Word Learning

Between birth and adulthood children must learn about 60,000 words. That translates to 8-10 words per day! Words rarely occur in isolation, many of them do not have observable meanings (what does the word "the" mean?") and even when they do it's not always clear what the speaker is referring to.

Because of this challenge, a lot of researchers have posited specialized mechanisms that help children analyze the auditory and visual inputs and learn what words mean. However, it is unclear what those mechanisms are and even if they are necessary. Recently the MACLab has begun a collaboration with Larissa Samuelson and Jessica Horst to understand on such mechanism: fastmapping. In fast-mapping situations children use words they know to help learn new words. Our investigations build on important behavioral studies by Horst & Samuelson (in press) that demonstrate that finding the referent of a new word may be distinct from actually learning the word. Our work extends this by building a computational model that instantiates this principle, without any specialized mechanisms. Word learning model

If such learning mechanisms are not required, though, how do you account for the fact that children's learning accelerates so dramatically in the first year of life, the so-called "vocabulary explosion"? Typically after learning the first 50 words or so many children's word learning literally takes off as they start learning 8-10 words per week. Researchers have long assumed that mechanisms like fast-mapping are required to explain this word spurt, but computational work is (soon to appear in the journal Science) investigating this acceleration and finding that such mechanisms may not be required. (More details will be available when the paper is published--in the meantime, contact Bob McMurray for more information about this project).

If specialized mechanisms are not required to explain the acceleration, what is responsible? Currently we hypothesize that it primarily derives from the number of easy and hard words in the language. Easy and hard, of course, are hard to define--they could derive from the sound pattern of the word (e.g. a Vocabulary explosion word like "blast" may be harder than "bob"), the part of speech (noun/verb/adjective), how often the word appears, and what contexts it is used for. Most interestingly, this difficulty could arise from the properties of the child--the kinds of experiences they have, and the skills that they bring to the table. Work in progress with Larissa Samuelson is trying to pin down exactly this question.

Finally, classic explanations of the word spurt have assumed that as you learn one word, you can use that knowledge to learn the next. Thus, every word would allow the system to speed up, just a little. However, there is at least one place where this is not the the case: in speech perception. Pioneering work by Janet Werker has shown that learning a new word like "buk" makes it harder to learn similar ones like "puk". The MACLab has been investigating this with Gwyneth Rost of the department of speech pathology, to understand how the sound patterns of the words make this difficult, and the learning processes that might overcome it.

Word Learning Projects

Modeling Fastmapping

“Fast Mapping” refers to children (and adults!) ability to quickly link a novel name to a novel object. As an example, consider a situation in which a child is presented with a duck, a novel object and a car. The child is asked to get the fode. By around two years of age, many children can correctly infer that since duck and car have names, fode must refer to the novel object.

In the developmental literature, fast mapping like this is often cited as one of the mechanisms by which young children learn words so rapidly. However, recent research by Horst and Samuelson (in press, Infancy), suggests that fast mapping is more complex than once believed.

Horst and Samuelson tested 2-year-old children’s fast mapping abilities by presenting them with several objects and asking them to pick out either an object for which they knew a name or one for which they didn’t. For example, children were presented with a duck, a car and a bird-toy and asked to “get the duck” or “get the fode.” Children were very good at both the known and novel names. However, five-minutes later, children were presented with only novel objects and now failed to show any evidence that they remembered which object was the fode. This finding stands in stark contrast to the majority of literature on children’s word learning.

This is a somewhat difficult finding to digest--infants clearly can identify the target, but don't seem to be remembering anything about it. An important question is whether they are learning anything at all? It is possible that they are in fact learning nothing--that, their in the moment behavior is completely dissociable from learning. However, an subtler alternative is that they are learning something, but that learning is slow--it may ultimately take hundreds of exposures to really learn the name. Since this can be difficult to examine in the lab, we have developed a neural network model of this problem to help understand the relationship between learning and online processing. In these networks input units represent either names or objects. Like the children, the networks are presented with three objects and one name and we look at which object the networks fixate on. Unlike the children, however, we can go deeper and explore the amount of learning the networks undergo….

This research has demonstrated that learning does occur during a fast mapping trial, however, the amount of learning from a single fast mapping trial is insufficient to support full-blown word learning (see Horst, McMurray and Samuelson, 2006). This work also shows that-with enough exposure-associative learning can lead to the full-blown word learning often accredited to fast mapping. Finally, the associative learning can explain how fast mapping in a single trial and word learning over many trials are related. New directions for this research include modeling changes to the input and other aspects of the task.

References
Horst, J., and Samuelson, L. (in press) Fast Mapping But Not Fast Word Learning: Referent Selection Without Retention. Infancy.

Papers
Horst, J.S., McMurray, B., and Samuelson, L.K. (2006) Online processing is essential for learning: Understanding fast mapping and word learning in a dynamic connectionist architecture. The proceedings of the 28th meeting of the Cognitive Science Society.

Graduate Alumni
Jessica Horst 		Collaborators
Larissa Samuelson

Back to top

Speech perception and early word learning

By around a year of age, infants have pretty sophisticated phonetic abilities. They can discriminate the sounds of their language (and fail to discriminate other sounds), and segment words from streams of speech. They can recognize speakers, and seem to be able to account for differences in the way people pronounce the same sounds. Most importantly, it is about this time that infants start learning words.

A critical question, then, is how these early phonetic abilities interact with infants early word learning. Christine Stager and Janet Werker conducted a seminal study on this topic (Stager & Werker, 1997). They used the Switch Task in which infants see on object paired with a novel name, and a second object paired with a novel name. After they habituate to this, the name/object mappings are switched. If infants have learned the names they should be surprised and look longer. On the other hand, if they have failed to learn the names, they should continue habituating.

Stager and Werker discovered that while 14 month-olds can learn dissimilar words like "lif" and "neem" fairly quickly in this task, they could not learn similar words, "bih" and "dih.". Nonetheless, outside of a word learning task, infants at this age can discriminate /b/ from /d/ just fine. This would seem to imply that all of the phonetic abilities acquired in the first year are not being brought to bear on this word learning task.... or does it?

Work in the MACLab by Gwyneth Rost is challenging this assertion by examining infants performance when the training consists of multiple voices. If word learning is over-taxing these infants phonetic skills, this ought to make the task even harder, and result in worse learning. An alternative possibility, though, is that infants phonetic categories are not entirely developed. If this is the case, providing multiple speakers provides just the sort of information infants may need to quickly acquire the needed phonetic distinction.

Read the paper to find out how this one ends!

Papers
Rost, G., and McMurray, B. (in press) 14-month-old infants can learn lexical neighbors: speaker variability augments slow phonological development. Developmental Science     Email for a copy

Presentations
Rost, G., and McMurray, B. (2008, March) Phonetic variability and early word learning. In McMurray, B. (Chair) The emergence of phonological contrast from lexical and perceptual processes. Symposium presented at the International Conference on Infant Studies, Vancouver, BC.     PDF

Rost, G., and McMurray, B. (2008, July) Phonological variability and word learning: Infants can learn lexical neighbors. In A. Owen (Chair), The role of input variability on language acquisition and use. Symposium presented at the XI International Congress For The Study Of Child Language (IASCL): Edinburgh, Scotland. PDF

Students
Gwyneth Rost

Back to top



Site maintained by Bob McMurray, Cheyenne Munson, and Joe Toscano.