research Hal Daumé III
about me
cv + bio
publications
research
teaching
software
photos
calendar
contact me
links
I am interested in many problems. Below is a snapshot (more information on some topics is on the NLP blog. I'm always interested in collaborating: if you want to work with me on any of these topics (or others), contact me by email.

Structured prediction

Learning to produce outputs with complex combinatorial structure is hard; I want efficient ways of doing it with minimal requirements on the part of the user and problem.
Papers: [Structured Learning for NLP: Thesis], [Search-based Structured Prediction], [Search-based SP as Classification, NIPSWS05], [Learning as Search Optimization, ICML05]

Automatic document summarization

I want to produce short documents from collections of long documents. I'm particularly interested in models with user interaction, such as queries, personalization, etc.
Papers: [Bayesian Query-Focused Summarization, ACL06], [Alignments for Summarization, CL05], [Bayesian Summarization, MSE05], [Fusion is Ill-defined, Sum04], [Noisy-channel Compression, ACL02]

Bayesian learning

The Bayesian paradigm provides an attractive set of techniques for learning when prior knowledge is strong or data is sparse. I am particularly interested in nonparametric Bayesian techniques and their applications to large scale learning problems.
Papers: [Bayesian Typology, ACL07], [Bayesian Query-Focused Summarization, ACL06], [Supervised Clustering with the Dirichlet Process, JMLR05], [Bayesian Summarization, MSE05], [Bayes Factors for Feature Selection, Unpub04]

Named entity extraction and coreference resolution

This is the task of finding mentions of people, places and things in a document. I'm particularly interested in the use of real-world knowledge for improving performance on this task, and statistical learning techniques that solve the problem jointly.
Papers: [Structured Learning for NLP: Thesis], [Features for a Large-scale EDT System, HLT05]

Domain adaptation/transfer learning

Often we have lots of data from some domain we don't care about but only a little from a domain we do care about. How can we get the best performance on the new domain with the littlest work?
Papers: [Easy adaptation, ACL07], [Domain adaptation, JAIR06]

Typology and Language evolution

Can we automatically discover the evolutionary path langauges went through by inspecting their taxonomic properties? What properties undergo the most rapid change and how important is language contact? How can we perform statistical inference and evaluate such models?
Papers: [Bayesian Typology, ACL07]

Discourse analysis

A document is more than just a sequence of sentences. These sentences are strongly interrelated. Discourse analysis is the task of uncovering these relations.

Semi-supervised learning

How can we learn from a lot of unannotated data and just a little annotated data? And how can we do it in the context of structured prediction, perhaps when the annotated data is only "weakly" annotated?

Textual entailment (and generalizations)

When does one sentence imply another? When do they contradict? What sort of logic is necessary to make these decisions?
quick links
   nlp blog
   searn
   nlp/ml meeting
   anlp (cs5964)
   ml (cs5350)
   mlrg (cs7941)
   algo (cs7936)
   whattosee
   thesis
   jmlr
   haskell tutorial
conferences
   soda 08
   recomb 08
   acl 08
   icml 08
   uai 08
   aaai 08
   ismb 08
   sigir 08
   coling 08
   kdd 08
   emnlp 08
   cikm 08
   nips 08
last updated on nine may, two thousand eight; contact me AT hal3 DOT name