Wide-Coverage Deep Statistical Parsing Using Automatic Dependency Structure Annotation

A number of researchers (Lin 1995; Carroll, Briscoe, and Sanfilippo 1998; Carroll et al. 2002; Clark and Hockenmaier 2002; King et al. 2003; Preiss 2003; Kaplan et al. 2004;Miyao and Tsujii 2004) have convincingly argued for the use of dependency (rather than CFG-tree) representations for parser evaluation. Preiss (2003) and Kaplan et al. (2004) conducted a number of experiments comparing “deep” hand-crafted wide-coverage with “shallow” treebankand machine-learningbased parsers at the level of dependencies, using simple and automatic methods to convert tree output generated by the shallow parsers into dependencies. In this article, we revisit the experiments in Preiss (2003) and Kaplan et al. (2004), this time using the sophisticated automatic LFG f-structure annotation methodologies of Cahill et al. (2002b, 2004) and Burke (2006), with surprising results. We compare various PCFG and history-based parsers (based on Collins, 1999; Charniak, 2000; Bikel, 2002) to find a baseline parsing system that fits best into our automatic dependency structure annotation technique. This combined system of syntactic parser and dependency structure annotation is compared to two hand-crafted, deep constraint-based parsers (Carroll and Briscoe 2002; Riezler et al. 2002). We evaluate using dependency-based gold standards (DCU 105, PARC 700, CBS 500 and dependencies for WSJ Section 22) and use the Approximate Randomization Test (Noreen 1989) to test the statistical significance of the results. Our experiments show that machine-learning-based shallow grammars augmented with sophisticated automatic dependency annotation technology outperform hand-crafted, deep, widecoverage constraint grammars. Currently our best system achieves an f-score of 82.73% against the PARC 700 Dependency Bank (King et al. 2003), a statistically significant improvement of 2.18% over the most recent results of 80.55% for the hand-crafted LFG grammar and XLE parsing system of Riezler et al. (2002), and an f-score of 80.23% against the CBS 500 Dependency