Active Learning for Sense Annotation

Hector Martinez Alonso; Barbara Plank; Anders Trærup Johannsen; Anders Søgaard

Active Learning for Sense Annotation

Hector Martinez Alonso, Barbara Plank, Anders Trærup Johannsen, Anders Søgaard

Centre for Language Technology

Abstract

This article describes a real (nonsynthetic) active-learning experiment to obtain supersense annotations for Danish. We compare two instance selection strategies, namely lowest-prediction confidence (MAX), and sampling from the confidence distribution (SAMPLE). We evaluate their performance during the annotation process, across domains for the final resulting system, as well as against in-domain adjudicated data. The SAMPLE strategy yields competitive models that are more robust than the overly length-biased selection criterion of MAX.

Original language	English
Title of host publication	Proceedings of the 20th Nordic Conference of Computational Linguistics : NODALIDA 2015
Number of pages	5
Place of Publication	Linköping
Publisher	Linköping University Electronic Press
Publication date	2015
Pages	245-250
ISBN (Electronic)	978-91-7519-098-3
Publication status	Published - 2015

Series	NEALT Proceedings Series
Volume	23
ISSN	1736-6305

Cite this

Active Learning for Sense Annotation. / Martinez Alonso, Hector; Plank, Barbara; Johannsen, Anders Trærup et al.

Proceedings of the 20th Nordic Conference of Computational Linguistics: NODALIDA 2015. Linköping : Linköping University Electronic Press, 2015. p. 245-250 (NEALT Proceedings Series, Vol. 23).

Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review

@inproceedings{508eaa7739e344fc9e1d022181131378,

title = "Active Learning for Sense Annotation",

abstract = "This article describes a real (nonsynthetic) active-learning experiment to obtain supersense annotations for Danish. We compare two instance selection strategies, namely lowest-prediction confidence (MAX), and sampling from the confidence distribution (SAMPLE). We evaluate their performance during the annotation process, across domains for the final resulting system, as well as against in-domain adjudicated data. The SAMPLE strategy yields competitive models that are more robust than the overly length-biased selection criterion of MAX.",

author = "{Martinez Alonso}, Hector and Barbara Plank and Johannsen, {Anders Tr{\ae}rup} and Anders S{\o}gaard",

year = "2015",

language = "English",

series = "NEALT Proceedings Series",

pages = "245--250",

booktitle = "Proceedings of the 20th Nordic Conference of Computational Linguistics",

publisher = "Link{\"o}ping University Electronic Press",

}

TY - GEN

T1 - Active Learning for Sense Annotation

AU - Martinez Alonso, Hector

AU - Plank, Barbara

AU - Johannsen, Anders Trærup

AU - Søgaard, Anders

PY - 2015

Y1 - 2015

N2 - This article describes a real (nonsynthetic) active-learning experiment to obtain supersense annotations for Danish. We compare two instance selection strategies, namely lowest-prediction confidence (MAX), and sampling from the confidence distribution (SAMPLE). We evaluate their performance during the annotation process, across domains for the final resulting system, as well as against in-domain adjudicated data. The SAMPLE strategy yields competitive models that are more robust than the overly length-biased selection criterion of MAX.

AB - This article describes a real (nonsynthetic) active-learning experiment to obtain supersense annotations for Danish. We compare two instance selection strategies, namely lowest-prediction confidence (MAX), and sampling from the confidence distribution (SAMPLE). We evaluate their performance during the annotation process, across domains for the final resulting system, as well as against in-domain adjudicated data. The SAMPLE strategy yields competitive models that are more robust than the overly length-biased selection criterion of MAX.

M3 - Article in proceedings

T3 - NEALT Proceedings Series

SP - 245

EP - 250

BT - Proceedings of the 20th Nordic Conference of Computational Linguistics

PB - Linköping University Electronic Press

CY - Linköping

ER -

Active Learning for Sense Annotation

Abstract

Fingerprint

Cite this