; p(HF)= 0.07). Yet, I consider that, since the p value is very close to the significance threshold (0.05), the tendency was for the two groups of participants to process the distinction between SOV and OSV differently. As such, I decided to maintain results for the cognate and the non-cognate group separate in subsequent analyses. A series of paired t-tests were conducted to compare the mean scores for each word order in the cognate and the noncognate groups. For the cognate group, the difference in accuracy between SOV (M= 17.25, SD= 1.84) and OSV (M=17, SD= 1.84) sentences was not statistically significant (t(19)= -0.0526, p= 0.605). By contrast, learners of the non-cognate version of the artificial language responded significantly more accurately to subject-first sentences (M= 14.35, SD= 2.5) than to object-first ones (M= 12.3, SD= 4.32; t(19)= -2.175, p= 0.042). The impact of lexical co-activation through cognates on L2 rule learning 32 Reaction times An analysis of participants’ reaction times indicated that, overall, subjects retrieving cognates were faster at listening to and matching the auditory sentence to a picture than subjects accessing non-cognates (M= 4.40 sec, SD= 8.83 vs. M= 6.57 sec, SD= 2.34; t(25)= -4.14, p = 0.0003). When assessing the effect of word order on reaction times, a repeated measures ANOVA revealed that the difference in both groups’ processing of the alternation between SOV and OSV sentences was not statistically significant (F(1,38)= 2.650, p(HF)= 0.112). In addition, no main effect of word order was found (F(1,38)= 0.149, p(HF)= 0.701). This indicates that participants’ reaction times were not significantly faster for neither the SOV nor the OSV word order conditions (SOV: M= 5.51 sec, SD= 2.07; OSV: M= 5.46 sec, SD= 2.15). 3.5. Performance on post-test trials In the post-test, all forty participants listened to twenty sentences (10 SOV and 10 OSV) which were made up of previously unheard cognate vocabulary items. As for the test trials, both participants’ accuracy rate and reaction times were calculated, this time aiming to establish whether the use of cognate vocabulary significantly improved the noncognate group’s performance. Accuracy Those participants who had previously listened to cognates delivered a very good performance. They succeeded to match the sentences to their corresponding pictures in a mean of 91.25% of the trials (M= 18.25, SD= 1.37). Participants of the non-cognate version of the artificial language were also quite accurate in their responses, and correctly performed the task on 78.75% of the occasions (M= 15.75, SD= 2.91). Yet, the difference between both groups was statistically significant, t(19)= 3.387, p= 0.003. Just as reported for the test, a repeated measures ANOVA was conducted to determine the effect that the two word order conditions had on the accuracy rate of both groups of participants. There was an interaction between word order and group (F(1,38)= 4.086, p(HF)= 0.05), and this provided evidence that the cognate and non-cognate groups processed the distinction between SOV and OSV sentences differently. Thus, the accuracy scores for both groups of participants in each of the word order conditions were further analysed separately. In addition, the statistical test also yielded a main effect of The impact of lexical co-activation through cognates on L2 rule learning 33 word order (F(1,28)= 20.3032, p(HF)= 0.001), indicating that participants responded to SOV and OSV sentences differently. Participants of the cognate group responded correctly to 95.5% of SOV sentences (M= 9.55, SD= 0.80) and to 87% of OSV ones (M= 8.7, SD= 1.14). The accuracy rate of participants of the non-cognate group was 90% (M= 9, SD= 1.05) for SOV items and 67.5% (M= 6.75, SD= 2.59) for OSV ones. The difference between the two word order scores was significant for both the cognate group (t(19)= -2.602, p= 0.018) and the non-cognate group (t(19)=-3.684, p= 0.002). Reaction times Turning to reaction times, overall participants of the cognate group were faster at performing the sentence-picture matching task than participants of the non-cognate group (M= 4.67 sec, SD= 1.017 vs. M= 5.26 sec, SD= 1.56). Yet, the difference between both groups was not statistically significant (t(19)= -1.533, p= 0.142). When analysing the impact of word order on these results, an ANOVA revealed a tendency towards an interaction between word order and group, F(1,38)= 4.019, p(HF)= 0.052. By contrast, there was not a main effect of word order, F(1,38)= 3.00, p(HF)= 0.091. On the one hand, it took participants of the cognate group equally long to react to SOV items (M= 4.68 sec, SD= 1.13) and to OSV sentences (M= 4.65 sec, SD= 0.961), and the difference between word orders was not significant (t(19)=-0.282, p= 0.781). On the other hand, participants of the non-cognate group reacted slightly differently to SOV and OSV items (M= 5.02 sec, SD= 1.5 vs. M=5.5 sec, SD= 1.77). This time, the difference in reaction times was significant (t(19)= 2.133 p= 0.046). 