Parsing. The monolingual parsing features we use are simply parsing model scores under the parser of Petrov and ▇▇▇▇▇ (2007). While that parser uses heavily refined PCFGs with rule probabilities defined at the refined symbol level, we interact with its posterior distribution via posterior marginal probabilities over unrefined symbols. In particular, to each unrefined anchored production iAj iBkCj, we associate a single feature whose value is the marginal quantity log P(iBkCj iAj, s) under the monolingual parser. These scores are the same as the variational rule scores of Matsuzaki et al. (2005).6
Appears in 1 contract
Sources: Dissertation
Parsing. The monolingual parsing features we use are simply parsing model scores under the parser of Petrov and ▇▇▇▇▇ (2007). While that parser uses heavily refined PCFGs with rule probabilities defined at the refined symbol level, we interact with its posterior distribution via posterior marginal probabilities over unrefined symbols. In particular, to each unrefined anchored production iAj iBkCj, we associate a single feature whose value is the marginal quantity log P(iBkCj iAj, s) under the monolingual parser. These scores are the same as the variational rule scores of Matsuzaki ▇▇▇▇▇▇▇▇▇ et al. (2005).6
Appears in 1 contract
Sources: Dissertation