Sgardine writesup Poon 2007 Joint Inference IE

This is a review of poon_2007_joint_inference_in_information_extraction by user:Sgardine.

Summary

The authors propose to use Markov Logic to perform citation deduplication using joint inference. MLNs are reviewed briefly; inference will be performed by a satisfiability solver, namely MC-SAT; and weights will be learned with a voted perceptron. Various predicates for the task are introduced; the rules for entity resolution are modified slightly to prevent inference from over-using the contrapositive (which will occur often in the data since most citation pairs do not match). The model was tested on a CiteSeer dataset and Cora, and was found to outperform previous methods.

Commentary

I like this approach. It seems like there's a lot going on in designing the predicates and rules that is more of an artifact of how they're being used than what necessarily is the obvious way to encode the scenario ... it seems like those are coming from actually using the system, and their explanations left me uncertain that this was the only (or even best) way to approach the problem.