TRIE++: Towards End-to-End Information Extraction from Visually Rich Documents