Parsing TCT with a Coarse-to-fine Approach

A key observation is that concept compound constituent labels are detrimental to parsing performance. We use a PCFG parsing algorithm that uses a multilevel coarse-to-fine scheme. Our approach requires a sequence of nested partitions or equivalence classes of the PCFG nonterminals, where the nonterminals of each PCFG are clusters of nonterminals of the finer PCFG. We use the results of parsing at a coarser level to prune the next finer level. The coarse-to-fine method use hierarchical projections for incremental pruning. We present experiments which show that parsing with hierarchical state-splitting is fast and accurate on Tsinghua Chinese Treebank. In addition, we propose a multiple-model method that adds concept compound labels to the output of the simple PCFG model and gains higher bracketing recall from the simple model. This scheme can be implemented by training two models on different labeling styles.