Consequences of the comparative fallacy for the acquisition of grammatical aspect in Spanish

2.1 Tense and aspect in Spanish

2.2 Tense and aspect in German

As previous research on the topic has shown (González and Diaubalick 2020; Diaubalick and Guijarro-Fuentes 2019; González and Quintana Hernández 2018; but see also Díaz and Bekiou 2006 a.o), differences in the expression of tense and aspect can lead to very different acquisition scenarios. German belongs to the so-called non-aspect languages (Schwenk 2012), i.e., it does not possess purely grammatical means to express aspectuality. Although there are different forms available to express the past, these do not represent different aspectual meanings but rather mark different speech styles – as for a formal narrow description based in semantics, therefore, these differences are practically irrelevant. The most frequent past forms are found in the Present Perfect (ich habe gefunden ‘I have found’) and the Preterit (ich fand ‘I found’).^[12] With most verbs, the Present Perfect is preferred over the Preterit, the latter belonging to more formal contexts or written texts (Heinold 2015: 90). Only for a handful of frequent verbs,^[13] the Preterit is equally common in spoken language. Crucially, interchanging the forms is a mere question of formality (i.e., pragmatic), not of meaning (i.e., semantic). For example, the sentence ‘It was raining, when we landed in Munich’ can be translated in four different ways that at most depend on the formality of the linguistic context, but crucially have the exact same semantic meaning (13–15):

Summarizing, we can conclude that in German all grammatical aspect features (especially, perfectivity and imperfectivity) are bundled together. Different from Spanish, the forms do not denote aspectuality and are generally – at least in colloquial context – interchangeable. If semantics require a speaker to mark an aspectual contrast, this has to be done through lexical items (e.g. adverbs/particles) or by reformulating the sentence using another verb or periphrasis. This is, of course, an essential observation that also translates to languages in general: a contribution to the aspectuality by part of the lexicon is always possible, and the semantic value of any such elements as adverbs or temporal-aspectual attributes has always to be taken into account. Crucially, German has no standardized way to express (im)perfectivity analogue to the Spanish grammar. Based on these insights, L1 German learners of Spanish would be receptive to the idea that contextual information (e.g., adverbs) plays a critical role in the selection of aspectual markers. The challenge for a German speaker of L2 Spanish, however, is that (i) some conditions favor conventional, prototypical options that downplay the role of the larger linguistic context, and (ii) the inclination to rely on contextual information is needed but not sufficient to develop the conceptual knowledge of aspect (i.e., semantic contrast of perfective and imperfective relies on notion of distance from the event) (Doiz Bienzobas 2013; Llopis-García et al. 2012).^[14] More on this below.

3.1 Current approaches to native/non-native comparisons

Turning to Second Language Acquisition, the previously presented comparison between Spanish and German has revealed differences between the two language systems that lead to well-known challenges. However, previous research has shown that even native speakers do not always behave as one would expect. For instance, Slabakova and Montrul (2003) find low accuracy rates for some contexts in a Truth Judgment Task, and García and van Putte (1988) attest unexpected behavior in production tasks. Finally, Salaberry (1999) finds that even in free production natives may differ from expectations. These “unexpected results” whereby native speakers seem to fail to demonstrate knowledge of aspect, lead us to question the adequacy of currently adopted definitional construct of aspect. As shown earlier, every verb in the lexicon can occur with all available tense forms,^[15] and speakers choose according to their own perspective and contextual factors. While purely formal approaches adopt a partial description of aspect being an artifact of the methodology employed, in the present study we adopt a much broader description of aspect. Although it is mostly studies within the generativist realm that often strongly rely on comparations between non-native and natives, in what follows we will try to stay within the formal approaches, but embracing a definition of aspect that is broadly conceptualized in terms of contextual information as the ones described above.^[16]

In L1 acquisition, it is commonly assumed that the target system is fully acquired: speakers tend to a uniformity. Such claims are not entirely true for L2 learners. Using Bley-Vroman’s (2009) terms, SLA is characterized by unreliability – i.e., a difference between the acquired language (namely, the Interlanguage Grammar (ILG)) and the target system – and non-convergence (comparing ILGs in different learners).

