A mare in a pub? Nonnative facilitation in phonological priming

The differences between the mechanisms underlying native (L1) and nonnative word recognition are still not entirely understood. The models of native spoken word recognition generally agree about the stages involved in auditory lexical access: auditory input triggers the activation of lexical candidates that are compatible with the incoming input, several activated candidates compete for selection, and the candidate most compatible with the input is selected (Luce and Pisoni, 1998; Marslen-Wilson, 1987). However, it is less clear how lexical activation, competition, and selection operate in nonnative listeners (see Gor, 2018). There is emerging evidence that in visual masked priming experiments, nonnative speakers are more influenced by word form, i.e. orthographic overlap between the prime and the target, than native speakers. In a primed lexical decision task, the speed of a lexical decision, or the answer whether the target stimulus is a real word or a nonword, depends on the properties of the word preceding it (the prime). The prime may either speed up (facilitate) or delay (inhibit) the response to the target. In masked priming experiments, in which the prime is visually present for a very short time and is processed subliminally, L1 speakers do not show facilitation for orthographically overlapping primes and targets that are morphologically and semantically unrelated, but second language (L2) learners do (Diependaele et al., 2011; Feldman et al., 2010; Heyer and Clahsen, 2015; Li et al., 2017). Crucially, nonnative facilitation effects are not confined to the visual masked priming experimental paradigm, but have also been reported for auditory phonological priming. Word-initial overlap in three phonemes has resulted in nonnative facilitation in Russian auditory priming (Cook and Gor, 2015; Gor et al., 2010), a result that is in contrast with native phonological inhibition observed in the same presentation mode with three-phoneme initial overlap between the prime and the target (Slowiaczek and Hamburger, 1992). If form-based facilitation is a general modality-independent property of nonnative word recognition, it is likely to have a strong impact on nonnative lexical access, and needs an explanation. The present study provides new evidence on nonnative facilitation in phonologically related (but morphologically and semantically unrelated) prime-target pairs, as in кoбыла – КАБАК (/kabɨla/–/kabak/, mare – PUB), and offers a possible explanation for the nonnative phonological facilitation effect.

Three aspects of the study by Cook and Gor (2015) call for further research. First, it used a measure of familiarity that was established in a separate task, and was not directly included in the main statistical model. Second, it did not use a standard measure of L2 proficiency to ascertain that the L2 participant group was homogenous and, third, it did not have a native speaker control group. The present study improves on these aspects of the study by Cook and Gor (2015). It uses lexical frequency as an estimate of lexical entrenchment in L2, and adds a semantic priming condition to confirm that the L2 learners access the meaning of the primes and targets in the same frequency range as the one used in phonological priming. It should be noted that while the study by Cook and Gor (2015) demonstrated the role of L2 lexical familiarity in auditory word recognition, it also demonstrated the complexities associated with identifying and coding low-familiarity recognizable words in an offline translation task, and with using the information about familiarity in the statistical analysis of reaction time (RT) data in a priming experiment. The present study follows Diependaele and colleagues (2013) and treats lexical frequency as an estimate of familiarity. The selected frequency ranges have been validated using the existing empirical data on the correspondence between Russian lexical frequency ranges and L2 proficiency levels, at which the words in these ranges are familiar to L2 learners of Russian (Solovyeva et al., 2011).³ The study tests and controls L2 proficiency, and adds a native control group to establish the native baseline. It explores the hypothesis that nonnative phonological facilitation in priming experiments is a result of low entrenchment of L2 lexical entries in the mental lexicon.

The design of this study is based on the following logic: When L2 learners access low-frequency, low-familiarity, and weakly entrenched words, they will not be engaged in strong lexical competition because such words have fuzzy L2 phonolexical representations (Cook et al., 2016). In this case, L2 learners will take advantage of sublexical phonological similarity between the prime and the target, which is known to lead to facilitation (Slowiaczek and Hamburger, 1992).

