TY - CHAP

T1 - How difficult is the adaptation of POS taggers?

AU - Rodrigues, Irving Muller

AU - Fernandes, Eraldo Rezende

N1 - Conference code: 6

PY - 2017/6/28

Y1 - 2017/6/28

N2 - Domain adaptation is a difficult problem, but also a relevant one. Unsupervised domain adaptation focuses on adapting a model from a source domain, which includes plenty of labeled data, to a target domain that provides no labeled data. This is the most compelling setting of domain adaptation, but it is also the most difficult one. We perform an experimental analysis to highlight how difficult this problem is. We show that the best available unsupervised domain adaptation system for POS tagging can be outperformed by a simple POS tagger that has access to only 250 labeled sentences from the target domain. This is not a fair comparison between these two systems, of course; but it highlights that unsupervised domain adaptation is not well solved yet. Moreover, the best available systems are not yet practical, since they are complex, difficult to implement, and do not achieve significant improvements.

AB - Domain adaptation is a difficult problem, but also a relevant one. Unsupervised domain adaptation focuses on adapting a model from a source domain, which includes plenty of labeled data, to a target domain that provides no labeled data. This is the most compelling setting of domain adaptation, but it is also the most difficult one. We perform an experimental analysis to highlight how difficult this problem is. We show that the best available unsupervised domain adaptation system for POS tagging can be outperformed by a simple POS tagger that has access to only 250 labeled sentences from the target domain. This is not a fair comparison between these two systems, of course; but it highlights that unsupervised domain adaptation is not well solved yet. Moreover, the best available systems are not yet practical, since they are complex, difficult to implement, and do not achieve significant improvements.

KW - Informatics

KW - tagging

KW - training

KW - Feature extraction

KW - adaptation models

KW - natural language processing

KW - syntactics

KW - Business informatics

UR - http://www.scopus.com/inward/record.url?scp=85049536443&partnerID=8YFLogxK

U2 - 10.1109/BRACIS.2017.77

DO - 10.1109/BRACIS.2017.77

M3 - Article in conference proceedings

AN - SCOPUS:85049536443

SN - 978-1-5386-2408-1

SP - 360

EP - 365

BT - BRACIS 2017

PB - Institute of Electrical and Electronics Engineers Inc.

CY - Piscataway

T2 - Brazilian Conference on Intelligent Systems - BRACIS 2017

Y2 - 2 October 2017 through 5 October 2017

ER -