3.6. Comparison of performance across the pre-test, the test and the post-test Overall accuracy Table 2 reports the descriptive statistics associated with the accuracy rate of the cognate and non-cognate groups across the three experimental blocks. Because the number of data points per participant varied from one part of the experiment to the other (these were 16 for the pre-test, 40 for the test and 20 for the post-test), the accuracy scores were converted to a 1-10 scale in order to be able to compare the performance across the different blocks. The impact of lexical co-activation through cognates on L2 rule learning 34 Table 2. Descriptive statistics for the accuracy of the cognate and the non-cognate group across the experimental blocks. In Table 2, it can be seen that the pre-test was associated with the numerically smallest mean for the cognate and the non-cognate groups. The mean accuracy increased in the test and finally peaked in the post-test for both groups. In order to test the hypothesis that the experimental blocks and the word order conditions had an effect on participants’ accuracy, a repeated measures ANOVA was performed. The test yielded an interaction between experimental block and group (F(2,76)= 8.241, p(HF) < 0.001), indicative of the fact that the performance of both groups of participants across the pre-test, the test and the post-test was different. Similarly, the test revealed a main effect of experimental block, F(2,76)= 53.236, p(HF) < 0.001. This indicates that, overall, participants’ accuracy scores significantly differed from the pre-test to the test and the post-test (Figure 8). Figure 8. Mean correct responses of the cognate and the non-cognate group across the three experimental blocks. The impact of lexical co-activation through cognates on L2 rule learning 35 In terms of word order, there was no interaction for neither experimental block and word order nor for experimental block, word order and group. This indicates that both groups responded to the distinction between the word orders equally across the three experimental blocks. However, an interaction between word order and group was found, F(1,38)=6.917, p(HF)= 0.012. This interaction suggests that, if the experimental block factor is not taken into account, the cognate and the non-cognate group processed the distinction between SOV and OSV sentences differently. Similarly, a main effect of word order was also found, F(1,38)= 16.128, p(HF) < 0.001. This shows that, overall, one of the two word orders elicited a more accurate performance than the other one (see results for the test (3.4.) and the post-test (3.5.) above). From the pre-test to the test In order to assess how the accuracy rate of both groups of participants varied from the pre-test to the test, a series of ANOVAs were, again, conducted. A main effect of experimental block was found, F(1,38)= 34.895, p(HF) < 0.001, indicating that, overall, the accuracy rate of both the cognate and the non-cognate group improved from the pretest to the test. Additionally, there was a significant interaction between experimental block and group, F(1,38)= 12.957, p(HF)= 0.001, and this revealed that the accuracy of one of the two groups of participants increased significantly more than that of the other group. More precisely, learners of the cognate and the non-cognate version of the artificial language showed an improvement of 29.4% and 7.1% from the first to the second experimental block, respectively (see Table 2 and Figure 8 for mean scores). This improvement was statistically significant for the former, (t(19)= -7.484, p < 0.01) but not for the latter (t(19)= -1.493, p= 0.152). Additionally, there was no interaction between experimental block and word order nor between experimental block, word order and group. This reveals that both the cognate and the non-cognate group of participants processed the distinction between the word orders equally across the two blocks. Nevertheless, there was an interaction between word order and group, F(1,38)= 5.009, p(HF)= 0.031. This means that, without considering the