Error and expectation in language learning: The curious absence of mouses in adult speech

Abstract

As children learn their mother tongues, they make systematic errors. For example, English-speaking children regularly say mouses rather than mice . Because children’s errors are not explicitly corrected, it has been argued that children could never learn to make the transition to adult language based on the evidence available to them, and thus that learning even simple aspects of grammar is logically impossible without recourse to innate, language-specific constraints. Here, we examine the role children’s expectations play in language learning and present a model of plural noun learning that generates a surprising prediction: at a given point in learning, exposure to regular plurals (e.g. rats ) can decrease children’s tendency to overregularize irregular plurals (e.g. mouses ). Intriguingly, the model predicts that the same exposure should have the opposite effect earlier in learning. Consistent with this, we show that testing memory for items with regular plural labels contributes to a decrease in irregular plural overregularization in six-year-olds, but to an increase in four-year-olds. Our model and results suggest that children’s overregularization errors both arise and resolve themselves as a consequence of the distribution of error in the linguistic environment, and that far from presenting a logical puzzle for learning, they are inevitable consequences of it.

Learning to predict or predicting to learn?

“…Ramscar, Dye, and McCauley (2013) proposed a much simpler model, based on ideas from learning theory, that attempts to explain how children master morphology (rather than syntax). In particular, the model attempts to learn which of a scene's semantic features are the best predictors of its appropriate phonological label (e.g.…”

“…1 The successful completion of this enterprise will be greatly helped by the explicit computational models that have been developed in this area (e.g. Chang et al, 2006;Ramscar et al, 2013). These can generate accurate hypotheses about what children predict, and also about how children should behave when they receive evidence that mismatches their predictions.…”

“…The theoretical clarity and computational precision of the models described in the second section are impressive, and they will clearly play an important part in solving the puzzles posed by children's predictions. However, we would like to emphasise that these are not the only possible accounts by which prediction-based learning could help children acquire language, and that evidence for prediction-based learning should not necessarily be taken as evidence that children learn language in a manner that is either connectionist (Chang et al, 2006;Elman, 1990) or neo-behaviourist (Ramscar et al, 2013).…”

“…In particular, recent research into reinforcement learning has provided evidence for the so-called "model-based" reinforcement learning approaches, in which learners acquire explicit and structured representations of the environment that they are learning about (such as an explicit map of a spatial environment, Daw, 2012;Daw, Gershman, Seymour, Dayan, & Dolan, 2011;Doll, Simon, & Daw, 2012;Otto, Gershman, Markman, & Daw, 2013). These models are acquired via prediction about future states of the world followed by adjustment from an error signal, just as in the "model-free" approaches of, for example, Ramscar et al (2013). Model-based approaches are particularly promising because, when some of the basic parameters of the model can be specified in advance, then it is easy to learn an accurate representation of the world from relatively little data (Dayan & Daw, 2008;Doya, Samejima, Katagiri, & Kawato, 2002).…”

A learning perspective on individual differences in skilled reading: Exploring and exploiting orthographic and semantic discrimination cues.

“…This way we can differentiate between participants who are slowing down (those above the horizontal line) and those who are speeding up (those below the horizontal line) as they progress through the SRT trials. 7 Ramscar, Dye, and McCauley (2013) demonstrated that discrimination learning can have different consequences depending on the specific conditioning histories Ð i.e., earlier or later in learning, specific exposure can lead to diverging predictions and, consequently, behaviour. To control for this effect we randomized trials.…”