In chunks we trust . . . the problem of gender assignment in foreign language learning of French

Material and participants

In order to check for possible fossilisation or persistence of erroneous gender assignment, a corpus of learner productions was created. This analysis is, therefore, based on a corpus of 35 texts written by German-speaking learners. The learners participating in this research project – 23 girls and 12 boys between the ages of 14 and 19 – are Austrian students attending either secondary school or vocational school in Upper Austria. While 18 participants (51.43%) attend a so-called AHS (Allgemeinbildende Höhere Schule ⁴ ), 17 participants (48.57%) attend a so-called BHS (Berufsbildende Höhere Schule ⁵ ). Although these schools differ slightly in terms of their curriculum, French is a major subject for all students participating in this research project. Each participant has an average of 2–4 hours of French per week, which depends mainly on their school level as well as on their individual school programme. Since the types of school of the learners participating in this project differ greatly, and thus also the progression in the individual subjects, their knowledge of the French language also varies. When assessing their level of French according to the Common European Framework of Reference for Languages (2018), a work published by the Council of Europe (2018) with the aim of providing a definition of proficiency levels through detailed descriptors, the participants’ knowledge of French corresponds to levels A1 (2 subjects), A2 (13 subjects), and B1 (20 subjects; Council of Europe, 2018, p. 78).

This variety of proficiency levels makes it possible to examine the presence of different types of errors, such as gender errors at several levels. With regard to knowledge of other languages, it should be noted that all the participants used German as their first language. In addition to German, they speak English as their first foreign language. French is, therefore, the second foreign language for most of the participants or, in other words, the L3.

The corpus includes various texts of between 60 and 250 words. Themes covered include family, friendship, the world of work, new technologies, health, nutrition, sport, leisure activities, holidays, fairy tales, and Austria, among others. This wide range of topics, from the familiar to the everyday, corresponds de facto to the learners’ areas of interest.

Results

As there were only two participants corresponding to the A1 level; they were not included in the analysis.

Looking at the overall results, one can easily detect an outlier in the B1 group (see Figure 1). In Figure 2, the outlier has been removed and it becomes evident that both groups are working in a very similar and comparable way, and the parallelism between the A2 and the B1 group arises very clearly: both groups behaved relatively homogeneously with regard to errors in gender assignment and agreement.

OPEN IN VIEWER

Figure 1.

Occurrences of gender errors of all participants A2/B1.

OPEN IN VIEWER

Figure 2.

Occurrences of gender errors of all participants A2/B1 in percentage to the totality of all errors.

The analysis of the given data in a boxplot ⁶ also shows that both groups are relatively stable concerning their distribution around the median (see Figure 3), even if the interquartile range of the B1 group is slightly larger and the dispersion within the group is greater. Despite the small number of participants in each group, it is visible that both groups behave relatively homogeneously among themselves and their behaviour with regard to gender violations suggests a certain parallelism that might indicate fossilisation.

OPEN IN VIEWER

Figure 3.

Comparison of total incorrect answers within the group.

The groups fluctuate quite stably around a median of about 10%. These findings are consistent with studies by Carroll (1999) or Franceschina (2001), for example, who also confirmed gender assignment and agreement as a persistent learner problem, even in advanced learners. In their studies on gender assignment in foreign language learners of Spanish (L2), they also found fossilisation to be around 10%. Note that a 10% error rate equals an entire grade level, which is quite significant for learners during their school career.

In a further step, the errors concerning gender assignment and agreement were categorised according to the type of part of speech involved or the syntactic construction in which they occurred (see Figure 4).

OPEN IN VIEWER

Figure 4.

Violations in gender assignment and agreement, A2 and B1.

The analysis shows the differences between the groups: as the syntax becomes more complex and as the level increases, errors concerning the congruence between past participle and gender occurred more often in the B1 group. In the A2 group, there were only two occurrences of a past participle form, and both were correct in their agreement. Instead, the A2 group makes more errors in the article + noun sequence – the ‘pure’ gender assignment – that is to say, assignment without agreement or syntactic alignment. We, therefore, observe a persistent problem in the learners’ behaviour and there is only little progression. The question remains as to why gender assignment is so hard to acquire – even in groups that have another Romance language as an L1 as, for example, Kupisch et al. (2013) have shown for learners of French with Italian as their L1.