experimental block factor, the two groups of participants responded to the distinction between SOV and OSV stimuli differently. In addition, a main effect of word order was found, F(1,38)= 6.157, p(HF)= 0.018. This showed that the accuracy rate for SOV The impact of lexical co-activation through cognates on L2 rule learning 36 sentences was higher than that for OSV ones in the two experimental blocks (see sections 3.2. and 3.5. above). From test to post-test Turning to the variation in accuracy rates from the test to the post-test, the ANOVA yielded a significant main effect of experimental block, F(1,38)= 27.125, p(HF) < 0.001. Just as from the pre-test to the test, this effect indicates that, overall, participants’ performance improved from the test to the post-test. The interaction between experimental block and group (F(1,38)= 5.024, p(HF)= 0.031) further established that the accuracy score increased substantially more for one group of participants than for the other one. The cognate group showed an improvement 5.625% in respect to the test, while the accuracy of the non-cognate group rose by 11.2% (see Table 2 and Figure 8 for mean scores). The improvement from the second to the third part of the experiment was significant for both groups of participants, t(19)= -2.332, p= 0.031 (cognates); t(19)= -4.82, p< 0.001 (non-cognates). Yet, this significance was much larger for the noncognate group. On another note, no interaction was found for experimental block, word order and group. As such, the cognate and the non-cognate group of participants were found to respond to the distinction between both word orders equally across the test and the post-test. Yet, a significant interaction was found for word order and experimental block (F(1, 38)= 4.381, p(HF)= 0.043). This is indicative of the fact that, if all forty participants are taken together, then the distinction between SOV and OSV sentences was processed differently from the test to the post-test. More precisely, the accuracy scores for SOV stimuli significantly improved from the second to the third experimental block (t(39)= -6.641, p < 0.001), but this was not the case for OSV items (t(39)= -1.867 p= 0.069). An interaction between word order and group was also found (F(1,38)= 5.273, p(HF)= 0.027). Just as described in the previous section, this means that without taking the experimental block factor into account, the cognate and the non-cognate group responded differently to SOV and OSV items. A main effect of word order (F(1,38)= 15.639, p(HF) < 0.001) indicated that SOV items elicited a significantly better performance than their OSV counterparts in the two experimental blocks (see sections 3.4. and 3.5. above). The impact of lexical co-activation through cognates on L2 rule learning 43 That is, as learners get involved in deciphering the meaning of a word through the clues in the context, they make use of cognitive processes which help them retain the word for a long period of time. By contrast, the second mechanism possibly favours rote learning —the short-lived memorization of words based on repetition (2012: 72). Yet, explicitly learning vocabulary also has some benefits, especially for low-proficient L2 learners. Both the amount of vocabulary and the rate at which words are learned are much higher following this method than the incidental learning mechanism, which seems to be more effective for intermediate and advanced learners who manage basic skills such as reading and listening (Nation, 2001; Read, 2004; Tode, 2008, cited in Alipour et al., 2015). Participants of this study had never been exposed to Basque before. Combined with the fact that the words of the language were taught one at a time and deprived of context (see section 2.3.3.), it is very likely that non-cognate learners recurred to memorization techniques and, as such, did not grasp a long-lasting knowledge of the lexical items. Participants learning cognates, by contrast, did not have to memorize the words, since these overlapped in form and meaning with items in their Spanish lexicon. It was not amongst the goals of this study that participants’ knowledge of the vocabulary items matched that of native speakers of Basque. Instead, the study sought that they learnt a set of thirty words in a short period of time. The results of the picture-matching task have revealed that all forty participants were capable of remembering the words of their version of the artificial language when presented in isolation. Yet, in light of the findings of the study, it is possible that non-cognate learners’ poor mastery of the words could have difficulted their lexical retrieval, and this probably challenged their learning of the grammatical rule. A more robust knowledge of the words could possibly ease lexical access and improve their rule learning. This issue could usefully be addressed in future investigations (see section 5 below). 