Apappu writeup of Roth talk

• This work explores the traditional SRL problem and attempts to decompose the problem in a novel way to maximize certain objectives using ILP approach.
• Learning conditional models using declarative (FOP style) expressive constraints (domain specific/knowledge driven)
• Detect, classify and inference (where ILP has been incorporated) using probability distributions of classification and additional linguistic (domain specific) constraints.
• These are the kind of constraints, we have seen in our previous readings in different papers, in several forms specially embedded in the model itself.
• The idea of re-ranking (instead of dropping) the illegal assignments seems to do well in IE tasks (eg. citation entity extraction)

• Overall, it is a very good take on how to include external knowledge while training a model.