Wka writeup of Poon and Domingos 2007
From Cohen Courses
Jump to navigationJump to searchThis is a review of poon_2007_joint_inference_in_information_extraction by user:wka.
Better inference by jointly infering segmentation and resolution for citation matching. MLN approach "considerably simpler, and outperforms others".
- "The joint inference methods do not impact citations with good boundaries"
- Weighted satisfiability solving, with different learning rate for each weight
- To avoid having joint inference hurting accuracy (entity resolution can spoil segmentation), similarity predicates are specifically defined to pass information across the 2 stages
- Rule that enforces 2 citations to resolve to same entity Jnt-Seg-ER works better on dense datasets, while plain works better on sparse datasets.
[Cluster recall: the fraction of clusters that are correctly output by the system after taking transitive closure from pairwise decisions] [Cora clusters are much more difficult to discriminate than in CiteSeer]
Paper has a very nice and smooth progression of the solution!