Knowledge-based approaches to producing large-scale training data from scratch for Word Sense Disambiguation and Sense Distribution Learning