Suranah writeup for Bunescu 2007

This is a review of Bunescu_2007_learning_to_extract_relations_from_the_web_using_minimal_supervision by user:Suranah.

The paper presents a method for extracting relations using a handful of positive and negative examples. The problem is modeled as a multi-instance learning problem, which is then formulated as an optimization problem using a subsequence kernel. They discuss two types of biases, and propose some solutions to handle them.

Though the results can be said as encouraging, I am not comfortable with the basic premise of the paper. I don't see how this can be practically used to extract new relations. They have provided the results on some hand picked argument entities, but how will it be possible to pick a moderately good set while extracting information on the web. And if we have to search through an entire space of entities which have some higher degree of mutual information, then the result and the method of evaluation is almost meaningless. While I see that the technique is motivated in a novel way, but until I see some results on a much more open and random selection of entities (or from output of some other system), I am not convinced that their approach can have a meaningful impact.