Still a Pain in the Neck: Evaluating Text Representations on Lexical Composition