These results, even though they were analysed for a very small sample and therefore have only very limited significance, lead us back to our initial questions: why is gender so difficult to acquire and what could be done in the input to help facilitate the processing of precisely this cue? In order to approach these questions, previous attempts at an explanation from different perspectives of acquisition will be used to examine how valid these approaches are.

Modular approaches

Most modular approaches argue from a Universal Grammar (UG)-based perspective (Carroll, 1999; Franceschina, 2001, 2003, 2005; Granfeldt, 2003; Kupisch et al., 2013). These UG-based approaches usually attempt to explain errors via the lack of, or decreasing access to the UG or via the transferability of language knowledge, that is the possibility of positive or negative transfer (Meisel, 2011; Schwartz & Sprouse, 1996). For example, Carroll (1999) or Franceschina (2003, 2005) conclude that gender assignment is accessible to adult learners, but some seem almost to be blocked from acquiring it due to existing language knowledge. As a consequence, they lack the possibility of relying on bootstrapping effects for positive transfer. Most explanations of persisting problems with gender assignment ultimately end up in fossilisation, arguing with the Failed Functional Features Hypothesis (FFH, Hawkins & Chan, 1997), ‘which predicts that the acquisition of L2 functional features not instantiated in the learner’s L1 feature inventory will result in persistent divergence in adult learners’ (Franceschina, 2003, p. 97) and which is based on the distinction between universal and parameterised features. Franceschina’s works (2001, 2003), however, suggest that the acquisition of gender takes place in a different way than, for example, the acquisition of number and cannot fully be explained by the FFH. She concludes that there is no critical period for the acquisition of syntax but dismisses the possibility that the gender (here = article) + noun sequence might be learned as a multiword unit in the form of a construction with given slots (Franceschina, 2001, p. 244). The possibility of acquisition by storing chunks as a first step in this process is not discussed at all. She also seems to exclude a probable influence of different processing strategies between NS and non-native speakers (NNS):

However, gender has been shown to be a persistent problem for some adult L2 learners. Failure to acquire a property of the L2 for which there is unambiguous positive evidence in the input intriguing. Given that some L2 learners [. . .] are indistinguishable from NSs in their production, comprehension and metalinguistic judgements of grammatical gender, we can rule out the possibility that the mere fact that these speakers already have knowledge of a previous language prevents them from fully acquiring the relevant properties of the L2. This suggests that the concerns expressed by Cook (1992) and Ellis (1991) about L2 speakers being intrinsically different from NSs, and therefore not directly comparable to them, may not be justified. (Franceschina, 2005), p. 191)

To sum up the main findings, UG-based approaches do suggest that adult learners are able to acquire native-like knowledge of grammatical gender (Franceschina, 2005). They also argue that the learner’s L1 plays a crucial role concerning possible attainment of grammatical gender and that extralinguistic factors such as age, length of exposure, task modality and language skill have to be taken into consideration as possible influencing factors for some of the variability. Nevertheless, the strongest influence is exerted by the L1, from which it can be concluded that transfer effects should be expected to be high, especially for learners who already have a gender system in the L1. ⁷ Positive transfer should therefore be a possible strategy for learners.

Despite this, a qualitative look into our data seems rather to confirm that positive transfer ⁸ does not play an important role in gender assignment:

In Example (4), it would have been possible for this learner with L1 German to rely on the German word eine Tasse ( ‘a cup’), which is female in both German and in French. Obviously, the learner knew that an article must be put in front of the noun (possible explanations: concept of the sequence/construction stored as a whole or transfer of the +gender parameter of the L1) but could not bootstrap on the L1 system. The same case occurs in every example (4–10). It should be noted that learners do not forget to put the article, which may underline the already cited hypothesis that gender has to be learned lexically as part of every item and that these sequences are stored, at least in the first phases of L2 acquisition, as a whole, comparable to multiword units.

However, Examples (7–9) also show that the learners do not rely on morphological cues such as typically feminine endings like -tion or -ée, and there are also errors in gender assignment that might be caused by negative transfer from the L1 into the target language:

Interestingly, in both cases, these are constructions that could be grouped under so-called formulaic speech. It is possible that the entire chunk was transferred by the learners here – however, the data set is too small to be able to make valid statements. If the thesis that chunks are preferably transferred as a whole were to be confirmed in a larger data set, this would speak strongly in favour of usage-based modelling within the framework of CxG.

