Abstract
This paper describes a statistical method for Japanese dependency analysis. The method differs from conventional statistical models in the way of calculating statistical values. The conventional models calculate the probability of a correct dependency between two bunsetsus (phrasal units of Japanese) for each pair of bunsetsus. On the other hand, we propose the triplet/quadruplet model, in which the conditional part of the probability consists of information on a modifier bunsetsu and all its modification candidates, and the probability that a candidate is chosen as the modifiee is calculated. The number of candidates is restricted to three or less by an HPSG-based grammar and heuristics. With a maximum entropy estimation, our parser achieves high accuracy: 88.6% for the analysis of the EDR annotated corpus.
