Structured prediction models via the Matrix-Tree Theorem

Terry Koo; Amir Globerson; Xavier Carreras; Michael Collins

Structured prediction models via the Matrix-Tree Theorem

Terry Koo^*, Amir Globerson, Xavier Carreras, Michael Collins

^*Corresponding author for this work

Massachusetts Institute of Technology

Research output: Contribution to conference › Paper › peer-review

107 Scopus citations

Abstract

This paper provides an algorithmic framework for learning statistical models involving directed spanning trees, or equivalently non-projective dependency structures. We show how partition functions and marginals for directed spanning trees can be computed by an adaptation of Kirchhoff's Matrix-Tree Theorem. To demonstrate an application of the method, we perform experiments which use the algorithm in training both log-linear and max-margin dependency parsers. The new training methods give improvements in accuracy over perceptron-trained models.

Original language	English
Pages	141-150
Number of pages	10
State	Published - 2007
Externally published	Yes
Event	2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL 2007 - Prague, Czech Republic Duration: 28 Jun 2007 → 28 Jun 2007

Conference

Conference	2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL 2007
Country/Territory	Czech Republic
City	Prague
Period	28/06/07 → 28/06/07

Cite this

@conference{262a0a76b8f34ac9825934912259d642,

title = "Structured prediction models via the Matrix-Tree Theorem",

abstract = "This paper provides an algorithmic framework for learning statistical models involving directed spanning trees, or equivalently non-projective dependency structures. We show how partition functions and marginals for directed spanning trees can be computed by an adaptation of Kirchhoff's Matrix-Tree Theorem. To demonstrate an application of the method, we perform experiments which use the algorithm in training both log-linear and max-margin dependency parsers. The new training methods give improvements in accuracy over perceptron-trained models.",

author = "Terry Koo and Amir Globerson and Xavier Carreras and Michael Collins",

year = "2007",

language = "אנגלית",

pages = "141--150",

note = "2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL 2007 ; Conference date: 28-06-2007 Through 28-06-2007",

}

Structured prediction models via the Matrix-Tree Theorem. / Koo, Terry; Globerson, Amir; Carreras, Xavier et al.
2007. 141-150 Paper presented at 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL 2007, Prague, Czech Republic.

Research output: Contribution to conference › Paper › peer-review

TY - CONF

T1 - Structured prediction models via the Matrix-Tree Theorem

AU - Koo, Terry

AU - Globerson, Amir

AU - Carreras, Xavier

AU - Collins, Michael

PY - 2007

Y1 - 2007

N2 - This paper provides an algorithmic framework for learning statistical models involving directed spanning trees, or equivalently non-projective dependency structures. We show how partition functions and marginals for directed spanning trees can be computed by an adaptation of Kirchhoff's Matrix-Tree Theorem. To demonstrate an application of the method, we perform experiments which use the algorithm in training both log-linear and max-margin dependency parsers. The new training methods give improvements in accuracy over perceptron-trained models.

AB - This paper provides an algorithmic framework for learning statistical models involving directed spanning trees, or equivalently non-projective dependency structures. We show how partition functions and marginals for directed spanning trees can be computed by an adaptation of Kirchhoff's Matrix-Tree Theorem. To demonstrate an application of the method, we perform experiments which use the algorithm in training both log-linear and max-margin dependency parsers. The new training methods give improvements in accuracy over perceptron-trained models.

UR - http://www.scopus.com/inward/record.url?scp=80053344876&partnerID=8YFLogxK

M3 - ???researchoutput.researchoutputtypes.contributiontoconference.paper???

AN - SCOPUS:80053344876

SP - 141

EP - 150

T2 - 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL 2007

Y2 - 28 June 2007 through 28 June 2007

ER -

Structured prediction models via the Matrix-Tree Theorem

Abstract

Conference

Other files and links

Fingerprint

Cite this