Distributed representations of words play a crucial role in many natural language processing tasks. However, to learn the distributed representations of words, each word in the text corpus is treated as an individual token. Therefore, the distributed representations of compound words could not be directly represented. In this paper, we introduce a recurrent neural network (RNN)-based approach for estimating distributed representations of compound words. The experimental results show that the RNN-based approach can estimate the distributed representations of compound words better than the average representation approach, which simply uses the average of individual word representations as an estimated representation of a compound word. Furthermore, the characteristic of estimated representations of compound words are closely similar to the actual representations of compound words.