In recent UG approaches to SLA (see Rothman and Slabakova 2019), if, in an experiment on a concrete feature, learners behave like native speakers (e.g., in a GJT), they are deemed to have attained that feature. In change, if they differ from natives, then their grammars are assumed as not constrained by UG; hence, UG is not (completely) available/accessible. Nevertheless, the proof of UG access is not equal to the acquisition of a target-like feature configuration (see discussion of the feature reassembly below).^[17]

Thus, the question remains how to describe and evaluate observed contrasts. For instance, Slabakova (2009: 169) argues that differences between L1 and L2 speakers are not qualitative but occur rather gradually depending on the specific input. Still, both L1 and L2 are acquired as natural language for communication purposes (Slabakova 2009: 156). Bley-Vroman (2009) takes a different approach, introducing the concepts of patches and viruses. Analogously to how these terms are used in informatics, the former ones denote the conscious use of elements to compensate for acquisition failure, whereas the latter characterize sentences on which native speakers disagree regarding their (un)grammaticality. The existence of such phenomena proves that L1 speakers, too, can choose between grammatically defined and shallow representations (cf. Clahsen and Felser 2006). Sometimes, patches are even necessary to overcome uncertainties.

While recent approaches focus on persisting learnability problems (see e.g., Tsimpli and Dimitrakopoulou 2007), it has been revealed that the results of a conscious learning strategy may superficially resemble the target system, and thus compensate the absent features (e.g. Hawkins and Hattori 2006). However, the underlying representations are different.

The core system is not working (or not working well), so the patching system is taking the burden. […] The language faculty has means of dealing with strangeness and uncertainty, and these are heavily employed in foreign language learning. (Bley-Vroman 2009: 193)

Many researchers (see e.g. Judy et al. 2008) have already demonstrated that these observations make it possible to describe the L2 outcome without direct comparisons to native speakers. Apart from circumventing discussions on what counts as “native” at all (see Mesthrie 2010), complicated even further through globalisation and the broad availability of international communication, this proceeding respects proposals on the need to consider ILGs in their own right stated quite early is UG/SLA research (e.g. duPlessis et al. 1987; Schwartz & Sprouse 1994 to name just a few) Concentrating only on ILG properties, it can be shown that L2 learners may arrive at stable grammars licenced by UG which indeed account for the L2 input (albeit not in the same way as the grammar of a native speaker). The issue, then, is whether the ILG is a ‘possible’ grammar (i.e., licensed by UG), not whether it is equivalent to the target.

Since language variation is supposed to be located at the feature assembly (see Lardiere 2009 for a discussion), only narrow formal features and their mapping to semantic notions are variable, whereas the “general mechanisms of syntactic and semantic computation” is universal (Slabakova 2009: 165). A similar view is defended in the Feature Reassembly Hypothesis (FRH; Lardiere 2008, 2009), according to which SLA consists of several phases (see Hwang and Lardiere 2013): first the L1 feature configuration is transferred, then a subsequent reassembly of features takes place. Allegedly problematic phenomena are explainable through difficulties in mapping the forms to the correspondent features. As pointed out by Rothman (2008), the arduousness of this task may cause insecurities (detectable through an overgeneralization of consciously learned rules that overwrite/blur acquired knowledge). Eventually, the FRH predicts that achieving a target-like feature configuration is possible, while the way to this target is determined by L1 transfer. Of course, this achievement can also be indicted of being a victim of the comparative fallacy, when the target system is understood as the language spoken by native speakers. Important for the matter of this paper, however, the FRH emphasizes the learning process as such (not only the outcome) which is important when contrasting learners on different proficiency levels.