4.3. Processing canonical and non-canonical word orders In the present study I hypothesized that participants would process SOV sentences with greater ease than OSV items. This hypothesis was postulated on the grounds that 1) speakers have been found to process subject-first sentences faster than object-first ones (Erdocia et al., 2009, 2014) and 2) OSV structures require that more items are kept in syntactic working memory and have, as such, higher complexity than SOV sentences (Matzke et al., 2002). As expected, learners of the non-cognate version of the language The impact of lexical co-activation through cognates on L2 rule learning 44 showed a preference for SOV stimuli in both the test and the post-test. As for learners of the cognate group, they more accurately matched SOV sentences to a picture in the posttest but, contrary to expectations, showed no clear preference for any of the word orders in the test. Although this finding is quite hard to account for, a possible explanation is provided by the study by Tily et al. (2011). This work showed that, even if subject-first sentences are more easily processed than their object-first counterparts in natural languages, this tendency could not be that solid when speakers learn artificial languages. The authors had native speakers of American English learn six different artificial language types, each of them having a different basic word order (SOV, SVO, VSO, OSV, OVS, VOS). Each language contained transitive and intransitive sentences. Participants were first exposed to the lexical items (nouns and verbs) of their language and then they were tested on their learning by means of a sentence-video matching task. In addition, they were also tested on their ability to “speak” the language. Interestingly, the results of the study indicated that although subjects learning the SOV and SVO languages attained a better comprehension and production than participants learning the OSV language (comprehension: > 90% vs. 89%; production: ≥ 50% vs. 45%), the difference was not statistically significant in neither of the two cases. This finding matches the one yielded by our study and suggests that the preference for subject-first sentences may sometimes weaken when subjects learn constructed languages. 5. Open issues for future research The interaction between lexical and syntactic learning could further be explored in future investigations. In the following lines, I outline some of the questions that remain unanswered. Is learning eased when the rule is similar between the L1 and the L2? The present study could be conducted again with a slight modification to the rule to be learnt, so that instead of making participants differentiate between SOV and OSV word orders, they would need to distinguish between SVO and OSV sentences. Since Spanish is an SVO language, this change would allow to test whether rule learning is eased when the L2 structure matches the canonical one in the L1. According to MacWhinney (1992), The impact of lexical co-activation through cognates on L2 rule learning 45 similar structures would help learning through transfer from the L1 to the L2. Additionally, Steinhauer (2014: 409) claimed that “native-like proficiency in L2 is reached earlier for those structures that are more similar between the L2 and one’s mother tongue”. In light of these claims, it could be hypothesized that participants’ performance would improve in this second experiment. To what extent do results depend on vocabulary knowledge? Secondly, the study could also be replicated increasing participants’ mastery of the vocabulary items of the language. As mentioned in the previous section, in the present study vocabulary learning is likely to have relied solely on memory, since successfully completing the picture-matching task in the learning phase did not require participants to have attained a profound knowledge of the words. This shallow mastery of the vocabulary items could possibly be responsible, in part, for the fact that participants of the noncognate group had difficulty learning the grammatical rule. Future investigations could examine whether by increasing participants’ knowledge of non-cognate words, learning is improved. One way to do this