doi dblp MRECN: mixed representation enhanced (de)compositional network for caption generation from visual features, modeling as pseudo tensor product representation Chiranjib Sur 2020 Volume 9 Issue 4