Cognitive Linguistics meets multilingual language acquisition: What pattern identification can tell us

3.1 Detecting chunks and frame-and-slot patterns in (bilingual) data

In a series of recent studies (e.g. Quick et al. 2021a, Quick and Hartmann 2021), we have applied one particular method originally developed for investigating monolingual acquisition to corpora that document bilingual language acquisition, namely the traceback method (e.g. Dąbrowska and Lieven 2005). In the traditional application of this method, a longitudinal corpus documenting the language acquisition of one child is split into two parts: The so-called test corpus, which usually contains the last two recording sessions, and the main corpus, which contains all previous recordings. The goal of the method is to show that the vast majority of the child’s utterances in the test corpus have predecessors in the main corpus, i.e. they are either verbatim repetitions (so-called “fixed strings”), or they can be accounted for with the help of “templates” that are partially lexically specific and contain an open slot, such as [I want referent], so-called frame-and-slot patterns. In order to detect frame-and-slot patterns, the traceback method uses a limited number of operations. Different implementations of the method use different operations; typically, three operations are used: superimpose, substitute and add, exemplified in (5) with examples from Hummel (2006).

In practice, a traceback analysis proceeds as follows: First, the analyst, or the algorithm, searches for verbatim precedents of each target utterance in the test corpus. If a match is found, the utterance counts as successfully derived, i.e. it is assumed that the child has it readily available. If no verbatim match is found, we search for partial matches, accounting for the target utterances with the help of the above-mentioned operations. If, for example, the target utterance is I want my banana, and the main corpus contains the utterances I want my apple and I want my pear, then we can postulate the frame-and-slot pattern [I want my X]. If, additionally, the item(s) that fill(s) the slot is attested in the main corpus – here: the lexical item banana –, then the utterance is seen as successfully derived. The motivation behind this approach is that if it can be shown that the later, usually more complex utterances can be accounted for with the help of fixed chunks and frame-and-slot patterns, this lends support to the usage-based account of language acquisition as sketched above.^[6]

A number of studies employing the traceback method have shown that a large proportion of children’s target utterances can be successfully derived from precedents in the main corpus (see Lieven et al. [2003]; Dąbrowska and Lieven [2005]; Lieven et al. [2009]; Vogt and Lieven [2010] for English data; see Miorelli [2017] for Italian; see Koch [2019] for German). These studies consistently find that a large proportion of utterances in the test corpus can be traced back to earlier utterances. This provides evidence in favor of the usage-based hypothesis that language acquisition is strongly item-based. In addition to testing this hypothesis, however, traceback has also been come to be used as an explorative method to identify patterns in children’s speech (Hartmann et al. 2021; see Kol et al. [2014] and Koch et al. [2020] for critical evaluations of the traceback method).

In their application of the traceback method to bilingual acquisition, Quick et al. (2018a, 2018b, 2018c, 2021a, 2021b), Gaskins et al. (2021), and Quick and Hartmann (2021) have focused on code-mixing, i.e. the use of more than one language in one single utterance. As mentioned above, code-mixing might, at first glance, be seen as a challenge to constructionist approaches to language acquisition, as mixed utterances like ich habe drei wheels ‘I have three wheels’ (Fion, German-English, from Quick et al. [2018b]) seem to be examples of highly productive and creative, rather than formulaic, language use. On closer inspection, however, we may find that highly entrenched patterns also play an important role in bilingual utterances that show code-mixing. Just as early monolingual language acquisition is highly formulaic and characterized by low-scope schemas which seem to function as a gateway to greater complexity, a usage-based approach to multilingual acquisition would expect that code-mixed utterances arise from the combination of entrenched chunks from different languages and/or the productive use of formulaic patterns with one or more open slots. This is precisely what the studies reviewed below have set out to investigate.

3.2 Participants

In the following, we will first focus on studies investigating the acquisition of three German-English bilingual children. All of them grew up in one-parent-one-language households, lived in Germany, came from a middle-class household, and are simultaneous German-English bilinguals. The first child, “Fion”, is the second child to a German-speaking mother and an English-speaking father. Although the parents mostly adhered to the OPOL strategy when they talked to Fion, they did not settle on a family language and sometimes used both languages interchangeably when all family members were present. Fion was recorded two hours per week, yielding ca. 200 hours of recording covering the age range from 2;3 to 3;11.^[7]

The second child, “Silvie”, had an English-speaking mother and a German-speaking father. The father’s proficiency in English was very limited and the parents therefore spoke German to one another. The corpus covers recordings from 2;4 until 3;10 years of age, 135 hours in total averaging to about 2.5 h of recordings per week.