The study uses lexical frequency as an estimate for L2 word familiarity and the level of entrenchment, and divides the frequency continuum into two frequency ranges, high and low. This design with frequency as a categorical variable makes it possible to test the hypothesis that only high-frequency highly familiar primes, which are represented in the nonnative mental lexicon in a more nativelike way, will produce inhibition. Conversely, low-familiarity nonnative words, even if they are numerous, are predicted to produce facilitation. Based on this argument, weak lexical competition will allow sublexical facilitation due to prime-target form overlap to dominate in nonnative lexical access. Alternatively, if nonnative speakers experience an increased lexical competition as a consequence of spurious activation of phantom competitors (Broersma and Cutler, 2008), then strong inhibition in L2 will be observed. Semantic priming serves to ensure that the prime-target pairs in the same low-frequency range show semantic facilitation, i.e. L2 learners can access at least some of their lexical meaning that will be sufficient to form semantic associations. Taken together, low entrenchment hypothesized for low-frequency primes and targets, and some sensitivity to meaning demonstrated by the same L2 participants in semantic priming with words from the same frequency range ensure that the stimuli have low familiarity, but are not completely unknown to the L2 participants.

Forty phonologically and 40 semantically related pairs of primes and targets were selected in the frequency range with log₁₀ frequency of 0.1 to 2.33 (1 to 212 per million) in Sharoff’s Corpus, which is a Russian language corpus with approximately 90 million words at the time of use (see Table 1 and Appendix 1).⁴ The primes were always lower in frequency than the related targets. In order to compare the priming effects for the high- and low-frequency prime-target pairs, 40 phonological and 40 semantic prime-target pairs were divided into two sets via a median split (Mdn_log = 1.16, 15 per million): 20 high-frequency pairs (M_log = 1.91, Min_log = 1.28, Max_log = 2.33, 19–212 per million) and of 20 low-frequency pairs (M_log = 0.68, Min_log = 0.06, Max_log = 1.04, 1–11 per million). Word frequency in the datasets for the semantic and the phonological priming experiments did not differ for the entire item set: t(77.97) = 0.19, p = .847, separately for high-frequency items: t(37.16) = 0.22, p = .823, and low-frequency items: t(37.96) = 0.46, p = .646.

The average amount of initial overlap for phonologically related primes and targets was 3.2 phonemes, with all the pairs overlapping in, at least, three phonemes. Table 2 reports the average length of the stimuli in each condition. Each phonologically and semantically related prime was paired with an unrelated prime, which was matched with the related prime in lexical frequency. In addition to 40 phonologically and 40 semantically related pairs, 120 word pairs were added as fillers to the total of 200 pairs. Another 200 pairs of word primes and nonword targets were added to equal the number of word targets. The nonwords were created by manipulating real word onsets. Forty pairs with nonword targets designed to match the phonologically related condition had the same phonological overlap to avoid any phonological bias associated with the lexicality of the target (i.e. if the prime and the target have the same onset, the target will be a word). Two 400-word presentation lists were created for two counterbalanced versions of the task: with half of the phonological and semantic pairs related, and half unrelated in a quasi-random order. Each participant heard the prime and the target only once (half of the pairs were related and half unrelated). The materials were recorded by a male native speaker of Russian, and were digitized for presentation using PRAAT (Boersma and Weenink, 2013). The experiment was programmed in DMDX (Forster and Forster, 2003) and administered on a laptop PC computer with a Logitech Precision USB gamepad. Both the prime and the target were presented aurally through the headsets, with the interstimulus interval (ISI) of 320 milliseconds. RTs were recorded from the target onset. Participants were instructed to press one of the two trigger buttons on the gamepad depending on whether the second word in each pair was a word or a nonword. Each trial timed out at 4000 ms if no response was made. Experimental stimuli were presented in blocks of 20 items with instructions to take self-timed breaks after each block if needed. Participants preferred not to take breaks during the experiment. The main experiment was preceded by a 10-item practice session.

