A Corpus-Based Computational Model ofMetaphor Understanding Incorporating Dynamic Interaction

摘要

The purpose of this study is to construct a computational model of metaphor understanding based on statistical corpora analysis. The constructed model consists of two processes: a categorization process and a dynamic interaction process. The model expresses features based not only on adjectives but also on verbs using adjective-noun and three types of verb-noun modification data. The dynamic interaction is realized based on a recurrent neural network employing differential equations. Generally, in recurrent neural networks, differential equations are converged using a sigmoid function. However, it is difficult to compare the estimated meaning of the metaphor to the estimated meaning of the target which is represented with conditional probabilities computed through statistical language analysis. In the present model, the differential equations converge over time, which makes it possible to compare the estimated meaning. Accordingly, the constructed model is able to highlight the emphasized features of a metaphorical expression. Finally, a psychological experiment is conducted in order to verify the psychological validity of the constructed model of metaphor understanding. The results from the psychological experiment support the constructed model.