Web-Scale Features for Full-Scale Parsing

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 693–702, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Web-Scale Features for Full-Scale Parsing Mohit Bansal and Dan Klein Computer Science Division University of California, Berkeley {mbansal, klein}@cs.berkeley.edu Abstract Counts from large corpora (like the web) can be powerful syntactic cues. Past work has used web counts to help resolve isolated am- biguities, such as binary noun-verb PP attach- ments and noun compound bracketings. In this work, we first present a method for gener- ating web count features that address the full range of syntactic attachments. These fea- tures encode both surface evidence of lexi- cal affinities as well as paraphrase-based cues to syntactic structure. We then integrate our features into full-scale dependency and con- stituent parsers. We show relative error re- ductions of 7.0% over the second-order depen- dency parser of McDonald and Pereira (2006), 9.2% over the constituent parser of Petrov et al. (2006), and 3.4% over a non-local con- stituent reranker. 1 Introduction Current state-of-the art syntactic parsers have achieved accuracies in the range of 90% F1 on the Penn Treebank, but a range of errors remain. From a dependency viewpoint, structural errors can be cast as incorrect attachments, even for constituent (phrase-structure) parsers. For example, in the Berkeley parser (Petrov et al., 2006), about 20% of the errors are prepositional phrase attachment er- rors as in Figure 1, where a preposition-headed (IN) phrase was assigned an incorrect parent in the im- plied dependency tree. Here, the Berkeley parser (solid blue edges) incorrectly attaches from debt to the noun phrase $ 30 billion whereas the correct at- tachment (dashed gold edges) is to the verb rais- ing. However, there are a range of error types, as shown in Figure 2. Here, (a) is a non-canonical PP VBG VP NP NP … raising $ 30 billion PP from debt … Figure 1: A PP attachment error in the parse output of the Berkeley parser (on Penn Treebank). Guess edges are in solid blue, gold edges are in dashed gold and edges common in guess and gold parses are in black. attachment ambiguity where by yesterday afternoon should attach to had already, (b) is an NP-internal ambiguity where half a should attach to dozen and not to newspapers, and (c) is an adverb attachment ambiguity, where just should modify fine and not the verb ’s. Resolving many of these errors requires informa- tion that is simply not present in the approximately 1M words on which the parser was trained. One way to access more information is to exploit sur- face counts from large corpora like the web (Volk, 2001; Lapata and Keller, 2004). For example, the phrase raising from is much more frequent on the Web than $ x billion from. While this ‘affinity’ is only a surface correlation, Volk (2001) showed that comparing such counts can often correctly resolve tricky PP attachments. This basic idea has led to a good deal of successful …

Content truncated to 500 words