Exploiting distributional semantics to benefit machine learning in automated classification of Chinese clinical text

Abstract

Machine learning has been widely employed for the automated classification of clinical text to enhance the utilization of clinical information and benefit clinical applications. However, the conventional approaches for the vector representations of text in machine learning algorithms cannot model the connections between highly similar words and will also lead to high dimensionality. This study suggests a combination of distributional semantics and supervised classification algorithms in order to tackle these problems. Latent Semantic Analysis, Random Indexing and word2vec are adopted for distributional semantic representations to build classifiers using Support Vector Machines, Naive Bayes and k-Nearest Neighbors. As an initial study, we adopt Chinese diagnostic phrases as the clinical text to be classified and the 3-digit ICD-10 codes as the class labels. The evaluation results demonstrate that distributional semantic representations can better capture the meanings of text and can improve the accuracy of clinical text classification when the data for training and testing come from different sources and share less consistent language use. Consequently, distributional semantics can enhance the extensibility of the classifiers for clinical text classification.