Apappu writeup on Bunescu and Mooney 2007

This is a review of Bunescu_2007_learning_to_extract_relations_from_the_web_using_minimal_supervision by user:Apappu.

• This paper is about binary classification over a pair of arguments against a target relationship using relation kernels based on Multiple Instance Learning.

• One of the salient features of this work is using limited number of examples both positive and negative, supplemented by two types of bias (correlation and semantic based)

• An interesting insight from the paper is how MIL is treated as simple supervised learning because of the problem of fewer but larger bags of instances.

• Interesting to know that syntactic features that were part of the original subsequence kernel didn't seem to discount the performance as we could see SSK-T1 did better than the original kernel.