Effect of Data Imbalance on Unsupervised Domain Adaptation of Part-of-Speech Tagging and Pivot Selection Strategies
Proceedings of the First International Workshop on Learning with Imbalanced Domains: Theory and Applications, PMLR 74:103-115, 2017.
Domain adaptation is the task of transforming a model trained using data from a source domain to a different target domain. In Unsupervised Domain Adaptation (UDA), we do not assume any labelled training data from the target domain. In this paper, we consider the problem of UDA in the contact of Part-of-Speech (POS). Specifically, we study the effect of data imbalance on UDA of POS, and compare different pivot selection strategies for accurately adapting a POS tagger trained using some source domain data to a target domain. We propose the use of F-score to select pivots using available labelled data in the source domain. Our experimental results on using benchmark dataset for cross-domain POS tagging, show that using frequency combined with F-scores for selecting pivots in the source labelled data produces the best results.