70 likes | 197 Vues
Traditional Named Entity Recognition (NER) assumes that each entity type is an independent class. This approach overlooks the hierarchical structure of entities. We explore the Genia Corpus Models, all built on the Maximum Entropy Markov Model (MEMM) classifier. Our models implement different strategies: using ancestor types as features, classifying at each hierarchical level, and assigning both local and global weights to nodes in the entity tree. Acknowledgments go to David, Mihai, and our instructors and TAs for their invaluable contributions.
E N D
Motivation • Traditional NER assume that each entity type is an independent class. • However, they can have a hierarchical structure
Models • All models are based on MEMM classifier • Model 1 • Just take the ancestor types to be features • Model 2 • Train a classifier at each level • Do verterbi on paths in the tree
Model 3 • Every node in the tree has a local weight and global weight • Global weight is for classification. • The sum of the local weights from the root to the node • Example
Acknowledgement • Thanks David and Mihai for insightful discussions • Thanks instructors for excellent courses • Thanks TAs for hard work.