Finally, “Tim” is the only child of an English-speaking mother and a bilingual Spanish-Catalan father. Technically, Tim grew up trilingually, with the mother speaking mostly English to him and the father mostly Spanish, while a babysitter with whom he spent ten hours per week from 0;5 onwards spoke German to him, and German was also the language of the kindergarten he attended for more than 30 hours per week from 1;3 onwards. Effectively, Tim was largely bilingual as his exposure to and use of Spanish was very limited, as English was used as the family language. Tim was recorded 4 to 6 times per week, covering a period of six weeks between the age of 2;6 and 2;7, yielding a total of 30 hours of recordings.

3.3 Traceback and multilingual data

In order to check whether code-mixed target utterances can still be accounted for with the help of component units from a different subcorpus, Quick et al. (2018c) carve up their data differently from the traditional traceback studies mentioned above: Studying Tim’s bilingual language use, they use the set of all code-mixed utterances as their test corpus, and trace each code-mixed target utterance to earlier monolingual and bilingual utterances. For example, the bilingual target utterance this is kaputt ‘this is broken’ contains the pattern [this is X], which is attested multiple times in earlier utterances, and the open slot is filled by a German element.

Among the patterns identified by the method are also bilingual units like und this, which the authors interpret as temporarily entrenched units. Of course this does not mean that the child only has these items available as fixed chunks (which would be odd, given that we can hardly expect them to occur in the input); rather, the child has used them often enough that they have “grown together”, as it were.

Overall, the traceback of lexically fixed items was successful in 88% of the cases, which suggests that lexically restricted patterns play an important role in the production of code-mixed utterances. The authors conclude that the form and production of highly formulaic constructions in code-mixing closely follows the entrenchment of the children’s own patterns, presumably extracted from the input – but also repeated frequently by the children themselves. Table 1 shows the most frequent patterns identified in Tim’s code-mixed data, showing that some multilingual utterances can be accounted for with the help of patterns with monolingual fixed parts like this is x, while in other cases, bilingual word combinations are identified as patterns because they are so frequent in the data, e.g. ich want x.

In another study, Quick et al. (2021a) used the traceback method in a slightly different way. Again, all code-mixed utterances formed the test corpus, and they used a) the child’s monolingual utterances and b) the child’s input as their main corpus.^[8] This study, which investigates the language use of “Fion”, draws on an automatic analysis of utterance-initial uni-, bi-, tri-, and 4-grams. This means that an utterance like und das da go to school (Fion, 02;05.29; see Hartmann and Quick [2021]) is split into chunks of 1 word (und, das, da…), 2 words (und das, das da, da go…) etc. up to 4 words. For the n-gram analysis, only fixed chunks are taken into account, i.e. no slot-and-frame patterns are posited. This necessarily entails a lower traceback success than could be expected if frame-and-slot patterns were taken into account as well. To make the procedure even more conservative, the authors excluded primed utterances: Patterns that are attested in the 20 preceding utterances (regardless of the speaker, i.e. whether it is an utterance of the child himself or one of his caregivers) were seen as potential primes and therefore not taken into account. Also, the authors operationalize two different thresholds: Following the above-mentioned CDS studies by Cameron-Faulkner et al. (2003) and Stoll et al. (2009), they only consider n-grams to be (potentially entrenched) patterns if they occur at least four times in the test corpus, not counting the potential primes in the 20-utterance window just mentioned. In their analysis, they distinguish between patterns that occur at least twice in the main corpus and patterns that occur at least once in the main corpus.

Even with the most conservative parameter settings, a fairly large number of target utterances can be traced both to the input and to the child’s monolingual utterances. Overall, 82.2% of Fion’s utterances are identified as containing utterance-initial patterns, i.e. units that are attested at least four times in the test corpus. However, this number also contains one-word patterns, i.e. utterances in which only the first word is identical four or more times. If we limit the scope to multi-word units, the percentage is much lower, but still, more than half of all utterances (51.7%) can be accounted for. Zooming in on the multi-word units only, 63.6% of the identified patterns are attested at least twice in the monolingual data and 65.2% are attested at least twice in the input data (Quick et al. 2021a).

Given the fact that we are dealing with highly non-standard patterns like und this or nein this is, this can mean one of two things: Either the results lend strong support for the hypothesis that language is highly formulaic – or they show that the traceback method is actually too permissive and that it identifies patterns where it makes no sense to assume an entrenched pattern. There are at least two arguments that speak against the latter suspicion: First of all, the traceback method is characterized by an “inbuilt conservativeness” (Quick et al. 2018b: 298), and the operationalization of the method in this particular set of studies ensures that it is applied in an even more conservative way than in the case of the classic traceback studies. Secondly, the implementation of the method is very much data-driven in these studies, i.e. without positing any patterns in a top-down fashion.

