Task Refinement Learning for Improved Accuracy and Stability of Unsupervised Domain Adaptation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5895–5906 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5895 Task Refinement Learning for Improved Accuracy and Stability of Unsupervised Domain Adaptation Yftah Ziser and Roi Reichart Faculty of Industrial Engineering and Management, Technion, IIT syftah@campus.technion.ac.il, roiri@ie.technion.ac.il Abstract Pivot Based Language Modeling (PBLM) (Ziser and Reichart, 2018a), combining LSTMs with pivot-based methods, has yielded significant progress in unsupervised domain adaptation. However, this approach is still challenged by the large pivot detection prob- lem that should be solved, and by the inher- ent instability of LSTMs. In this paper we propose a Task Refinement Learning (TRL) ap- proach, in order to solve these problems. Our algorithms iteratively train the PBLM model, gradually increasing the information exposed about each pivot. TRL-PBLM achieves state- of-the-art accuracy in six domain adaptation setups for sentiment classification. Moreover, it is much more stable than plain PBLM across model configurations, making the model much better fitted for practical use.1 1 Introduction Domain adaptation (DA, (Daum´e III, 2007; Ben- David et al., 2010)) is a fundamental challenge in NLP, as many language processing algorithms re- quire costly labeled data that can be found in only a handful of domains. To solve this annotation bottleneck, DA aims to train algorithms with la- beled data from one or more source domains so that they can be effectively applied in a variety of target domains. Indeed, DA algorithms have been developed for many NLP tasks and domains (e.g. (Jiang and Zhai, 2007; McClosky et al., 2010; Titov, 2011; Bollegala et al., 2011; Rush et al., 2012; Schnabel and Sch¨utze, 2014)). A number of approaches for DA have been proposed (§ 2). With the raise of Neural Net- works (NNs), DA through Representation Learn- ing (DReL) where a shared feature space for the source and the target domains is learned, has 1Our code is publicly available at: https://github. com/yftah89/TRL-PBLM. become prominent. Earlier DReL approaches (Blitzer et al., 2006, 2007) were based on a lin- ear mapping of the original feature space to a new one, modeling the connections between pivot fea- tures – features that are frequent in the source and the target domains and are highly correlated with the task label in the source domain – and the complementary set of non-pivot features. This approach was later outperformed by autoencoder (AE) based methods (Glorot et al., 2011; Chen et al., 2012), which employ compress-based noise reduction to extract the shared feature space, but do not explicitly model the correspondence be- tween the source and the target domains. Recently, methods that marry the complementary strengths of NNs and pivot-based ideas (Ziser and Reichart (2017, 2018a), denoted here with ZR17 and ZR18, respectively) established a new state-of-the-art. Despite their strong empirical results, relying on NNs and on the distinction between pivot and non-pivot features, the models in ZR17 and ZR18 suffer from two limitations. These limitations stem from the fact that in order to create …

Content truncated to 500 words