would be, following McCarthy (1984), by training them to produce and use the lexical items in a wide range of language contexts. Do results vary when learning is implicit? In third place, this study could be replicated by making both vocabulary and rule learning implicit instead of explicit. In the last decades, a growing body of research has shown that adults can learn the lexicon and/or syntax of an artificial language through incidental exposure (e.g. Peña et al., 2002, Grey at al., 2014). Under this approach, learners are not informed that they are being tested nor that there are some specific rules or words that they need to learn. By contrast, learning takes place through meaning-focused tasks (e.g. semantic plausibility judgement tasks) or form-focused tasks (e.g. focusing on the order of words). By adapting the present study to make learning implicit, the question could be addressed as to in which way the degree of explicitness impacts learning outcomes. Do results vary when time pressure is introduced? In fourth place, to strengthen the validity of the reaction times taken in this study, further research could introduce time pressure as a way of eliciting the fastest response possible from participants. As mentioned in section 2.3.3., in the present study subjects could take as much time as they wanted to match each sentence to one of the two pictures displayed. The impact of lexical co-activation through cognates on L2 rule learning 46 This caused some of their response times to be quite long, and some trials even had to be removed from further analysis because they were judged not to reflect participants’ actual processing time (pre-test: 2.97%, test: 3.25%, post-test: 2.88%). By introducing time pressure into the experiment (for instance, following Walker et al. (2017), by asking participants to answer as quickly as possible), it is probable that the response times of both groups decrease. In addition, these could arguably more accurately reflect processing differences between the cognate and non-cognate groups. Nevertheless, this change would come with a cost, since it is very likely that, once time pressure is introduced, participants’ accuracy rates worsen. Can native-like processing be reached? Finally, future research could also assess differences in morphosyntax processing by means of event-related potentials (ERPs). Since the artificial language draws on Basque, both native speakers of Spanish and Basque could be asked to carry out the experiment. To do this, the vocabulary of the language would require some minor modifications, since not all the items used in the present study corresponded to actual words of Basque (e.g. salutatu ‘to greet’, empujatu ‘to push’). Hence, the materials would need to be adapted so that the test could be performed by a native speaker of the language. In addition, while the same procedure of this study would be followed by the Spanish group of participants, Basque natives would skip the vocabulary and rule learning to directly conduct the test. There is strong evidence that L1 transfer plays a role in early stages of L2 acquisition, and when learners reach a very high proficiency level in the second language, the L1 has been found to still be partly co-activated (Steinhauer et al. 2010). Whether at this stage the first language interferes with an appropriate L2 processing or not is a matter of debate. On the one hand, Steinhauer et al., (2009) claimed that in high proficient learners of a second language, co-activation of the L1 and the L2 does not necessarily result in interference and, as such, that native-like L2 processing can be achieved. These authors proposed a description of the stages learners go through in their processing of grammatical and ungrammatical sentences, as reflected by ERP components, since they first start learning a language up to the point they reach a high proficiency. According to them, once the highest proficiency level is reached, no differences in brain signatures should be attested between natives and non-natives. These stages could be taken as The impact of lexical co-activation through cognates on L2 rule learning 47 reference for testing whether native-like processing of morphosyntactic violations in the artificial language can be attained. On the other hand, Erdocia et al. (2014) found that high proficient Spanish-Basque bilinguals processed the distinction between SOV and OSV sentences differently than native speakers of Basque. More precisely, in non-natives OSV items elicited a P600 at second DP position that did not show in natives’ processing. This suggests that, even at high proficiency, learners’ processing differs from native-like. Taking this into account, further research could also explore to what extent native-like processing of SOV and OSV sentences in the artificial language can be reached. 