How infants learn language – III

A very complicated and very clever experiment used Harvard and Penn undergrads to study how infants learn language. It’s worth plowing through the whole paper [ Proc. Natl. Acad. Sci. vol. 108 pp. 9014 – 9019 ’08 ] as I can only summarize the main points here.

The work totally destroyed the current leading theory which is that infants form hypotheses about what words mean which are confirmed or denied by subsequent exposures to the word. This works pretty well when 4 – 5 objects are presented at a time on a computer screen along with a nonsense syllable. Multiple exposures to 4 – 5 other objects with other nonsense syllables, sometimes including the original syllable confirms or denies the infant’s hypothesis. Experiments done this way tend to confirm the leading theory (which is why it is the leading theory).

The problem with the leading theory is that real life isn’t like that at all.  A picture is given of a fairly typical middle-class living room with toys, furniture, books, pictures all over the place — literally hundreds of objects.  Then there is the problem of verbs (which can be shown by actions, but usually aren’t) AND  the problem that some objects that a word refers to aren’t all that similar (the different breeds of dogs) AND the problem that many objects referred to in speech aren’t physically present (let’s go to the store) AND the problem that most utterances contain many words, few of which are understood by the infant (and on and on and on).   You get the idea.

The authors exposed 37 undergrads to randomly selected video vignettes of naturally occurring contexts (the meal, the play room) in which parents utter words to 12 – 15 month old infants.  One vignette showed a mother opening a bag of toys (in a room with a lot of other stuff) saying “who else is in the bag”.  Participants watched the vignettes with audio muted and just a beep or a nonsense word when bag appeared in the sentence.  Why did they mute the audio — because infants at this age don’t understand many words.

The vignettes differed in informative content — e.g. a parent staring at something the kid was holding and saying ‘this is a horse’ has a lot of information, vs. ‘Let’s go see the horses at the farm’ when no horses are in sight.

The undergrads were supposed to guess the word (all of which are frequently used in parental speech to infants)  They didn’t do well.  There were 288 vignettes, only in 7% of them did the undergrads achieve over 50% accuracy (as a group) in guessing the word.  These vignettes were called highly informative (HI).  90% of the vignettes  had an accuracy score of under 33% — this is called lowly informative (LI).  As you might think, none of the HIs were for verbs (half of the 288 vignettes) — they are far more abstract than nouns.

Interestingly, they ran the same experiment with 12 children ranging in age for 3 to 5, and got the same levels of accuracy (and they say an Ivy League education is only for the best and the brightest). The kids were no worse than the undergrads (but at least they were no better).

Then they took a new bunch of undergrads and exposed them to 5  vignettes, each of which had the same word to guess.  Some series of vignettes had only 1 HI and 4 LIs, others had 5 LIs.  Subjects were asked to write down their guesses as to what the beep or nonsense word meant after each vignette.

The really interesting part was in the group exposed to 1 HI and 4 LIs.  The 1 HI could be first, third, or fifth in the presentation of the vignettes.  According to the statistical correlation theory this should make no difference.  If a hypothesis is formed and remembered for each vignette, as more and more vignettes are seen there is more information making it more likely to confirm the hypothesis.  That is not at all what happened.

If HI was presented first in the series of 5 the accuracy of the guesses was 40%, if it was presented third or fifth, the accuracy was 20%.  So it looks as if instead of forming multiple hypotheses and confirming or rejecting them based on subsequent information, the students form a hypothesis based on the first thing they see, and don’t form others, just confirming or rejecting it with further experience.

This does make sense (or rather the multiple hypothesis theory doesn’t make sense).  The circumstances of word presentation are too varied and the backgrounds too rich to form a zillion hypotheses, remember them all and weigh each new piece of information against all of them.

There’s a lot more in this paper.  The really hard part of psychology is thinking of the correct experiment to do — doing this one must have been simplicity itself.

Anyone trying to learn a new language as an adult knows that this is an impressive intellectual feat. If you find the way infants learn language interesting, you might be interested in two other posts on the subject.

Post a comment or leave a trackback: Trackback URL.


  • luysii  On March 4, 2012 at 10:54 am

    [ Proc. Natl. Acad. Sci. vol. 109 pp. 3253 – 3258 ’12 ] Apparently this is news to language researchers, but infants know the meanings of many common nouns at 6 – 9 months (don’t language researchers have kids?). The current theory was that word learning is possible only when infants can grasp a speaker’s referential intentions and understand language as a motivated communicative activity. Since infants only begin to understand intentional activity at 9 months, it was thought that they wouldn’t understand words till then. Fortunately, someone decided to look

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: