Stochastic Dependency Parsing Based on A* Admissible Search

Dependency parsing has gained attention in natural language understanding because the representation of dependency tree is simple, compact and direct such that robust partial understanding and task portability can be achieved more easily. However, many dependency parsers make hard decisions with local information while selecting among the next parse states. As a consequence, though the obtained dependency trees are good in some sense, the N-best output is not guaranteed to be globally optimal in general. In this paper, a stochastic dependency parsing scheme based on A* admissible search is formally presented. By well representing the parse state and appropriately designing the cost and heuristic functions, dependency parsing can be modeled as an A* search problem, and solved with a generic algorithm of state space search. When evaluated on the Chinese Tree Bank, this parser can obtain 85.99% dependency accuracy at 68.39% sentence accuracy, and 14.62% node ratio for dynamic heuristic. This parser can output N-best dependency trees, and integrate the semantic processing into the search process easily.