Children use analogy rather than rules when learning language

A lot of research into children’s language development is on English. But English is a peculiar language when you compare it to the languages of the world. Unlike many languages in which the order of words can be varied without changing the meaning, word order really matters in English (Dog bites man is much less interesting than Man bites dog!). Many other languages use changes to words to change meaning: we do this in English too but to a much lesser extent (so we change run to runs to indicate that we are talking about another individual and not about I , you, we or they). Other languages do this to a much greater extent.

Since children can learn any language they are exposed to from an early age, this means that they are using the same brain mechanisms as we use to learn English. Therefore in order to fully understand how these mechanisms work, it is really important to look at how children learn languages that are very different to English and present children with different types of complexity. This is what we set out to do by studying how Polish and Finnish children learn how to change verbs to change meaning.

We tend to see language as a set of rules that every speaker follows in order to be understood. But as everyone knows, there is no rule without an exception, and often it is not possible to formulate a rule at all. For example, when forming the English past tense, liked follows a rule of adding -d to the present tense form like, but sleeped is not correct and should be slept instead. There is a long tradition in linguistic research that assumes that learners have the predisposition to find a general rule like “add -d” for the majority of verbs, and treat exceptions such as sleep/slept, put/put, fight/fought in a different way. One possible way to deal with exceptions would be to just memorise all irregular verbs. Another way could be to come up with more tiny rules that apply to some specific words that sound similar. For example, for sleep/slept, there are similar sounding verbs – so-called neighbours – that build their past tense in the same way: keep/kept; creep/crept; weep/wept. Thus, knowing these examples, the past tense of sleep can be “guessed” by analogy.

In English, memorising exceptions or building smaller rules for irregular groups of verbs may seem possible. But the English verb system is very simple compared to other languages. For example, in the English present tense, with the exception of the third person singular (she like-s) all forms are the same (I like; you like; we like; they like). In Polish, however, each form has to be used with a different ending: lub-ię/-isz/-i/-imy/-icie/-ą. And these endings even differ for other verbs, depending on which conjugation class they belong to. Finnish, on the other hand, also uses different endings for each form (-n/-t/-u/-mme/-tte/-vat) but most of them are the same for all verbs. However, the word stem may change: While the second person singular is nuku-t, the third person is nukku-u.

For complex languages like these, it becomes hard to formulate a general rule or to group exceptions into classes with their own rules. Including differences between endings and changes to the stem, Finnish verbs would have about 15 classes and Polish verbs about 12, but the number of classes is not necessarily agreed on in either of the two languages. But how is it possible that children learn these and more complex languages just fine, despite the lack of generalising rules? If children are able to “guess” word forms by analogy across similar sounding neighbours like for sleep/slept, do we need to assume rules at all? In our study, we wanted to know whether children’s ability to learn verb forms – even for verbs that seem to follow a rule – can be explained purely by memorising the most common examples and using analogy across them to guess the less common forms.

We tested 77 Finnish and 81 Polish children between the ages of three and five years on their ability to produce the correct verb form that describes a short video sequence. For example, if the video showed someone dancing, the child would have to say “(he) dances”, whereas if a picture of the child’s face was pasted on the dancer’s head, he/she would say “(I) dance”, with a picture of the experimenter’s face it would be “(you) dance”, and so on. Then, in order to find out what information the children’s knowledge is based on, we analysed the errors they made.

Frequency and analogy

If children rely on memorising examples, we should find lower error rates for forms that occur often in the child’s environment compared to rare forms. If children use analogy between similar forms, we should find lower error rates for forms that have more similar sounding neighbours compared to forms with fewer neighbours. And indeed, we found both effects in both Finnish and Polish.

Types of errors

Not only when errors were made but also what kind of errors were made by the children points towards the use of memorising and analogy. When children did not know the required verb form they often used another form that they know and hear often (for example the 3rd singular ajattele-e “(he) thinks” in Finnish instead of the 1st singular ajattele-n “(I) think”). Another type of error that may happen is that children may get lured into using the wrong ending through false friends in the neighbourhood. For example, children used macz-asz instead of macz-esz “(you) dip” because the stem macz- has many similar sounding neighbours that use the ending -asz in the 2nd singular, although the less frequently occurring -esz would be the correct ending for this particular verb. This phenomenon is called “overgeneralisation”.

Our experiments seem to suggest that, instead of learning hard rules, children produce verb forms by using analogy across forms they hear often. And this seems to be the case not only for Polish, which has very irregular endings, but also for the more regular verb system of Finnish.

Neural networks

The next step was to test whether memorising and analogy is not only necessary but also sufficient for learning complex verb systems like in Finnish and Polish without the need for explicit rules. For this, we used a computational model called an artificial neural network. This kind of model is very useful for studying whether a task is learnable by nothing except a number of examples. While other models need to be told what to do, a neural network figures this out by itself. We trained it example by example, giving it a verb stem like macz and telling it to produce, for example, the 3rd person singular (“you…”). If the model did not produce maczesz, it was told that the output was incorrect. Over time, it learned to adjust in order to produce more and more correct words. In order to make the model’s input similar to the children’s learning environment, the examples and the model’s output were representations of their sounds. Also, individual forms that are found to occur more often in the children’s environment were also used more often in the training of the model.

That way the neural network only had the sound of a word and its frequency of occurrence for learning the Finnish and Polish present-tense verb system. Over training, the model learned to attend to the important similarities in the sounds of different words that are useful clues for producing the correct target form. With enough time, both the Finnish and the Polish verbs could eventually be learned correctly by a neural network. This means that it is indeed possible to learn these complex language systems just based on the similarity of sounds, without the predisposition to find a general rule.

Like the children in the experiment, the model made fewer errors on verbs that occurred more often and verbs that had more similar sounding neighbours. In addition, the types of errors the model made were very similar to the children’s. The model tended to use a more frequent verb form in place of a rare form or used the wrong ending due to false friends in the neighbourhood.

Conclusion

In summary, our study demonstrates not only that children seem to rely on analogy across similar sounding verbs they hear often in order to produce the correct verb form. We also showed that it is possible to learn complex verb systems using only similarities in sound, without any predisposition to learn a rule. This could mean that sound-based analogy is enough to explain even learning of the simple English verbs with no distinction between rules and exceptions. More research is needed to study to what extent a model of memorising and analogy is sufficient to explain not just the present tense or the past tense but an entire verb system, complex noun systems or language acquisition in general.

For more information about this research, see our paper:

Engelmann, F., Granlund, S., Kolak, J., Szreder, M., Ambridge, B., Pine, J., Theakston, A., and Lieven, E. (2019). How the input shapes the acquisition of verb morphology: Elicited production and computational modelling in two highly inflected languages Cognitive Science https://doi.org/10.1016/j.cogpsych.2019.02.001