Hebrew dependency parsing: Initial results

We describe a newly available Hebrew Dependency Treebank, which is extracted from the Hebrew (constituency) Treebank. We establish some baseline unlabeled dependency parsing performance on Hebrew, based on two state-of-the-art parsers, MST-parser and MaltParser. The evaluation is performed both in an artificial setting, in which the data is assumed to be properly morphologically segmented and POS-tagged, and in a real-world setting, in which the parsing is performed on automatically segmented and POS-tagged text. We present an evaluation measure that takes into account the possibility of incompatible token segmentation between the gold standard and the parsed data. Results indicate that (a) MST-parser performs better on Hebrew data than MaltParser, and (b) both parsers do not make good use of morphological information when parsing Hebrew.