Modeling Bilingualism

When children are brought up speaking two languages, they often go through a stage of 'mixing' where they appear to be unable to separate their languages. For instance, a Welsh word might be inserted into an English sentence: As an example, when I first realised the implications of death, my parents told me that I cried and said "I don't want to go into the pridd" (earth, dirt).

Several theories have been put forward to explain this. Firstly, I may simply not have known the word for 'dirt', and had to rely on a word in another language. Back then, Welsh was probably my stronger language, so this would be an example of mixing into my weaker language. Alternatively, I had not yet learned to tell the difference properly between Welsh and English.

However, both my parents speak Welsh and both languages are used, probably with quite a lot of mixing. Therefore, I may have known the English word, and been aware that I was mixing, but I knew that using a bilingual code was permissible, given my interlocutors.

Indeed, Montanari (2008) finds that the child she studies mixes some words even when they know the word in the language of context. Does this suggest, then, that the child simply didn't know which words belonged to which language? I argue that this isn't necessarily the case.

Adults mix their languages for many reasons. In fact, it's often difficult to decide which language a word belongs to without a lot of context (e.g. 'zeitgeist'). Let's forget about languages for a minute and ask 'to what extent has the child acquired the communicative code of its parents'? By this, I mean how closely does the child's output mirror the parent's input?

To do this, let's look at Quay's (2008) study of a trilingual child. Japanese is the language of the environment, the father is strongest in English and also speaks Japanese and the mother strongest in Chinese and also speaks English and Japanese. Weekly recordings were made from 1;10 to 2;4 years. The utterances of both the child and the parents were coded along with the addressee. The summary of the data is very detailed - containing the proportions of mixing between any two people in Japanese/English, Japanese/Chinese, Chinese/English and Japanese/Chinese/English.

Let's model the child's mixing proportions as a function of the parent's mixing proportions. Each cell in the table below contains the correlation between the model’s predictions and the child’s actual mixing proportions. The first two models use the mother and father’s data separately. The third model is an additive model which combines the parents’ utterances and the fourth uses the difference between the parents’ mixed utterance types. The difference model was provided as a conceivable, but unlikely model. The correlations in the first column correspond to a model using the total input, whereas the last two columns correspond to a model using only utterances directed to the child (direct) and utterances directed to the other parent (indirect).

Although the mother spends more time with the child than the father, the total mixing behaviour of the child is equally predicted by the mother and the father. However, the best model is an additive model of the direct utterances to the child. That is, the child's output is closest to a model which tries to imitate the mixing behaviour of both parents.

Interestingly, the highest correlation between the mixing proportions is between the parents (0.999), which is nearly perfect. Perhaps, then, the child is simply trying to acquire the adult’s mixing strategies or 'Code'.

We can look at the data in more detail by calculating the correlations between mixing proportions for each interlocutor separately:

When addressing the mother, the child's mixing proportions reflect the mother’s total mixing proportions better than the father’s and vice versa, indicating pragmatic differentiation to each parents’ mixing. When addressing the father, the child’s mixing proportions reflect the mother’s indirect input. This could indicate that the child is mimicking the mother’s interaction with the father. The opposite isn't true, but any mimicry may be masked since the child spends so much time alone with the mother.

These two analyses conclude that the child’s mixing reflects the mixing of the parents from a very young age. Modelling allows us to gain extra insights on the potential learning mechanism for the child, but it relies on detailed data, as in Quay (2008). The model could be taken further to include considerations of location, the societal status of each language and the parent's tactics (Negative evidence, implicit allowance of mixing, teaching of translation equivalents etc.).

Now for the ambitious, unfounded part: Considering a communicative code, there may be no qualatative difference between mono- and bi-lingual language acquisition. How, then, do bilinguals select words? One possible solution is to use a sort of Bayesian probability distribution over the linguistic, social and pragmatic contexts for each word that represents the best estimation of when to use a word. If a mapping between words and pragmatic and social contexts is acquired, a discrete mapping between words and ‘languages’ becomes irrelevant. This approach works equally well for acquiring one ‘language’, or several levels of tone or dialect.

In this sense, the ‘remarkable’ ability to keep languages separate (Costa & Santesteban, 2004) seems less remarkable and less specific to bilinguals: We don’t find it remarkable that an adult refrains from using terms of endearment during a boardroom speech.

This approach would be extended to syntactic acquisition by assuming that, as the mapping between words and meanings developed, strings of words themselves became a context which was encodable in the probability distributions of words. This is essentially a constructivist approach to bilingual acquisition: Before linguistic acquisition, infants first learn an embodied perceptual ‘language’ – an iconic mapping between form and meaning – which allows them to relate structure in the world to an interaction between sensory and motor activity. The mapping between structure in the world and symbolic, linguistic representations would build itself on top of this system in the same way as syntactic (Bernardini & Schlyter) and lexical (Nicoladis & Secco) acquisition can build on pre-existing structures.

Following from this, the ‘difficult’ bit of language acquisition is not the segmentation of strings into words or words into lexicons, but the initial segmentation of the world into functional concepts. The development of this more fundamental understanding of the world may be an additional factor in the qualitative differences between mixing in children and adults.