6. Conclusions This study provides an insight into the adult capacity to learn a set of cognate and noncognate lexical items and a grammatical rule from an artificial language when explicit information about the lexicon and the syntax is provided. The aim of the research was to assess, for the first time, if rule learning was eased by cognates. The study also investigated whether subjects showed a preference for SOV sentences over OSV ones during learning. The most significative finding to emerge from this study is that rule learning is facilitated when the vocabulary items of the artificial language are cognates with the L1. When the vocabulary was non-cognate, participants had more difficulty applying the grammatical rule. This was attributed to the fact that retrieving non-cognates was very costly, and this caused non-cognate learners to dispose of few resources to apply the grammatical rule. By contrast, participants of the cognate group could easily retrieve the words, since these matched the corresponding items in their Spanish lexicon. Due to the cognate facilitation effect, participants’ lexical access was virtually cost-free, and this allowed them to easily apply the grammatical rule. One way to possibly ease non-cognate learners’ lexical retrieval and, as such, to improve their rule learning could be to increase the robustness of their knowledge of the words, which the explicit learning mechanism used in this study made quite shallow. The impact of lexical co-activation through cognates on L2 rule learning 48 Additionally, an analysis of participants’ performance in terms of word order revealed that, as predicted, overall SOV sentences were processed more accurately than OSV ones. Yet, the subject-first preference was not met by participants of the cognate group during the test. Although this finding is hard to explain, previous research has found that the preference observed in natural languages can weaken in artificial language learning (Tily et al., 2011). Overall, this study provides evidence that the cognate facilitation effect attested in lexical and syntactic processing also favours rule learning. Yet, the pioneer nature of this work makes the interaction between word and rule learning to remain an intriguing one which could be usefully explored in further research. References Ahmad, Jameel. 2012. “Intentional vs. incidental vocabulary learning”. ELT Research Journal. 1(1): 71-79. Alipour, Fatemeh M., Manije Youhanaee, Hossein Barati & Arman Nasirahmadi. 2015. “Intentional vs. incidental vocabulary learning through games by young EFL Persian speakers”. International Journal of Research Studies in Educational Technology. 4(1): 23-34. Aslin, Richard & Elissa Newport. 2012. “Statistical learning: From acquiring specific items to forming general rules”. Current Directions in Psychological Science. 21:170–176. Bastarrika, Ainhoa & Douglas J. Davidson. 2017. “An event related field study of rapid grammatical plasticity in adult second-language learners”. Frontiers in Human Neuroscience. 11: 12. Bates, Elizabeth, & Judith Goodman. 1997. “On the inseparability of grammar and the lexicon: Evidence from acquisition, aphasia, and real-time processing”. Language and Cognitive Processes. 12: 507–584. Boersma, Paul & David Weenik. 1990. “Praat: doing phonetics by computer”. Ear Hear. 32, 266–268. Bonatti, Luca L., Marcela Peña, Marina Nespor & Jacques Mehler. 2005. “Linguistic constraints on statistical computations: the role of consonants and vowels in continuous speech processing”. Psychological Science, 16:451–459. The impact of lexical co-activation through cognates on L2 rule learning 49 Costa, Albert, Michele Miozzo, & Alfonso Camarazza. 1999. “Lexical selection in bilinguals: Do words in the bilinguals’ two lexicons compete for selection?” Journal of Memory and Language, 41: 365-397. Costa, Albert, Alfonso Caramazza & Núria Sebastián-Gallés. 2000. “The cognate facilitation effect: Implications for models of lexical access”. Journal of Experimental Psychology: Learning, Memory and Cognition. 26: 1283-1296. Costa, Albert & Mikel Santesteban. 2004. “Lexical access in bilingual speech production: Evidence from language switching in highly proficient bilinguals and L2 learners”. Journal of Memory and Language. 