Twenty-nine paid volunteers provided informed consent and participated in the study: 18 nonnative speakers of Russian (L2, 4 female; mean age: 34.29, SD: 12.82), and 11 native speakers of Russian (NS, 8 female; mean age: 25.56, SD: 2.42). L2 learners were all native speakers of American English and learned Russian in a formal classroom. The average age of onset of Russian was 18.4 years (range: 13–27), and the length of formal study was 12 years (range: 3–39). L2 participants were screened into the study using a standard global test of proficiency: the Interagency Language Roundtable (ILR) oral proficiency interview (OPI). They all scored at ILR 3 (equivalent of ‘Superior’ on the American Council on the Teaching of Foreign Languages (ACTFL), and C2 on the Common European Framework of Reference for Languages (CEFR) scales).

A linear mixed-effects modeling approach was used to analyse log-transformed RT data, and a logistic mixed-effects model (glm function) to analyse accuracy data for phonological and semantic priming. All data analyses were done in R programming environment (R Core Team, 2015), using lme4 package for R (Bates et al., 2014); the degrees of freedom and p-values were generated using lmerTest package (Kuznetsova et al., 2015). The data for phonological and semantic priming were fitted separately because both the primes and the targets were different across these two priming sets. However, we used the same model structure to achieve better comparability of the results between the two models. The final model both for phonological and semantic priming included three fixed factors (group: NS, L2; frequency: high-frequency, low-frequency; and condition: related, unrelated), and two random effects for participants and items entered as varying intercepts. The levels of the fixed factors were treatment-coded with the intercept estimating the dependent variable (accuracy or RT) for the unrelated condition in high-frequency words for native participants. The models’ estimated coefficients for each factor and interaction terms, standard errors, degrees of freedom, the t statistic (or z statistic), and the p values are presented in Table 3. Pairwise comparisons were made within the full model by changing the intercept. Table 4 contains the descriptive statistics.

The study documents a reversal of the phonological priming effect in L2 from inhibition for high-frequency prime-target pairs to facilitation for low-frequency pairs. At the same time, for native participants it shows the expected inhibition in the high-frequency condition that weakens in the low-frequency condition. In semantic priming, facilitation is observed both in L2 participants and NSs in both frequency ranges. Therefore, the study confirms that facilitation in form priming in the absence of morphological or semantic relations is a more general property of nonnative word recognition that transcends the visual domain and is present in phonological priming as well (see Heyer and Clahsen, 2015, and for the visual modality Li et al., 2017). The direction of the priming effect in nonnative speakers depends on lexical frequency, and presumably, the level of familiarity and lexical entrenchment of the prime-target pair (Diependaele et al., 2013).

Importantly, nonnative facilitation in semantic priming indicates that nonnative participants accessed the lexical meaning of low-frequency primes and targets in the same frequency range as the ones that produced phonological facilitation. Therefore, a certain level of familiarity with both the primes and targets must have been present. According to Cook and Gor (2015), it is this low level of familiarity resulting from low exposure (not completely unknown, but not well-known words) that leads to weak and fuzzy phonolexical representations, and ultimately, to weak lexical competition.

The reported results suggest that nonnative auditory word recognition is less influenced by lexical competition and more influenced by sublexical phonological properties of words than native word recognition. Crucially, this nonnative property of lexical access emerges only for less familiar, or low-frequency L2 words characterized by low entrenchment (see Perfetti, 2007). The degree of familiarity depends on the level of L2 proficiency: the higher the L2 proficiency, the more words are well-known to L2 speakers. Therefore, the role of sublexical processing in auditory word recognition should decrease at higher L2 proficiency levels when lexical competition becomes robust. This study argues that a larger cohort of competitors, which appears because of spurious, or phantom activation, does not necessarily lead to strong nonnative lexical competition observed in native lexical access (see Broersma, 2012; Cook and Gor, 2015). Its results support the claim that nonnative low-frequency low-familiarity words do not compete strongly for lexical access, possibly because they have low-resolution (or fuzzy) phonolexical representations. The study does not argue for a different mechanism underlying L2 lexical access, but rather shows that L2 speakers do not strongly engage lexical competition in processing low-familiarity words.