Awni Hannun • 11/8/2019

The Label Bias Problem

This technical article delves into the label bias problem, a subtle issue affecting discriminative sequence models such as Maximum Entropy Markov Models (MEMMs). It explains how label bias can cause models to ignore current observations during prediction, using part-of-speech tagging as an example. The piece outlines the mathematical formulation of MEMMs and sets the stage for discussing solutions like Conditional Random Fields (CRFs).

0 comments

#Sequence Labeling #Conditional Random Fields #Maximum Entropy Markov Model