Modeling medical texts for distributed representations based on Skip-Gram model

Medical text is complex while its semantic expression is variable. Producing a vector for each medical word to encode its semantic information, as the cornerstone of medical semantic understanding, is quite different from the tasks of natural language processing. In this paper, we focus on the vectorization of Chinese medical text words. To recognize Chinese medical entities, a framework is constructed for Chinese word segmentations by jointly using a built medical dictionary and a trained hidden Markov model. Then, each medical text word is learned by Skip-Gram model into a 128 dimension vector for the distributed representation, and the robustness of the vectorization model is strengthened by negative sampling. Comparing with the one-hot method, the word vectorization method based on Skip-Gram model can improve the performance of the CNN classifier significantly due to the data sparsity and vector dimension decreased. The accuracy is also improved about 15% than that of the one-hot method.