3.2 Redefining the role of native participants in SLA studies

There are studies (see Birdsong and Gertken 2013: 118 for a review) where – maybe unsurprisingly – native speakers showed an even higher variation in their answers than the experimental L2 group. Cook (1997) notes that in the research literature, L1 and L2 learners are treated rather differently in respects of terminology. Concretely, expressions generally avoided in L1 studies such as “false” and “incorrect” are relatively common in the SLA context. This is contradictory (Cook 1997: 43): although the term “interlanguage” suggests a system that can be described in its own rights (see Klein and Purdue’s 1997 basic variety), terms such “error” and “failure” are surprisingly frequent.^[18] Furthermore, in a later publication, Cook (1999) criticises that often the term “native speaker” is equated to “monolingual speaker” (recall Mesthrie 2010 for a redefinition of the concept of ‘native speaker’ in L2 varieties). The only hard criterion to count as native speakers does not concern proficiency but consists of having acquired the language in question from birth, a criterion unfulfillable by L2 learners by definition. Given that bilingual speakers are not only the sum of two monolinguals (Cook 1999: 191; see also Birdsong and Gertken 2013: 111), analyses of differences between L2 speakers and monolingual native speakers a priori correspond to a mismatched or unfair comparison. Birdsong and Gertken (2013: 121), therefore, suggest speaking of high-proficient instead of near-native speakers. As proficiency is an exclusive property of L2 speakers, it allows the consideration of other cognitive variables (such as working memory and attention, but also motivation and emotional factors, see Dörnyei 2006) known to have a general impact on language use.

One might, of course, wonder where a pure analysis of ILGs will lead us. Admittedly, most learners commonly have a certain motivation to improve their proficiency level; i.e., the learners themselves support the idea to compare their performance to native speakers, often supported by external pressures such as possible stigmas attached to foreign accents or similar properties.^[19] Admitting that the interlanguage is a system in its own right would not automatically prohibit to carry out comparisons. However, linguistic research must evidently consider both comparative differences and similarities which contribute to understanding the nature of bilingualism (without ignoring null findings, see Birdsong and Gertken 2013).^[20]

In the following, we will show that the richness of tense-aspect phenomena would help us to air the comparative fallacy with new empirical evidence. Thus, possible consequences for deviations from the target L1 grammar must be considered when learners are analysed to avoid the comparative fallacy. We will not, however, exclude the native speakers’ data but use them differently: since in many contexts of the Spanish past tenses a high degree of subjectivity is at play (Salaberry 2008), native speakers do not converge on their judgments and often indicate uncertainties about the grammaticality status. Consequently, we will show that analysing learner data individually can also reveal knowledge about variation and make it possible to conclude general processes for language acquisition instead of contrasting them only to native speakers’ mean values.

3.3 Defining research questions

Bringing together the thoughts on the variability of tense-aspect phenomena and the reflections on the comparative fallacy, it seems unsurprising that numerous studies on L2 Spanish have confirmed learning difficulties for past tenses (Comajoan 2014). The L1 and the differences it presents in its tense-aspect system is in some cases decisive: for instance, Díaz et al. (2008) ^[21] show that Greek, French or Portuguese-speaking learners whose L1 has a similar system to Spanish have some advantages over Chinese or Japanese learners. As for the reassembly task (see above), the learners must detach the configuration from their L1 and reconfigure the features to achieve the target assembly. Given the role of perspectivity within the realm of tense-aspect, a translation to purely formal features is not easy.^[22] However, since studies especially within the generativist approach rely on comparisons to control groups (as argued above), we will try to apply the analysis described in 2.1. The learning task could then be described as a mapping of the abstract feature [+perfective] to the Preterit and [+continuous], [+habitual] and of [+progressive] to the Imperfect. Crucially, it is only here where the systematic role of perspectivity comes into play: it is within those abstract aspectual features which are then mapped to the concrete morphophonological forms. Whereas for Anglophone learners, for instance, this represents a standard reassembly task in the sense of the FRH, German learners, in a first step, must comprehend that different past tenses express different forms of aspectuality. That is, they have to understand and define what those abstract aspectual features are in the first place. As pointed out above, important in this context is a focus of the process, and not on comparing the outcome to native speakers. That is, it is also perfectly possible that a fully-functioning system might be developed which, although different from native Spanish, does still work as a language with well-defined grammatical rules. In essence, a difference in the use of forms (deviation from the target system) is not equal to failure.

By concentrating on the contrast between the Imperfect and Preterit in L2 Spanish by L1 German learners,^[23] we will address three research questions presented here together with our predictions:

1. Concerning individual variability within the native speakers: do natives behave as a homogeneous group? If not, under which conditions? Since the expression of perspective is not detachable entirely from a subjective (broad) evaluation a speaker has over the situation, we expect the variation within the native group to be quite high. The categories “ungrammatical” and “grammatical” are, thus, in fact questionable as unequivocal labels for a research experiment.

