Computing and Informatics

A known way to improve the accuracy of dependency parsers is to combine several different parsing algorithms, in such a way that the weaknesses of each of the models can be compensated by the strengths of others. For example, voting-based combination schemes are based on variants of the idea of analyzing each sentence with various parsers, and constructing a combined output where the head of each node is determined by ``majority vote'' among the different parsers. Typically, such approaches combine very different parsing models to take advantage of the variability in the parsing errors they make. In this paper, we show that consistent improvements in accuracy can be obtained in a much simpler way by combining a single parser with itself. In particular, we start with a greedy implementation of the Nivre pseudo-projective arc-eager algorithm, a well-known left-to-right transition-based parser, and we combine it with a ``mirrored'' version of the algorithm that analyzes sentences from right to left. To determine which of the two obtained outputs we trust for the head of each node, we use simple criteria based on the length and position of dependency arcs. Experiments on several datasets from the CoNLL-X shared task and the WSJ section of the English Penn Treebank show that the novel combination system obtains better performance than the baseline arc-eager parser in all cases. To test the generality of the approach, we also perform experiments with a different transition system (arc-standard) and a different search strategy (beam search), obtaining similar improvements in all these settings.