Quick et al. (2018b) take an in-depth look at a sample of the Fion data. They manually identify frame-and-slot patterns to investigate whether Fion’s code-mixed utterances are complete or partial repetitions of earlier utterances. As with the “Tim” data in Quick et al. (2018c) mentioned above, each code-mixed utterance was traced back to all utterances prior to the target utterance. Here, however, the traceback results actually only constitute the starting point for a follow-up analysis in which they investigate how the identified frame-and-slot combinations develop: For instance, they show that the predominant language in the open slot changes over time from English to German (and vice versa for the fixed lexical items in the fixed part of the frame-and-slot pattern), e.g. ich bin eine bee ‘I am a bee’ (Fion, 02;07.17) – and that’ s rueckwaerts ‘and that’s backwards’ (Fion, 03;11.24). This is in line with Fion’s linguistic development in general, which includes a shift in his dominant language from German to English (see Figure 2) after a trip to Ireland and a visit from his monolingual English grandparents (Quick et al. 2018b: 289–290).

Figure 2:

Language proportions in the “Fion” data, indicating a clear shift towards English around the age of 03;05.

Quick et al. (2018b) also go on to investigate discourse priming on the basis of their traceback results, showing that much of the code-mixing turns out to be primed by the occurrence of the same forms in the immediately preceding context, as in the two directly adjacent utterances in (6). In addition, they show that at an earlier age, the entire chunks tend to be primed, while at later age spans, only the open slots tend to be primed.

This suggests that different input situations lead to considerable individual differences in the children’s output. In general, the role of individual differences is a crucial aspect that has to be taken into account when investigating child language acquisition. It is therefore quite instructive to compare the traceback results obtained for different children. Quick and Hartmann (2021) do exactly this for Fion’s and Silvie’s data. They find that the traceback success differs considerably between the two children, as Fion’s data contain much more code-mixed chunks that he re-uses multiple times. Code-mixing is thus quite a heterogeneous phenomenon that shows considerable individual differences. Even though it was not originally conceived for this purpose, the traceback method can arguably contribute valuable insights on this aspect.

In a similar vein, Gaskins et al. (2021) apply the traceback method to the Silvie data as well as to the data of two other bilingual children: “Sadie”, a Polish-English bilingual, and “Eetu”, a Finnish-English bilingual. Gaskins et al. (2021) use the children’s code-mixed multi-word units as their test corpus and trace them back to the same children’s monolingual constructions. Unlike Quick et al. (2021a), they manually identify frame-and-slot patterns, and they show that between 63 and 65% of all code-mixed target utterances can be traced back to the monolingual data and that across all children in more than 90% of all cases, monolingual frames are combined with slot fillers from the other language. They also show that, although the three children, in frame-and-slot patterns, have clear preferences for using fixed parts from one language and slot fillers from the other, there is also a fair share of frame-and-slot patterns in the data where the opposite direction can be observed: “17% of Sadie’s frames were Polish; 4% of Sylvie’s were English and 9% of Eetu’s frames were Finnish” (Gaskins et al. 2021).

These examples show that the traceback method can provide interesting insights regarding bilingual language acquisition. The findings from Gaskins et al. (2021) and Quick and Hartmann (2021) also hint at various ways in which the method can be used to investigate individual differences. However, the method does have its limitations. It is, after all, based on arbitrary frequency thresholds, and it cannot take into account many of the more subtle aspects that play a role in the acquisition and use of constructional patterns. Thus, not all patterns detected by the method are cognitively plausible to the same extent. For instance, [nein X] is identified as a frame-and-slot pattern on a par with, say, [this is X], although it might be more straightforward to assume that nein is a “pivot” in the sense of Tomasello as explained above, which is flexibly combined with other linguistic material. Another limitation, at least in the way the method is implemented in the studies reviewed above, is that it operates at the level of words and (adjacent and non-adjacent, but always utterance-internal) word combinations. Thus, the results are only informative about constructional patterns at the syntactic level, although it would also be interesting to take morphological constructions into account as well, considering code-mixed utterances like firewehr ‘fire brigade’ (Fion, 02;05.24) or I’m suching my Uhr ‘I’m looking for my watch’ (Fion, 03;08.10).

The traceback results therefore have to be interpreted with some caution. Nevertheless, traceback provides one way to reach one of the central goals of usage-based approaches to language acquisition: It helps identify recurrent patterns and, by proxy, constructions. As such, it also paves the way for further analyses. We have discussed some examples above, but there are a number of other questions that could be pursued, particularly with regard to the frame-and-slot patterns. For instance, it would be interesting to investigate whether the slot fillers of a particular frame-and-slot pattern share certain formal features (e.g. their prosodic pattern), and whether items from the other language are more likely to occur in the open slot of frame-and-slot patterns in code-mixed utterances if they exhibit those features as well.