Compounds are very common in many kinds of language. Most of the research in this field is from the view of morphology, while artificial neural network is seldom concerned. Based on Hopfield model, we create a novel neural network to simulate the recognition process of compounds in English and Chinese. Our model is composed of two layers: abstraction layer and recognition layer. The first layer can extract the common features of the training samples and represent it as a new attractor, which can be transferred into the next layer. This step imitates morpheme abstraction of compounds. Recognition layer is constructed as an improved Hopfield network, in which two existing attractors can merge into a new one. This step reflects the cognition of a new compound when all the morphemes are memorized. One specific example ‘raincoat’ is demonstrated, and the results provide strong evidence to our model.