WIND: Weighting Instances Differentially for Model-Agnostic Domain Adaptation

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2366–2376 August 1–6, 2021. ©2021 Association for Computational Linguistics 2366 WIND: Weighting Instances Differentially for Model-Agnostic Domain Adaptation Xiang Chen, Yue Cao, Xiaojun Wan Wangxuan Institute of Computer Technology, Peking University Center for Data Science, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {caspar,yuecao,wanxiaojun}@pku.edu.cn Abstract Domain Adaptation is a fundamental prob- lem in machine learning and natural language processing. In this paper, we study the do- main adaptation problem from the perspective of instance weighting. Conventional instance weighting approaches cannot learn the weights which make the model generalize well in target domain. To tackle this problem, inspired by meta-learning, we formulate the domain adap- tation problem as a bi-level optimization prob- lem, and propose a novel differentiable model- agnostic instance weighting algorithm. Our proposed approach can automatically learn the instance weights instead of using manually de- signed weighting metrics. To reduce the com- putational complexity, we adopt the second- order approximation technique during train- ing. Experimental results1 on three different NLP tasks (Sentiment Classification, Neural Machine Translation and Relation Extraction) illustrate the efficacy of our proposed method. 1 Introduction Domain shift is a challenging problem which is commonly encountered in Natural Language Pro- cessing (NLP). Due to the data distribution dis- crepancy between source and target domain, the model trained on the data from source domain may fail to achieve satisfying performance in target do- main. Therefore we face the domain adaptation problem. In some real-world situations, we may only focus on the performance of our model on a specific domain. To maintain the performance, we need labeled training data for supervised learning. However, we often cannot collect enough labeled training data relevant to the domain we are inter- ested in (in-domain). Thus, we need to introduce more labeled data from other different domains 1The code is available at https://github.com/ CasparSwift/WIND (out-of-domain). We aim to leverage the general knowledge from out-of-domain dataset to enhance the in-domain performance of our model. We consider a specific domain adaptation sce- nario in this work, where we have a few labeled in-domain training data and meanwhile we have sufficient labeled out-of-domain training data from other general domains. Training on these two datasets jointly is a straightforward solution for this scenario, but not all samples from out-of-domain dataset has equal effect during the training procedure. Several stud- ies (Koehn and Knowles, 2017) on neural machine translation (NMT) task show that, out-of-domain instances relevant to the in-domain data are benefi- cial while the instances irrelevant to the in-domain data may be even harmful to the translation quality. Apart from that, for sentiment classification task, some general expressions such as “I’m truly im- pressed by the design.” may appear in all domains. Taking them as training samples can help the model to learn general syntactic and semantic knowledge, which improves the cross-domain sentiment clas- sification performance. But using examples like “This chair is solid.” (negative sentiment, furniture domain) may reduce the accuracy of classifying …

Content truncated to 500 words