50(4): 491-511. Costa, Albert, Marco Calabria, Paula Marne, Mireia Hernández, Montserrat Juncadella, Jordi Gascón-Bayarri, Alberto Lleó, Jordi Ortiz-Gil, Lidia Ugas, Rafael Blesa & Ramon Reñé. 2012. “On the parallel deterioration of lexico-semantic processes in the bilinguals' two languages: Evidence from Alzheimer's disease”. Neuropsychologia. 50: 740-753. Cuetos, Fernando, Maria González-Nosti, Analía Barbón & Marc Brysbaert. 2011. “Spanish word frequency based on film subtitles”. Psicológica. 32: 133-143. de Groot, Annette & Gerard Nas. 1991. “Lexical representation of cognates and noncognates in compound bilinguals”. Journal of Memory and Language. 30(1): 90-123. de Groot, Annette & Rineke Keijzer. 2000. “What is hard to learn is easy to forget: The roles of word concreteness, cognate status, and word frequency in foreignlanguage vocabulary learning and forgetting”. Language Learning. 50 (1): 1-56. de Diego-Balaguer, Ruth & Diana Lopez-Barroso. 2010. “Cognitive and neural mechanisms sustaining rule learning from speech”. Language Learning. 60(2): 151-187. de Rijk, Rudolf. 1969. “Is Baque an S.O.V. language?”. Fontes Linguae Vasconum, 1–3: 319–351. Dijkstra, Ton, Jonathan Grainger. & Walter J. B. van Heuven. 1999. “Recognition of cognates and interlingual homographs: The neglected role of phonology”. Journal of Memory and Language, 41: 496-518. Dijkstra, Ton, Koji Miwa, Bianca Brummelhuis, Maya Sappelli & Harald Baayen. 2010. “How crosslanguage similarity and task demands affect cognate recognition”. Journal of Memory and Language. 62: 284-301. The impact of lexical co-activation through cognates on L2 rule learning 50 Endress Ansgar & Luca L. Bonatti. 2007. “Rapid learning of syllable classes from a perceptually continuous speech stream”. Cognition. 105:247–299. Endress, Ansgar & Luca L. Bonatti. 2016. “Words, rules, and mechanisms of language acquisition”. Cognitive Science. 7: 19-35. Erdocia, Kepa, Itziar Laka, Anna Mestres-Missé & Antoni Rodriguez-Fornells. 2009. “Syntactic complexity and ambiguity resolution in a free word order language: Behavioral and electrophysiological evidences from Basque”. Brain & Language. 109: 1-17. Erdocia, Kepa, Adam Zawiszewski & Itziar Laka. 2014. “Word order processing in a second language: from VO to OV”. Journal of Psycholinguistic Research. 42 (6): 475-494. Friederici, Angela, Karsten Steinhauer & Erdmut Pfeifer. 2002. “Brain signatures of artificial language processing: Evidence challenging the critical period hypothesis”. PNAS. 99(1): 529-534. Gollan, Tamar & Lori-ann Acenas. 2004. “What is a TOT? Cognate and translation effects on tipof-the-tongue states in Spanish-English and Tagalog-English bilinguals”. Journal of Experimental Psychology: Learning, Memory, and Cognition. 30: 246-269. González-Nosti, María, Analía Barbón, Javier Rodríguez-Ferreiro & Fernando Cuetos. 2015. “Effects of the psycholinguistic variables in the lexical decision task in Spanish: A study with 2765 words”. Brain Research Methods. 46(2): 517-525. Green, David. 1998. “Mental control of the bilingual lexico-semantic system”. Bilingualism: Language and Cognition, 1: 67-81 Greenberg, Joseph. H. 1963. “Some universals of grammar with particular reference to the order of meaningful elements”. In Joseph. H. Greenberg (Ed.), Universals of language. Cambridge, MA: MIT Press, 73–113. Grey, Sarah, John N. Williams & Patrick Rebuschat. 2014. “Incidental exposure and L3 learning of morphosyntax”. Studies in Second Language Acquisition. 36: 611-645. Hopp, Holger. 2017. “Cross-linguistic lexical and syntactic co-activation in L2 sentence processing”. Linguistic Approaches to Bilingualism 7 (1): 96-130. Izura, Cristina, Fernando Cuetos & Marc Brysbaert. 2014. “Lextale-Esp: A test to rapidly and efficiently assess the Spanish vocabulary size”. Psicológica. 35: 49-66. Laka, Itziar. 1996. A Brief Grammar of Euskara, the Basque Language. University of the Basque Country. The impact of lexical co-activation through cognates on L2 rule learning 51 Lemhöfer, Kristin & Mirjam Broersma. 2012. “Introducing LexTALE: A quick and valid lexical test for advanced learners of English”. Behaviour Research Methods. 44(2): 325-343. Loebell, Helga & Kathryn Bock. 2003. “Structural priming across languages”. Linguistics. 41: 791-824. McCarthy, Michael. J. 1984. “A new look at vocabulary in EFL”. Applied Linguistics. 5: 12-22. MacWhinney, Brian. 1992. “Transfer and competition in second language learning”. Advances in Psychology. 83: 371-390. Marcus, Gary. 1996. “Why do children say breaked?” Current Directions in Psychological Science. 