Reverse Revision and Linear Tree Combination for Dependency Parsing

Deterministic transition-based Shift/Reduce dependency parsers make often mistakes in the analysis of long span dependencies (McDonald & Nivre, 2007). Titov and Henderson (2007) address this accuracy drop by using a beam search instead of a greedy algorithm for predicting the next parser transition. We propose a parsing method that allows reducing several of these errors, although maintaining a quasi linear complexity. The method consists in two steps: first the sentence is parsed by a deterministic Shift/Reduce parser, then a second deterministic Shift/Reduce parser analyzes the sentence in reverse using additional features extracted from the parse trees produced by the first parser. Right-to-left parsing has been used as part of ensemble-based parsers (Sagae & Lavie, 2006; Hall et al., 2007). Nivre and McDonald (2008) instead use hints from one parse as features in a second parse, exploiting the complementary properties of graph-based parsers (Eisner, 1996; McDonald et al., 2005) and transition-based dependency parsers (Yamada & Matsumoto, 2003; Nivre & Scholz, 2004). Also our method uses input from a previous parser but only uses parsers of a single type, deterministic transition-based Shift/Reduce, maintaining an overall linear complexity. In fact both the ensemble parsers and the stacking solution of NivreMcDonald involve the computation of the maximum spanning tree (MST) of a graph, which require algorithms of quadratic time complexity (e.g. (Chu & Liu, 1965; Edmonds, 1967)). We introduce an alternative linear combination method. The algorithm is greedy and works by combining the trees top down. We tested it on the dependency trees produced by three parsers, a Leftto-Right (LR ), a Right-to-Left (RL ) and a stacked Right-to-Left parser, or Reverse Revision parser (Rev2 ). 1 The experiments show that in practice its output often outperforms the results produced by calculating the MST.