Apappu writeup on Sutton and Maccullum

This is a review of sutton_2004_collective_segmentation_and_labeling_of_distant_entities_in_information_extraction by user:Apappu.

• This paper aims to solve the co-reference problem that appear in talk-summary emails.
• They describe skip-chain CRFs as extended linear chain CRFs that can allow long-distant edges between tokens which represent a sense of co-reference in case of Named Entities.
• They also mention that inference in skip-chain CRFs is intractable hence they reside with approximate inference using tree reparametrization algorithm.
• They show that in overall performance skip-CRF does better than the linear one due to improved recall over the recognition of speaker names when former was used.

• NO reason was given to explain the drop in performance over recognizing s-time and e-time. Does, approximate inference hurt short-distant connections ?