Conditional Random Fields

Conditional random fields (CRFs) are probabilistic models that predict related labels jointly, using input-dependent evidence and factors such as neighboring-label or spatial interactions. They are used for sequence labeling tasks like named-entity recognition and part-of-speech tagging, as well as image segmentation and conversational emotion recognition. Common approaches include linear-chain and semi-Markov dynamic programming, mean-field inference and its neural unrollings, and energy-based models with learned potentials. A central challenge is making structured inference efficient at scale while ensuring its added dependencies improve on simpler neural predictions, as evaluations on settings such as MS COCO segmentation and IEMOCAP emotion recognition illustrate.