Sgardine writesup Bellare 2009

This is a review of Bellare_2009_generalized_expectation_criteria_for_bootstrapping_extractors_using_record_text_alignment by user:Sgardine

Summary

A model is constructed that attempts to find an alignment of a text sequence against a token-tag sequence extracted from a database. A CRF is used to learn such an alignment. Its output is used for training an additional linear-chain CRF for use as an extractor. The alignment CRF outperforms previous methods on the alignment task. The extraction CRF is competitive with or better than previous methods, and in some cases not much worse than a CRF trained on (relatively expensive) human-annotated data.

Commentary

I was a little surprised that the AlignCRF was zero-order. I'd also have to go through again the discussion of training its parameters.