Modelling Salient Features as Directions in Fine-Tuned Semantic Spaces