Deep Encodings vs. Linguistic Features in Lexical Complexity Prediction