A very striking example from the corpus shows that there are violations in gender assignment even if the semantic gender is not arbitrarily assigned, but is transparent due to the properties of the noun:

These are certainly isolated examples, but they are repeated within the corpus and at least call into question the role of transfer in relation to grammatical gender. With regard to foreign language learning explanations that only argue by the possibility of access to UG, these explanations seem to be insufficient and unsatisfactory as they cannot explain either cases such as in Example (4) or the very similar results of the –gender group and the +gender group in the study by Franceschina (2003), as she herself highlights.

Usage-based approaches

It has been already mentioned that there are some phenomena in the dataset that might indicate that the framework of CxG and a usage-based perspective on foreign language acquisition could help to shed light on the difficulties learners have in acquiring gender. Unfortunately, there are only a few studies that examine gender assignment from a usage-based perspective and the arguments mostly refer to frequency effects (Wegener, 1995, 2005, 2016). Both studies show that acquisition takes place. Both are working with children who are acquiring German as a foreign language and who possess different heritage languages, and both attribute this to frequency in input, but they give no explanation for the fact that children continue to make errors in gender assignment in their studies. In the end, processing is hardly taken into account.

However, the works of Binanzer (2017, 2020; Binanzer & Wecker, 2020) can possibly be located in the area of a usage-based, and above all, strongly input-focused framework. Although she primarily investigates the acquisition sequences for gender in younger learners (target language German), she clearly presents in her studies that learners must first acquire patterns so as to be able to acquire agreement patterns, the so-called ‘consistent agreement pattern’ (Corbett, 1991), in order to find out which linguistic units are gender-sensitive targets (Binanzer & Wecker, 2020; Corbett, 1991) and how the genre information has to be mapped there (Binanzer & Wecker, 2020). Her findings certainly indicate that L2 learners may also be able to process probabilistically and filter the input for statistically probable patterns, even if they no longer seem to be as sensitive to this as children in first language acquisition. Nevertheless, implicit statistical learning seems to still be possible (Binanzer, 2020; Binanzer & Wecker, 2020). In consequence, she concludes that the most important step is to access the semantic congruent patterns:

Once the semantic congruent patterns and plural schemata are securely available, meta-linguistic knowledge and language-practical skills about the formal congruent patterns and gender classification regularities or paradigmatic relations between singular and plural forms can be built up in the same way, i.e. through awareness and subsequent proceduralisation and automation. (Binanzer, 2020, p. 24, translation JW) ⁹

The study does not address how these patterns can be made more accessible to learners via possible input optimisation and manipulation, however.

If we follow general assumptions of CxG and usage-based approaches regarding the entrenchment and establishment of schematisation during language acquisition, ¹⁰ it is important to examine why gender so stubbornly resists these processes. After all, it is an almost overrepresented category in the input, but frequency effects do not seem to occur, as learners obviously do not manage to process it sufficiently. The idea that patterns may play a role in gender assignment and, therefore, also the ability of probabilistic pattern prediction during processing can probably be seen in the previously cited study by Ayoun (2020): she investigated the gender assignment of L1 French speakers and found that they were uncertain about the correct gender in about 25% of the cases. She then investigated the strategies speakers used to assign the correct gender and found that speakers used both combinatoric pattern and rule knowledge, which suggests that explanatory potential from both construction grammar and UG is at work while acquiring morphological categories like gender. In her study, NS first looked for prototypical multiword units by adding adjectives or building a whole phrase, and then checked these units on the phonological level (Ayoun, 2020). ¹¹ However, Ayoun (2020) argues from a purely UG-based perspective and concludes that NS only process via phonotactic rules, which they try to (re)find via these strategies. She seems to overlook the possibility that these strategies may aim to process first via pattern-driven and combinatorial probability to re-access the rule-based processing that adult L1 speakers seem to have ‘lost’.

This leads us back to the question of how the processing in NS of morphological categories with little salience may be imitated in learners and what might be useful for optimising learners’ input.