2. Concerning individual variability within the non-native speakers: do non-natives behave as a homogeneous group? If not, under which conditions? Learners shall not be compared to natives in a direct way. However, we expect that the special characteristics which define a learner (proficiency level, access to explicit pedagogical rules) manifest themselves in the data. Since German as L1 lacks grammatical aspect, we expect learners to use the different forms independently of encoding (im)perfectivity. In other words, we prognosticate a certain L1 effect here.

3. Are different uses of the past tenses as illustrated in Section 2.1 (e.g. completed single events or habits representing frequent usages vs. interrupted actions or coerced verbs representing infrequent or non-standard usages) treated differently? As shown in the examples above, sometimes it is quite complicated to tease apart ungrammaticality from inappropriateness both from a theoretical as well as from a more practical point of view, i.e., when the interpretation of a syntactically well-built sentence turns illogical due to the verb form used. It does, thus, depend on the participant’s definition of what acceptability means, i.e., whether a seemingly grammatical sentence with an odd interpretation is seen as acceptable or not. We consequently expect high subjectivity in these cases.

4.1 Participants

Sixty participants with three different language backgrounds were recruited for the study. The native group is formed by 30 native Spanish speakers (median age 23, x‾ = 29.4, s = 10.1), consisting of 22 female and 8 male participants whose variety of Spanish is the European one. For the group of L2 learners, thirty L1 German speakers with a self-reported advanced level of Spanish (i.e., C1-level or higher) were recruited for the study (22 female and 8 male speakers; median age 27, x‾ = 25.2, s = 7.9). Additionally, their proficiency level was confirmed through a 50-items cloze test adapted from the test provided by the Oxford University Language Center. All participants were living in Spain during the data collection where they had spent at least three months before the test date and reported to use Spanish daily. Eight participants, however, had to be excluded from further analysis, because they fell outside the advanced level according to the proficiency test. Note that the variety of Spanish that German learners are exposed to is European Spanish. Therefore, it seemed logical to us to focus on this variety here, too. Although, due to a stronger representation of the Present Perfect in European Spanish language varieties, this might, to a certain point, induce more of an unfamiliarity with the Preterit;^[24] the perfectivity contrast between Imperfect and Preterit is still nevertheless valid and present in day-to-day communicative situations.

4.2 Experimental tasks

In addition to an ethnolinguistic questionnaire and the proficiency test, all participants were given the same offline tasks in written form. Data for this study were collected via a Grammaticality Judgment Task (GJT) consisting of 48 items. Whereas 36 items were experimental items with a Preterit or Imperfect form, the other 12 were distracters. The items were constructed in a way that, in accordance with our research questions, a high level of subjectivity (i.e., speaker’s perspective) was allowed for. The participants were simultaneously confronted with other tasks belonging to a greater study, in which this GJT was embedded (Diaubalick and Guijarro-Fuentes 2017, 2019). Instructions given to the participants were held to a minimum. Participants were asked to evaluate the items on a Likert-Scale from 1 to 5, and correct ungrammatical items if they wanted to.^[25] The decision for a Likert-Scale follows the idea that, due to illogical combinations, sentences can turn odd without being univocally ungrammatical. No time limit was given, but in average data collection took about 25–30 min per participant for the GJT, and around the same time for the other tasks of the experiment. Nevertheless, participants were allowed to go back and forth.

Items were constructed in four categories (examples below) which are based on previous studies on tense-aspect-phenomena (e.g. Guijarro-Fuentes 2005; Slabakova and Montrul 2003, 2008): ten items contrast single actions and habitual events and will be called ‘standard context items’ in the following (comprising the fact that these usages are the most frequent and allegedly the least disputable ones, being usually the first ones presented in a textbook of Spanish as foreign language). These items contrast with those of less frequent uses that aim to study the acquisition of all the nuances connected to the phenomenon: ten items are ‘coercion contexts’ (stative verbs with and without an eventive meaning), ten items contain ‘impersonal subjects’ (inspired by Slabakova and Montrul 2008), and further six items represent non-finished telic actions in an imperfective context which, due to their aspectuality, we will call ‘conflicting feature items’. In each category, half of the items were constructed as ungrammatical. Importantly, although we distinguish between more frequent and less frequent cases here, this does not mean that judgments a priori must be less homogeneous in the latter. In every item, we have a clear expectation based on previous studies cited above.

