From the Information Extraction Pipeline to Global Models, and Back

Опубликовано 12 августа 2016, 3:07

Decisions in information extraction (IE), such as determining the types and relations of entities mentioned in text, depend on each other. To remain efficient, most systems make decisions in a sequential pipeline fashion, even if later decisions could help earlier ones. In this talk I will show how we used Conditional Random Fields to make these decisions jointly, substantially outperformed less global approaches and ranked first in several international IE competitions. I will then present relaxation methods we developed and applied to scale up (exact) inference in such models. In the final part of my talk I will argue why we should not dismiss the pipeline and present an exact beam-search algorithm, based on column generation, to overcome the pipeline's greedy nature.