Michael Collins is an Irish-American computer scientist at Columbia University whose work on statistical parsing, discriminative models, and structured prediction has been among the most influential in computational linguistics. His PhD thesis on statistical parsing introduced head-driven models that achieved state-of-the-art accuracy and established new standards for the field.
Early Life and Education
Born in Dublin, Ireland, in 1967, Collins studied computer science and mathematics at University College Dublin before earning his PhD from the University of Pennsylvania in 1999 under Mitchell Marcus. His dissertation on statistical parsing models was immediately recognised as a landmark contribution. He held positions at AT&T Labs and MIT before joining Columbia University.
Born in Dublin, Ireland
Completed PhD at the University of Pennsylvania
Published influential head-driven statistical parsing models
Developed the structured perceptron for NLP
Introduced parameter estimation methods using large-margin training
Received the ACL Fellowship
Key Contributions
Collins's head-driven statistical parsers (Models 1, 2, and 3) used lexicalised probabilistic context-free grammars where the probability of each rule depended on the head word of the phrase. By conditioning on head words, these models captured crucial selectional preferences and subcategorisation information that unlexicalised models missed, achieving dramatic improvements in parsing accuracy on the Penn Treebank.
He introduced the structured perceptron for NLP tasks, adapting the classical perceptron algorithm to work with structured outputs such as parse trees, tag sequences, and translation hypotheses. This provided a simple, effective alternative to maximum entropy and conditional random field models. His work on discriminative reranking — training a model to select the best parse from an n-best list produced by a generative parser — introduced powerful feature engineering techniques and demonstrated the value of discriminative training for structured prediction.
"The key to good parsing is capturing the right statistical dependencies — and head words are the most important single source of information for syntactic disambiguation." — Michael Collins
Legacy
Collins's parsing models set accuracy records that stood for years and influenced virtually all subsequent work on statistical parsing. The structured perceptron became a standard tool for NLP practitioners. His clear technical writing and tutorials on discriminative models trained a generation of NLP researchers. The methods he developed paved the way for modern neural approaches to structured prediction.