To avoid the connection to an alleged ideal L1 speaker (see Cook 1997: 41), however, the terms ‘grammatical’ versus ‘ungrammatical’ must be treated with high caution. As we will demonstrate in the next section, they do not represent unequivocal labels for any participant (neither native nor non-native), since arguably acceptability, grammaticality and appropriateness are individually interpreted very differently. Thus, maintaining the terminological distinction between ‘grammatical’ and ‘ungrammatical items’ can be likewise as misleading as to speak of an error analysis. Nonetheless, Ragheb and Dickinson (2011: 120) affirm that, when error is used as a neutral term devoid of content – i.e., as a method and not as a judgment – then its analysis might eventually lead to a result. In the same sense, we will use especially the term ungrammatical to express a deviation from our expectations, not necessarily meaning inaccuracy. Crucially, it is not intended to express any normative notion of falseness but rather a label. Nevertheless, some methodological consequences remain: whereas so-called grammatical items are constructed as expectedly acceptable, the so-called ungrammatical items contain an expectedly inacceptable form. This determines the use of statistics as we will translate our expectations to numerical judgments in order to calculate deviations. Considering what has been presented above, of course, any such statistical deviations would be little surprising.

Examples from our test battery are given in (16)–(19).^[26]

All test items are constructed with the intention that the tense forms used cannot be replaced with an alternative one. That is, in designing the task we were very careful in creating items with no room to choose both forms without changing the meaning of the item.

5.1 Native data – group results

The data from the native group show, items constructed as grammatical and ungrammatical are judged in a very different way. While there is generally a high consistency in accepting grammatical items, the ungrammatical ones are not rejected with the same strength. This confirms our safeguard to use the term ungrammatical as a non-strict label. Table 1 (organised according to the categories shown in examples 16–19) shows means and standard deviation for the two groups of native speakers.

Two main observations arise: first, the only case where we find standard deviation values greater than 1 is in the ungrammatical items of contrasting levels. Second, the means found here were not greater than 3. This means that for all the grammatical items, most speakers agree on their status. This is important because by using a five-point Likert scale, the value of 3 is the one expected to be selected if a speaker is uncertain about an item; other values, on the contrary, indicate tendencies or certainties. Newly, the difference between ungrammaticality and inappropriateness can have an influence on the certainty about the status of an item. Note that especially the judgments within the categories “impersonal subjects” and “contrasting levels” often deviate from expectations (see below).

Observing the native group, Figure 1 shows convincingly that not all categories are treated homogeneously. It seems that, only in standard contexts (i.e., habitual/episodic events) speakers sharply distinguish between grammatical and ungrammatical items, thus showing a greater unanimity than in all other categories. Conversely, our participants do not show clear tendencies in the judgments of impersonal subjects, contrasting levels and coercion contexts.

Figure 1:

Native data in graphical form, where the mean is represented as a cross, the quantiles by vertical lines, and outliers as dots.

Whereas in other research on this topic (including Diaubalick and Guijarro-Fuentes 2019), standard ANOVAs have been performed for finding differences between groups, in the following, we will take a different approach. Instead of asking if groups behave differently, we want to know if items are treated differently according to the context in which they are given and contrast those for each participant group separately. Consequently, we will contrast items for each separate category, asking whether speakers rate these items differently.

In a first step, however, we need to know if those items constructed as grammatical are indeed treated as such; the same later goes for the ungrammatical ones. One way to do this is by performing one-tailed t-tests (assuming normal distribution within the Likert scale). In order to possibly translate the threshold limits required by the test, in the following we only use natural numbers. Table 2 shows the results of these tests that strengthen the impression that only in some cases, items are accepted or rejected with a high level of certainty. Especially in those cases in which the mean values are higher than 2, but lower than 4, we can assume either a high proportion of variability or a general uncertainty – this concerns the grammatical items with an impersonal subject and those with contrasting aspectual levels. Therefore, for these two categories another right-tailed t-test is not applicable. Instead, a left-tailed t-test with 2 as threshold value was conducted.

For standard and coercion contexts, the obtained values are significantly lower than 2, such that we can assume the items are indeed acceptable. Also for the contrasting levels, although the obtained value is slightly higher, there is not enough evidence to reject the idea that the speakers regard the items as grammatical – certainly, they are not unacceptable (with a mean lower than 4). Thus, only in the case of the impersonal subject contexts, the significant result indicates that the items are not consistently judged as grammatical. Nonetheless, they are clearly not judged to be ungrammatical, either.

