I agree with the comments made about the need for a more detailed description of the week learner. Furthermore, i agree with Joe's comment about the lack of analysis of the results. For example BWI was observed to be highly biased towards precision; the short explaination of the authors: "because the learned contextual patterns are highly accurate" was not convincing to me. For several tasks BWI achieved both high precision and recall; what are the characteristics of these tasks compared to other tasks where the recall was significantly lower?

I found the usage of the "lexical" setting for the wildcards appropritate to achieve better performance, but i found the comparisons with other methods not having this information unfair.

What i really liked in the paper was the precise and clear formulation of both the IE task as a classification problem and the definition of a wrapper.

Overall, i found the paper interesting, but i would not consider the approach to be novel, but rather a good usage of boosting.