Abstract
In this paper, we present our new experimental system of merging dependency representations of two parallel sentences into one dependency tree. All the inner nodes in dependency tree represent source-target pairs of words, the extra words are in form of leaf nodes.
We use Universal Dependencies annotation style, in which the function words, whose usage often differs between languages, are annotated as leaves. The parallel treebank is parsed in minimally supervised way.
Unaligned words are there automatically pushed to leaves. We present a simple translation system trained on such merged trees and evaluate it in WMT 2016 English-to-Czech and Czech-to-English translation task.
Even though the model is so far very simple and no language model and word-reordering model were used, the Czech-to-English variant reached similar BLEU score as another established tree-based system.