search

sustcsonglin / TN-PCFG

source code of NAACL2021 "PCFGs Can Do Better: Inducing Probabilistic Context-Free Grammars with Many Symbols“ and ACL2021 main conference "Neural Bilexicalized PCFG Induction"
45 stars 6 forks source link

A question about the MBR decoding #5

Closed speedcell4 closed 1 year ago

speedcell4 commented 1 year ago

To my understanding it means replacing MaxSemiring(log_potential) with MaxSemiring(marginal), is this correct? But, why does this work better?

sustcsonglin commented 1 year ago

Yep it's correct! MBR decoding uses marginals, which depends on global contexts, while Viterbi decoding uses local scores. As such i think it is very intuitive that MBR decoding's better than Viterbi decoding. This is a common finding in the parsing literature, e.g. in TreeCRF based dependency parsing https://aclanthology.org/2020.acl-main.302.pdf

speedcell4 commented 1 year ago

I got it, thank you~