5: 81–85. Matzke, Mike, Heinke Mai, Wido Nager, Jascha Rüsseler & Thomas Münte. 2002. “The cost of freedom: An ERP-study of non-canonical sentences”. Clinical Neurophysiology. 113: 844–852. Meuter, Renata & Alan Allport. 1999. “Bilingual language switching in naming: asymmetrical costs”. Journal of Memory and Language. 40: 25-40. Nation, Paul. 2001. Learning vocabulary in another language. Cambridge: England; Cambridge University Press. Peña, Marcela, Luca Bonatti., Marina Nespor, & Jacques Mehler. 2002. “Signal-driven computations in speech processing”. Science, 298: 604–607. Perruchet, Pierre & Annie Vinter. 1998. “PARSER: A model of word segmentation”. Journal of Memory and Language. 39(2): 246-263. Perruchet, Pierre, Michael D. Tyler, Nadine Galland & Ronald Peereman. 2004. “Learning nonadjacent dependencies: no need for algebraic-like computations”. Journal of Experimental Psychology: General. 133(4): 573-583. Read, John. 2004. Assessing vocabulary. Cambridge: Cambridge University Press. Saffran, Jenny, Richard Aslin, & Elissa Newport. 1996. “Statistical learning by 8-monthold infants”. Science, New Series. 274 (5294): 1926-1928. Santesteban, Mikel & John W. Schwieter. (in press). “Lexical selection and competition in bilinguals”. In Roberto R. Heredia & Anna. B. Cieślicka (eds.). Bilingual lexical ambiguity resolution. New York, NY: Cambridge University Press. Schoonbaert, Sofie, Robert J. Hartsuiker & Martin J, Pickering. 2007. “The representation of lexical and syntactic information in bilinguals: Evidence from syntactic priming”. Journal of Memory and Language. 56: 153-171. The impact of lexical co-activation through cognates on L2 rule learning 52 Soares, Ana. Helena Mendes, Montserrat Comesaña & Ana Santos. 2018. “Lexicosyntactic interactions in the resolution of relative clause ambiguities in a second language (L2): The role of cognate status and L2 proficiency”. Psicológica. 39: 164-197. Steinhauer, Karsten, John E. Drury, Nicholas Bourguignon & Kristina Kasparian. 2010. “ERP evidence of transfer effects: L1 grammar influences L2 sentence processing”. Linguistic Society of America (LSA) Annual Meeting. Baltimore, ML. Steinhauer, Karsten, Erin J. White & John E. Drury. 2009. “Temporal dynamics of late second language acquisition: evidence from event-related brain potentials”. Second Language Research. 25(1): 13-41. Steinhauer, Karsten. 2014. “Event-related potentials (ERPs) in second language research: A brief introduction to the technique, a selected review, and an invitation to reconsider critical periods in L2”. Applied Linguistics. 35(4): 393-417. Tily, Harry J., Michael C. Frank & T. Florian Jaeger. 2011. “The learnability of constructed languages reflects typological paterns”. Proceedings of the Annual Meeting of the Cognitive Science Society. 33: 1364-1369. Tode, Tomoko. 2008. Effects of frequency in classroom second language learning. Bern, Switzerland: Peter Lang. Toro, Juan M., Luca L. Bonatti, Marina Nespor & Jacques Mehler. 2008. “Finding words and rules in a speech stream: functional differences between vowels and consonants”. Psychological Science. 19:137–144. van Hell, Janet G. & Anette M. B. de Groot. 2008. “Sentence context modulates visual word recognition and translation in bilinguals”. Acta Psychologica. 128: 431-451. Walker, Neil, Christine Schoetensack, Padraic Monaghan & Patrick Rebuschat. 2017. “Simultaneous acquisition of vocabulary and grammar in an artificial language learning task”. Proceedings of the 39th Annual Conference of the Cognitive Science Society: 1307-1312. Weber-Fox, Christine M., & Helen J. Neville. 1996. “Maturational constraints on functional specializations for language processing: ERP and behavioural evidence in bilingual speakers”. Journal of Cognitive Neuroscience, 8(3): 231– 256. The impact of lexical co-activation through cognates on L2 rule learning 59 The impact of lexical co-activation through cognates on L2 rule learning 60 The impact of lexical co-activation through cognates on L2 rule learning 61 The impact of lexical co-activation through cognates on L2 rule learning 62 The impact of lexical co-activation through cognates on L2 rule learning 63 The impact of lexical co-activation through cognates on L2 rule learning 64 The impact of lexical co-activation through cognates on L2 rule learning 65 Post-test pictures: The impact of lexical co-activation through cognates on L2 rule learning 66 The impact of lexical co-activation through cognates on L2 rule learning 67 The impact of lexical co-activation through cognates on L2 rule learning 68