Discussion

It seems important to pay more attention to the fact that the acquisition of morphological categories in processing depends strongly on salience and attention. This is where the comparison between first language acquisition and foreign language acquisition becomes significant. Studies have already shown that children tend to process morphologically and prosodically, while adults process semantically (Lew-Williams & Fernald, 2007; Havron & Arnon, 2021). This leads to a different behaviour in the perception of salient features during language processing. With this, however, gender also loses the function of predictability about properties of the noun, since learners process primarily lexically-semantically (Grüter et al., 2012; Lew-Williams & Fernald, 2007). It was also demonstrated that children process and acquire faster when they process multiword units (Arnon & Christiansen, 2017).

In addition, Eisenbeiß (2003) was able to illustrate in her work that the gender system in first language acquisition (L1 German) was only reliably acquired when the case system was also stably acquired. Consequently, case and gender seem to support each other during the automatisation process and by this way facilitate the storage of the mental representations of morphological categories (Eisenbeiß, 2003).

With this, Eisenbeiß (2003) also demonstrated that children benefit not only from multiword units, but also from the fact that several mutually supporting features were present in the input. This is in line with the previously mentioned different cognitive processing routes that can be observed in children and adult learners (Clahsen & Felser, 2018; Frenck-Mestre et al., 2009; Havron & Arnon, 2021; Jost & Christiansen, 2016; Pulvermüller et al., 2013; Schimke & Hopp, 2018; Zwitserlood, 2018).

Given all these findings, the assumption for the gender problem is, therefore, that morphological features are no longer sufficiently salient in the processing by adult learners to attract enough attention for them to be stored. The learners pay too much attention to the semantic features of the input; they process the noun label and oversee the article + noun sequence, perhaps due to economic constraints and their processing experience in the already acquired language systems. To optimise processing, it is therefore necessary to find a method in the input that triggers both combinatorically and rule-based processing routines. The aim in doing so is to control the perception and processing of the input as well as possible. This seems to work best through the presentation of formulaic sequences and multiword units (Arnon & Ramscar, 2012). However, is this already being implemented in textbooks?

Problems with the didactic input?

It has been a long time since input enhancement was discussed and tested in Foreign Language Learning, but a look in current manuals for French as a foreign language shows that input enhancement is much more difficult than one might think, if one considers the cognitive processing and the problem of salience of the cue that should be learned. Figure 5 shows examples of input enhancement that is meant to draw the learner’s attention to gender assignment in difficult cases – the gender is not marked on the phonetic level in the following cases. ¹²

OPEN IN VIEWER

Figure 5.

Examples of input enhancement that should draw attention to difficult gender assignment; À plus! Nouvelle édition 1. Gymnasium Bayern. 2017. Cornelsen.

Using the examples of the French words for ‘the friend’ (l’ami, masculine versus l’amie, feminine) in which gender is not audible if the direct article is used, the manual tries to explain that in every such case, the male word is written in blue and the female word in red to facilitate the gender differentiation. It is easy to detect that this solution triggers the noun label processing that might block the processing of the morphological cue as has been showed by Arnon & Ramscar (2012).

Figure 6 shows an additional confusing factor of the method proposed by the manual À plus: besides the problem that, once again, the noun label processing is triggered by colouring the whole lexeme, the colour red is also used here to mark the irregularly formed plural in French. These are several false runs for the learners to process correctly.

OPEN IN VIEWER

Figure 6.

Examples of input enhancement that should draw attention to difficult gender and number assignment; À plus! Nouvelle édition 1. Gymnasium Bayern. 2017. Cornelsen.

Interestingly, the interrogative and the possessive pronoun are not marked at all, whereas in the case of the interrogative pronoun there is even an overt realisation on the graphic level.

The aim in designing learning materials is to take into account the different processing routines – statistical ones and rule-based ones – and, above all, to try to integrate the benefits of multiword units for the learners in order to support them to develop the right processing strategies. A first approach would be to create more salience through the provision of multiword units in which attention is drawn to the feature to be acquired in order to achieve improved processing and thus more learning success. In the following, some examples of possible prototypical chunks are suggested to illustrate the idea of input manipulation that follows linguistic findings:

Following the suggestions of contextual optimisation (Motsch & Berg, 2017), vocabulary learning should concentrate more on article + noun sequences and additionally on giving as many opportunities as possible to process morphological cues by increasing the amount of target cues – gender should be overtly marked whenever possible and in several items and features. In doing so, learners can begin by acquiring chunks that can eventually be transformed into rules or smaller units; in this case, morphological language knowledge would be stored in the course of their acquisition.