Turning to the ungrammatical items, the numbers indicate a stronger tendency towards unforeseen judgments. Only regarding the standard contexts, the mean value is higher than 4, whereas in all other contexts it is lower. Thus, the standard contexts can be analysed by means of a left-tailed t-test, whereas in all other cases we must conduct right-tailed t-tests with the value 4 as threshold (Table 3).

The statistical analyses show that only standard contexts constructed as ungrammatical are consistently rejected. As for coercion contexts and contrasting levels, we cannot say that they are absolutely rejected (means significantly lower than 4). In fact, in the case of the contradicting levels, the examples even show to be generally acceptable.^[29] Likewise, the non-significant value in the impersonal subject items means we cannot dismiss the hypothesis that the participants are rejecting the items. In summary, the results concerning the ungrammatical items suggest that there is clearly less certainty and less consistency among the native speakers in rejecting ungrammatical items than in accepting grammatical ones.

What seems essential is to distinguish sharply between frequent and less frequent usages. Consequently, two additional ANOVAs contrasting the different test categories (Table 4) show that neither in the grammatical nor in the ungrammatical categories, items are rated homogeneously across test conditions.

To track these differences down, we applied a post-hoc test using Tukey’s HSD deviding the categories in homogeneous subgroups (Tables 5 and 6). In sum, in non-standard items there is a high effect of subjectivity. The next logical step must thus be to look closer at the individual data.

5.2 Native data – individual results

Again, a clear unanimity among native speakers cannot be expected. Table 7 summarizes the number of individuals behaving as intended. Different from the analyses applied to the group results, the means presented here are calculated for every participant individually. Again, we see that the labels “grammatical”/“ungrammatical” are not categorical. These observations complicate designs involving comparisons between native speakers and L2 learners.

Most observations from the group results are confirmed. However, two values deserve a further comment here. Similar to Slabakova and Montrul’s (2003) findings, we can attest no systematicity; quite an interesting result when trying to consider the FRH (see below). Second, the Preterit in those items of conflicting features triggers a contradiction to the context. However, it is not entirely clear if this contradiction deserves the status of ungrammatical. The illogic produced by the combination of the Preterit with an event that, according to the context, is not completed clearly affects the acceptability, but speakers do not seem to equate this with ungrammaticality. As for standard and coercion context, the results from the group analyses are confirmed.

Taking together the results, and providing an answer to our first research question, that is, do natives behave as a homogeneous group? If not, under which conditions?, the data presented here have shown that none of the natives acted homogeneously. Note that this observation mainly proves that an individually stable system in each speaker does not equal a systematicity on the group level. We find a certain degree of variation regarding acceptance or rejection of items that feature tense-aspect phenomena; this, thus, confirms a problem that had been identified in many previous empirical studies. A prejudiced comparison between non-native and native speakers is thus, a priori, a mistaken one. In the following, we will therefore describe the non-native data in its own right by using the same means as with the native group, i.e., without speaking of failure or errors, but of variation.

5.3 Non-native data – group results

Table 8 summarizes the mean judgments of the 22 highly advanced speakers of Spanish.

We detect various mean judgments close to the value of 3. Again, this may express a general uncertainty. But is it that different from what happened in the native group? Let us now concentrate on the distinction of grammatical versus ungrammatical items. Figure 2 in which Table 8 is graphically visualized makes the contrast visible: In non-standard contexts, the judgments do not clearly diverge. The figure illustrates why we opted for a contrast between the different contexts, first for the grammatical, then for the ungrammatical items, instead of contrasting grammatical with ungrammatical items in each category: it definitely seems to be the context that defines how judgments are made.

Figure 2:

Graphic representation of mean judgments in the German group.

To confirm the impression of a high uncertainty with most non-standard items, the data will be analysed using the same statistical test as in the case of native speakers. Since no mean judgment is passing the threshold (under 2 in the grammatical, over 4 in the ungrammatical items), all t-tests are conducted under the null hypotheses that overall learners still behaved as expected, i.e., t-tests for grammatical items are left-tailed, t-tests for ungrammatical items are right-tailed. A significant result is, therefore, a deviation from expectation. According to Tables 9 and 10, our first impressions are entirely confirmed.

