Improved Parsing for Arabic by Combining Diverse Dependency Parsers

Recently there has been a considerable interest in dependency parsing for many reasons. First, it works accurately for a wide range of typologically different languages. Second, it can be useful for semantics, since it can be easier to attach compositional rules directly to lexical items than to assign them to large numbers of phrase structure rules. Third, robust machine-learning based parsers are available. In this paper, we investigate two techniques for combining multiple data-driven dependency parsers for parsing Arabic, where we are faced with an exceptional level of lexical and structural ambiguity. Experimental results show that combined parsers can produce more accurate results, even for imperfectly tagged text, than each parser produces by itself for texts with the gold-standard tags.