Back to the previous topicElman |
|---|
| Elman (1991), at the Center for Language Research, UCSD, worked with NN that did
different types of analyses and predictions with language. In one case, the networks received a
sentence, and then tried to determine if it belonged to a grammar they had been trained on. In
other experiments, the networks received a sequence of words, and had the task of predicting the
next word of the sentence. In both cases the input to the network was at the level of words, i.e.
words were presented one at a time. Words were represented by orthogonal strings of binary bits.
This model was able to effectively learn context-free and context-sensitive grammars. The
sentences used in this experiment fall into categories that linguists believe cannot be understood
without using abstract representations. Elman reached the conclusion that a high-level analysis
process was taking place inside the network:
It is reasonable to believe that in order to handle agreement and argument structure facts in the presence of relative clauses, the network would be required to develop representations which reflected constituent structure, argument structure, grammatical category, grammatical relations, and number . (At the very least, this is the same sort of inference which is made in the case of human language users, based on behavioral data. He also reported that the networks captured lexical category structures, and that the relevance to grammatical structure was still not fully understood. However, as Elman pointed out, these networks had acquired the ability to process a sentence lexically but had not been asked to perform any kind of semantic analysis. |
Back to the Table of Content
To the next topic