Developing a psychologically realistic generalisation mechanism within MOSAIC

点击翻译按钮获取中文摘要

How do children learn the grammatical categories of their language? For example, how do children learn that "dog" is a Noun and "chase" is a Verb, and work out that in a novel utterance such as "The wug is meeking the blik", "wug" and "blik" are Nouns and "meeking" is a Verb? Recent research with computer models of language learning has shown that one way of doing this is to group words together on the basis of the words that come before and after them. For example, in English, words that come after "a" and "the" and before "is" and "can" tend to be Nouns, whereas words that come after "is" and "can" and before "a" and "the" tend to be Verbs. However, at the moment, computer models that group words together in this way tend to be unrealistic as explanations of human language learning because they do not learn gradually like children, and so cannot mimic the behaviour of children at different points in development. The aim of this research is to develop a more child-like model of category learning. This will be done by taking ideas from recent computer models that group together words at a single point in time, and building them into a model called MOSAIC that learns language more gradually. MOSAIC is a computer program that takes as input speech directed at language-learning children, and produces as output child-like utterances that get longer as the model processes more input. These utterances can therefore be compared with those of children at different points in development. In the current project, we will use MOSAIC to develop a more realistic way of grouping words into categories in four ways. First, we will measure differences in the productivity of different categories (e.g. Noun and Verb) in transcripts of children's speech. This will give us a description of differences in how much children know about particular categories at different points in development, which a realistic model of language learning should be able to mimic. Second, we will build several different versions of the model that group words into categories in different ways. Since the only difference between these versions of the model will be the way in which they group words into categories, this will allow us to see which grouping method is most effective. Third, we will evaluate the different versions of the model by comparing their performance with respect to each other. This will be done in two ways. First we will look at how good the models are at grouping words into the kind of categories used by adults. Second, we will look at how good the models are at mimicking the pattern of generalisation in children's speech at particular points in development. Finally, we will look at whether it is possible to get better results by taking account of how much information humans (and particularly young children) are able to deal with at any one time. An interesting feature of MOSAIC is that one of the reasons why it learns so slowly is that, like humans, when learning from an utterance it hasn't heard before, it can only deal with information at the beginning and the end of the utterance. MOSAIC can therefore be used to look at whether limiting the amount of information that is used to group words together makes the output of the models look more realistic.