Apappu writeup on Poon and Domingos
From Cohen Courses
Jump to navigationJump to searchThis is a review of the paper poon_2007_joint_inference_in_information_extraction by user:Apappu.
- My first impression of this paper is that it is trying to imitate PCFGs through predicates and rules.
- Solving segmentation and entity-extraction problem (in citation realm) jointly with the help of Markov Logic Network (set of weighted first oder clauses).
- Authors show that individualy systems comparatively easy to assemble show it does indeed improve extraction accuracy
- In this work, MC-SAT is used for inference, discriminative weight learning using voted perceptron for learning.
- Authors have tried standalone segmentation and entity extraction treating them as baseline.
- Segmentation model is a HMM with enhanced ability to detect field boundaries, whereas Entity resolution model is an MLN described in previous work (Singla and Domingos '06)
- Predicates and rules are employed to transfer information from one stage to another.
- To perform joint segmentation (as opposed to isloated one) a predicate has been defined over a trigram with additional constraints over field(like title, author etc.) boundary.
- Essentially, this is a combination of appropriate linguistic cues and statistical model which performs better than isolated inference methods.
- I like this paper !
