Coarse-to-fine n-best parsing and MaxEnt discriminative reranking

Coarse-to-fine n-best parsing and MaxEnt discriminative reranking

June 2005 | Eugene Charniak and Mark Johnson
This paper introduces a novel method for constructing high-quality $n$-best parses using a coarse-to-fine generative parser, followed by a MaxEnt discriminative reranking process. The coarse-to-fine approach generates 50-best parses with an $f$-score of 96.8 on the Penn Treebank, significantly improving over the 89.7 $f$-score of the base parser. The reranking parser uses a MaxEnt model to select the best parse from the 50-best list, achieving an $f$-score of 91.0 on sentences of length 100 or less. The paper details the implementation of the coarse-to-fine parsing algorithm, the feature extraction process, and the MaxEnt parameter estimation method. Experimental results show that the proposed system outperforms existing parsers, achieving a 13% reduction in error over the best parsers available at the time.This paper introduces a novel method for constructing high-quality $n$-best parses using a coarse-to-fine generative parser, followed by a MaxEnt discriminative reranking process. The coarse-to-fine approach generates 50-best parses with an $f$-score of 96.8 on the Penn Treebank, significantly improving over the 89.7 $f$-score of the base parser. The reranking parser uses a MaxEnt model to select the best parse from the 50-best list, achieving an $f$-score of 91.0 on sentences of length 100 or less. The paper details the implementation of the coarse-to-fine parsing algorithm, the feature extraction process, and the MaxEnt parameter estimation method. Experimental results show that the proposed system outperforms existing parsers, achieving a 13% reduction in error over the best parsers available at the time.
Reach us at info@study.space
[slides and audio] Coarse-to-Fine n-Best Parsing and MaxEnt Discriminative Reranking