Thomas L. Griffiths
Thomas L. Griffiths works on Bayesian methods for learning linguistic structure from unannotated text. The papers here ask how a learner can recover categories and grammars when no labelled examples are available, and they answer by replacing single best-fit parameter estimates with full Bayesian inference. Typical tools include Dirichlet priors, Markov chain Monte Carlo sampling and dynamic programming over parse charts, applied to problems such as part-of-speech tagging and grammar induction.
The most cited of these works is A fully Bayesian approach to unsupervised part-of-speech tagging (2007). It keeps a standard trigram hidden Markov model but integrates out its transition and emission parameters, so that tags are chosen because they stay probable across many parameter settings. It also argues for sparse priors suited to the skewed distributions of natural language, and for variation of information as a label-free way to score clusterings. Close behind is Bayesian Inference for PCFGs via Markov Chain Monte Carlo (2007). That paper offers alternatives to Inside-Outside estimation for probabilistic context-free grammars, through a Gibbs sampler and a collapsed Hastings sampler that both draw whole trees from an inside chart.
These papers, both from 2007, share a common direction. Each treats sampling-based Bayesian inference as a practical substitute for maximum-likelihood training in unsupervised language learning. Each also attends to convergence techniques such as annealing. The work on tagging and the work on grammars apply the same idea to different levels of linguistic structure.
Griffiths is currently affiliated with Princeton University and the University of California, Berkeley. Earlier bylines include Brown University, Stanford University and the Massachusetts Institute of Technology. Co-authors on the papers here are Sharon Goldwater and Mark Johnson.
Thomas L. Griffiths has 2 papers in our library, in 2007, most of them on Part-of-speech tagging; often with Sharon Goldwater, Mark Johnson.
current affiliationPrinceton University, Princeton, United States; University of California, Berkeley, Berkeley, United States
OpenAlex · ORCID · h-index 97 · 697 works · 47 742 citations OpenAlex, 2026-10-14
2 works in our library · 266 citations in our library · 5 more works in our library · ⭐ Paper Star rank #2409
Affiliations
- Princeton University · Princeton, United States — current · 2018–2026 on 355 papers
- University of California, Berkeley · Berkeley, United States — current · 2001–2026 on 228 papers
- Georgia Institute of Technology · Atlanta, United States — 2024–2025 on 2 papers
- Stanford University · Stanford, United States — 2000–2019 on 14 papers
- University College London · London, United Kingdom — 2009–2017 on 2 papers
- University of California System · Oakland, United States — 2010–2016 on 4 papers
- The University of Western Australia · Perth, Australia — 2000–2013 on 3 papers
- Brown University · Providence, United States — 2005–2012 on 18 papers
- University of Louisiana at Lafayette · Lafayette, United States — 2007–2008 on 2 papers
- Massachusetts Institute of Technology · Cambridge, United States — 2002–2007 on 9 papers
Topics
Works by year
2000s
Papers
- A fully Bayesian approach to unsupervised part-of-speech tagging (2007 · 380 citations)
The paper addresses unsupervised part-of-speech tagging. Standard methods fit one "best" parameter setting by maximum-likelihood estimation (MLE) and then pick tags under that setting. Goldwater and Griffiths instead use a fully Bayesian approach. They keep the familiar trigram HMM structure but place symmetric… - Bayesian Inference for PCFGs via Markov Chain Monte Carlo (2007 · 352 citations)
The paper offers Bayesian alternatives to the usual maximum-likelihood way of learning probabilistic context-free grammars (PCFGs) from unannotated strings with the Inside-Outside (IO) algorithm.
Also in our library
Co-authors here
Cited by
This page in Markdown → https://md.research.elamur.ai/free/en/authors/thomas-l-griffiths/