Chinese Word Meaning Acquisition On Visual Information Based On Hellinger Distance

Cognitive science research suggests that perceptive information have an important supporting role on natural language acquisition. Because the machine can't independently associate words and perceptive world, so grounded semantics appeared, that is a research of the connection between language symbols and perceptive symbols.ViMac system (Visual Information based Meaning Acquisition of Chinese Words) is the first built pictures to talk system based on the simple geometry pictures and build the system of Chinese lexical semantics based on visual features. The Chinese lexical semantics which ViMac system output can be used by an automatically generate image descriptions system which called ViMacs system (ViMac Application System). ViMacs system can output some natural language description statement about a number of images which contain simple geometric description, that is, realize visual information to the expression of natural language information mode transition process.In ViMac system, distance metric have played an important role in several key stages of the construction of visual information's lexical semantics. Distance metric is an important factor affecting system performance, the current distance metric is the KL distance metric, but KL distance metric has some problems, can't fit for the needs of ViMac system well.Based on the analysis of the learning processes of ViMac system a and analysis the problems of the distance measure of ViMac system, and compare a variety of distance measurement algorithm, this paper proposed by using Hellinger distance as the distance measure metric. Using Hellinger distance to improve the key technology module of ViMac system, and extended one-dimensional Hellinger distance to multi-dimensional to improve ViMac system feature selection module. System evaluation results show that the Hellinger distance as a new distance metric effectively improved ViMac system's performance. This paper also improved the semantic modeling's performance; apply k-nearest neighbor semantic modeling instead of the Gaussian semantic modeling. The results show that, k-nearest neighbor semantic modeling effectively improved ViMacs system's performance.