[17] investigated the use of structural support vector machines (SSVMs), an algorithm that combines the advantages of both CRFs and SVMs, and word representation features, which contain word-level backoff information over large unlabelled corpus by unsupervised algorithms, for clinical entity recognition.