The psychometric, behavioral, and neurological role of empirically-identified semantic components

点击翻译按钮获取中文摘要

My research interest is in understanding how words have meaning (semantics). In the past many explanations of semantics have been circular since they have been cast in terms of semantic rules, judgments, or features. I am particularly interested in naturalizing' semantics, e.g. explaining it in terms that are compatible with other work in the biological sciences and that do not make any special assumptions about language as special'. This application focuses specifically on studying the relation of human behavior to the principal components extracted from a computational model of semantics derived from predicting word context in a large corpus (a co-occurrence model). Co-occurrence models represent a target word's context in a large corpus of text as a vector (a string of numbers) that encodes the target word's context, thereby bootstrapping word meaning from word usage. In co-occurrence models, the distance between word vectors can be used to estimate semantic similarity of two words. Distances of a single word's vector from the average vector of multiple words from a single semantic category (which I call the category-defining vector, or CDV) are good measure of category membership. For example, the closest neighbours of the vectors for ten mammals will be other mammals. Recent co-occurrence models are prediction models, in which a surprisingly simple computational tool (a simple three-layer neural network) is used to predict a word's context. This model is closely related to an animal learning model, the Rescorla-Wagner model. The new prediction models are therefore exciting because they suggest that lexical semantics may be explicable using standard discriminant learning principles from comparative psychology. I will undertake a series of experiments that can help us understand how semantic models are organized, and whether their organization is mirrored in human psychological organization. These experiments look for either behavioral or neurological correlates of semantic components that have been derived mathematically (using principal components analysis) from a prediction co-occurrence model of language trained on a corpus of three billion words of text. The behavioral correlates are patterns of human response that may reflect the structure of the semantic components, i.e. experimental participants should be faster to make decisions about the words most strongly loaded on the most important axes of variance that we have found mathematically. The neurological correlates are patterns of brain activity that correlate with those axes of variance. The proposed research is important because it will provide tests of the behavioral and neurological plausibility of animal learning theory accounts of lexical semantics, thereby opening a path to unifying higher-order cognition with these well-established, fully-specified, and simple models of learning.