Liuliu writeup of Poon 2007

This is a review of the paper poon_2007_joint_inference_in_information_extraction by user:Liuliu.

This paper proposed a joint citation matching method for citation extraction in the context of Markov logic network. The two key points are (1) rules which capture the logistics of citation segmenting and matching; and (2) efficient and good inference and weight learning method.

For (1), I feel the proposed rules are intuitive and easy to understand. This is due to the power of ML, so that we can define predicates as we want. They solved the isolated segmentation problem using HMM but by using rules, and proposed two joint methods (Jnt-seg and Jnt-seg-ER) by setting different rules. Results show that Jnt-seg works better on sparse data set while Jnt-Seg-ER works better on dense data set.

For (2), they use MC-SAT algorithm for inference and voted perceptron algorithm for learning weight. They didn't introduce slice-sampling method very detailed in this paper as it's out of its scope. However, as I am not very familiar with this sampling method, I don't quite understand how the inference step works.