Using RL for Non-Greedy Dependency Parsing

Neural networks were introduced for dependency parsing by Chen and Manning (2014) and through using neural networks, they were able to to make massive strides in transition-based dependency parsing. However, these transition-based dependency parsing mostly rely on greedy decoding at inference stage which means that if the model makes an error early on, then it parser will continue to diverge further and further away from the ground truth. Thus, we reconstruct a reinforcement learning algorithm to perform non-greedy decoding for transition-based parsers first shown by Shen et al [2016]. Our RL-based parser was tested on the English Penn Treebank (PTB) dataset and it achieved a 89.2% accuracy and improves by about 0.4 percentage points over a supervised dependency parser.