Specializing Unsupervised Pretraining Models for Word-Level Semantic Similarity