Learning to predict or predicting to learn?

ABSTRACT

Humans complete complex commonplace tasks, such as understanding sentences, with striking speed and accuracy. This expertise is dependent on anticipation: predicting upcoming words gets us ahead of the game. But how do we master the game in the first place? To make accurate predictions, children must first learn their language. One possibility is that prediction serves double duty, enabling rapid language learning as well as understanding. Children could master the structures of their language by predicting how speakers will behave and, when those guesses are wrong, revising their linguistic representations. A number of prominent computational models assume that children learn in this way. But is that assumption correct? Here, we lay out the requirements for showing that children use “predictive learning”, and review the current evidence for this position. We argue that, despite widespread enthusiasm for the idea, we cannot yet conclude that children “predict to learn”.

KEYWORDS:

• Language acquisition
• language processing
• prediction
• error-driven learning
• dialogue

Acknowledgements

Many thanks to Alex Doumas for helpful comments on the manuscript.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. Note that Fernald's work raises a potential alternative route by which prediction could influence language development. So far, we have contrasted views in which prediction is specifically used as a learning mechanism, with views in which prediction is a characteristic of more expert systems and is not used for learning. But Fernald's work suggests that, under the expert system account, prediction could still facilitate language acquisition, by increasing processing speed and thereby acting as a crutch for acquiring knowledge of words and grammar. Prediction can therefore facilitate learning even if children do not use predictive, error-driven learning.

Additional information

Funding

The work was supported by Research Project Grant RPG-2014–253 from the Leverhulme Trust (to H.R. and M.P.), and ESRC Future Research Leaders award ES/L01064X/1 (to H.R).