Bilingual Semantic Network Construction

This article proposes a neural network for building Chinese and English semantic resources connection. Abundant monolingual semantic information is stored into its bipartite graph structure respectively. Two hidden layers are also set in every part, word layer and concept layer. Every word associates with different concepts separately; every concept includes different vocabularies; and these two layers also independently connect to their counterparts through bipartite graph. These distributed characteristics in hidden layers meet the need of parallel network computing. The unsupervised method is used to train the network, and samples are translation lexicons, results of the bilingual word-level alignment algorithm. The training principle comes from the inspiration of bilingual semantic asymmetry. Every translational equivalent contains the unambiguous information by comparison between source and target languages. These translation lexicons are viewed as a kind of special context. They almost have definite meaning. Every input will activate and suppress various kinds of potential connections by the interaction of hidden layers, and modify their connective weights. Finally a demo test presents.