Combining strategies for tagging and parsing Arabic

Abstract

We describe a simple method for combining taggers which produces substantially better performance than any of the contributing tools. The method is very simple, but it leads to considerable improvements in performance: given three taggers for Arabic whose individual accuracies range from 0.956 to 0.967, the combined tagger scores 0.995–a sevenfold reduction in the error rate when compared to the best of the contributing tools. Given the effectiveness of this approach to combining taggers, we have investigated its applicability to parsing. For parsing, it seems better to take pairs of similar parsers and back off to a third if they disagree.