The power of prediction in language comprehension

predicting conversations

How good are we at predicting upcoming words during a conversation? If someone begins a sentence with the words 'You never forget how to ride', you'll be surprised if it doesn't end with 'a bike'. But if the sentence begins with 'You never forget how to ride an', then, phonologically, the bike is out of the question – but an elephant isn't.

In a new study published in the journal eLife, an international team including Oxford's Dr Matt Husband investigated, using neuroimaging techniques, whether our brains have the capacity to make very specific predictions about upcoming words, such as their initial sound.

Dr Husband, from the Language and Brain Laboratory in Oxford's Faculty of Linguistics, Philology and Phonetics, talks to Arts Blog about the research.

What was the prevailing wisdom on our ability to predict upcoming words?

For the last decade, our ability to predict upcoming words has become integral to our understanding of language comprehension. The idea that we predict upcoming words has gone from the fuzzy intuition that we all, more-or-less, seem to experience during conversations to a detailed mechanism proposing that we comprehend language in part by predicting what will be said next.

The key empirical questions of the last decade have moved from 'do we predict?' to 'what is it that we predict?' We might, for instance, only predict a very abstract, high-level meaning of what will be said, or our predictions could be quite fine-grained – perhaps even all the way down to the probability for what specific form a word will take.

Arguably the most high profile evidence for this 'specific word' view of prediction is DeLong, Urbach and Kutas' 2005 Nature Neuroscience paper titled 'Probabilistic word pre-activation during language comprehension inferred from electrical brain activity'. They investigated how the brain responded to the expectations set up by sentence fragments such as 'The day was breezy so the boy went outside to fly…' when it continued with the expected phrase 'a kite' or an unexpected phrase like 'an airplane'.

DeLong and colleagues cleverly manipulated the indefinite article in their study, taking advantage of a phonological rule of English in which the indefinite article is realised as 'a' before consonant-initial words and as 'an' before vowel-initial words. This allowed them to ask whether the brain was predicting the sound form of a word before actually seeing that word. They examined electrical brain activity elicited by articles that were comparable with the highly expected yet unseen noun ('a', followed by 'kite'), or by articles that were incompatible with the highly expected noun and heralded a less expected one ('an', followed by 'airplane'). They reported increased brain activity for articles that were incompatible with the highly expected noun compared with those that were compatible with the highly expected noun, suggesting that predictions could be very specific and fine-grained.

What does this new research find?

It is perhaps surprising given how foundational and widely cited the DeLong, Urbach and Kutas study is that a direct replication of this effect on 'a/an' had not been reported in the published, peer-reviewed experimental literature. The study reported in our paper is the first large-scale attempt to directly replicate these results, using both their original methods and analysis and new analysis techniques currently available to us. This was a massive effort, coordinated between nine UK university labs and collecting 334 participants – ten times what an experimental study of this type would normally have. The participants read sentences that were presented one word at a time, while electrical brain activity was recorded at the scalp. Each sentence contained an expected or unexpected combination of an article and a noun, such as in the kite/airplane example given earlier.

Surprisingly, even with such high statistical power, our study did not replicate the key finding of DeLong, Urbach and Kutas, suggesting that either the effect on 'a/an' is, in general, too weak to detect using current neuroimaging techniques, or that the predictions we make during language comprehension are not so specific as to single out a particular word.

What are the implications of this study?

Our failure to find the 'a/an' prediction effect stands as an important reminder that it should take a large body of research to fully convince a field to adopt a new theoretical proposal. The results of a single published study, even in a high-impact journal, need to be considered with care.

Our failure to find the 'a/an' prediction effect should also caution us when it comes to how detailed our predictions can be. It may be that we still predict broad meanings and perhaps some specific features of upcoming words, but very fine-grained, phonetically detailed predictions may be out of reach for the human mind.

The study was led by Mante Nieuwland, cognitive neuroscientist at the Max Planck Institute for Psycholinguistics (MPI) and the University of Edinburgh, and involved researchers from nine UK laboratories: University of Birmingham, University of Bristol, University of Edinburgh University of Glasgow, University of Kent, University College London, University of Oxford, University of Stirling, and University of York.

The Oxford laboratory research was supported by the John Fell Fund.