For all non-standard contexts, grammatical items are not strictly accepted, and ungrammatical items are not rigorously rejected. Though recalling the importance of non-significant findings (Section 3.1), affirmations on the standard contexts need more data to avoid possible type II errors. Conducting ANOVAs for the several categories, there is no indication that grammatical items are indeed treated heterogeneously. In the case of ungrammatical items, conversely, learners do seem to differentiate between categories (Table 11).

However, a post-hoc test, using Tukey’s HSD, shows there are no disjoint subgroups (Table 12). It is thus safe to say that most grammaticality judgments among our learners indicate a high variation and possible uncertainty.

5.4 Non-native data – individual analyses

The analyses on an individual level confirm the group data shown above (Table 13). In non-standard contexts, less than half of the participants behaved as expected. Also in the standard contexts, although to a lesser extent, there is high variation.

In the same way as we argued above that individual stability does not imply group homogeneity, here, we can argue the opposite: the lack of homogeneity on a group level does not mean that each individual could not have a stable interlanguage system. Overall, it seems impossible to expect participants to sharply distinguish allegedly grammatical from allegedly ungrammatical items. Since this applies both to native and non-native speakers, there is not objective reason to take differences between the mean values of the two groups as indication of access problems. Instead, we should focus on revealing possible effects a category can have on the judgment. Thus, answering our second research question, that is, do non-natives behave as a homogeneous group? If not, under which conditions?, one needs to consider possible (explicit) learning strategies besides the information that our statistical analyses may provide. For L2 learners whose L1 lacks the perfectivity contrast, strategies known from previous literature on the topic concern the consideration of the lexical aspect (e.g., stativity, telicity, see De Miguel 1992; Salaberry 2008) or the focus of temporal adverbs (Diaubalick and Guijarro-Fuentes 2019; Rothman 2008). This observation is directly connected to our third research question; namely, are different uses of the past tenses treated differently? Meant as guideline, the presence of adverbs supposedly facilitates the choice of the adequate verb form. For instance, ayer (‘yesterday’) denotes a completed time interval that often co-appears with the perfective aspect and, thus, seems to favour the Preterit. Adverbs as frecuentemente (‘frequently’), in contrast, are more often combined the Imperfect. Again, it is crucial at this point to remind ourselves that, of course, adverbials generally do contribute to the aspectual readings of the sentence (as well as lexical aspect does). However, an alleged correlation between some selected adverbs and (im)perfectivity are only rules-of-thumb which do not invariably lead to a target-like sentence; other elements contribute to aspectuality, too. Since in German aspectual differences are not expressed via verbal morphology but can only be made explicit via participles, spontaneous/non-grammaticalized periphrases and adverbs, it is plausible to assume that German-speaking learners base their strategies on similar lexical items. As revealed repeatedly in usage-based studies (see e.g. Wulff and Ellis 2018), the focus on adverbials can have a blocking effect for further learning, since learners know lexical cues often to be a reliable tool in any language. Therefore, the next logical step of analysis is to focus on those items in which the appearing marker is misleading; for instance, an imperfective context co-appearing with the adverb ayer (‘yesterday’). If in such a case the item contains a form of the Imperfect, it is grammatical, but may be rejected due to too rigorous a learning strategy. For an illustration, recall our third research question.

In our experiment, such misleading markers were present in all four categories (see examples 16–19), and therefore reflect the learners’ answers in the standard- and non-standard contexts simultaneously. Our data seem to confirm that indeed in such cases, grammatical and ungrammatical items are not sharply distinguished anymore (Table 14).

A further t-test on these values confirms the deviation from expectation to be a significant one (Table 15). Since learners base their decisions on those misleading markers, the item is not recognized as unequivocally grammatical or ungrammatical.

These results effectively stem from the application of learning strategies, as revealed by the individual performance (Table 16). Whereas only five learners treat grammatical items with a misleading temporal marker as grammatical, there is only one learner who rejects the ungrammatical items.

In light of the answers of our three main research questions, the data are congruent with observations of previous studies: in empirical experiments specifically constructed to reveal the learners’ accuracy with tense-aspect forms, native speakers included as a control group do often not behave as expected. We next discuss our findings in more detail.