Conversion and Exploitation of Dependency Treebanks with Full-Tree LSTM

Abstract

As a method for exploiting multiple heterogeneous data, supervised treebank conversion can straightforwardly and effectively utilize linguistic knowledge contained in heterogeneous treebank. In order to efficiently and deeply encode the source-side tree, we for the first time investigate and propose to use Full-tree LSTM as a tree encoder for treebank conversion. Furthermore, the corpus weighting strategy and the concatenation with fine-tuning approach are introduced to weaken the noise contained in the converted treebank. Experimental results on two benchmark datasets with bi-tree aligned trees show that (1) the proposed Full-Tree LSTM approach is more effective than previous treebank conversion methods, (2) the corpus weighting strategy and the concatenation with fine-tuning approach are both useful for the exploitation of the noisy converted treebank, and (3) supervised treebank conversion methods can achieve higher final parsing accuracy than multi-task learning approach.