IITK at SemEval-2021 Task 10: Source-Free Unsupervised Domain Adaptation using Class Prototypes

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 438–444 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 438 IITK at SemEval-2021 Task 10: Source-Free Unsupervised Domain Adaptation using Class Prototypes Harshit Kumar∗ Jinang Shah∗ Nidhi Hegde∗ Priyanshu Gupta∗ Vaibhav Jindal ∗ Ashutosh Modi Indian Institute of Technology Kanpur (IIT Kanpur) {hakumar, sjinang, nidhih, guptap, vaibhavj}@iitk.ac.in ashutoshm@cse.iitk.ac.in Abstract Recent progress in deep learning has primar- ily been fueled by the availability of large amounts of annotated data that is obtained from highly expensive manual annotating pro- cesses. To tackle this issue of availability of annotated data, a lot of research has been done on unsupervised domain adaptation that tries to generate systems for an unlabelled target domain data, given labelled source do- main data. However, the availability of anno- tated or labelled source domain dataset can’t always be guaranteed because of data-privacy issues. This is especially the case with med- ical data, as it may contain sensitive informa- tion of the patients. Source-free domain adap- tation (SFDA) aims to resolve this issue by us- ing models trained on the source data instead of using the original annotated source data. In this work, we try to build SFDA systems for se- mantic processing by specifically focusing on the negation detection subtask of the SemEval 2021 Task 10. We propose two approaches - ProtoAUG and Adapt-ProtoAUG that use the idea of self-entropy to choose reliable and high confidence samples, which are then used for data augmentation and subsequent training of the models. Our methods report an improve- ment of up to 7% in F1 score over the baseline for the Negation Detection subtask. 1 Introduction The availability of large scale datasets has been the main driving factor behind the success of super- vised deep learning in the recent times. However, the process of data annotation is very expensive and time consuming, being one of the major chal- lenges in extending deep learning techniques for new tasks. One possible way to solve this problem is to train a machine learning model using an annotated ∗ Authors contributed equally to the work. Names in alphabetical order. source dataset to assist the annotation process over some unlabelled target dataset. However, there may be differences in the source and target domain dis- tributions, which may lead to inaccuracies. Thus the challenge is to update the weights of the source classifier to generalize it well on the target domain. This aligns with the well studied problem of Un- supervised Domain Adaptation (UDA) (Kouw and Loog, 2019; Wang and Deng, 2018; Ramponi and Plank, 2020) A common denominator across many popular UDA methods is their dependence on large amounts of labelled source domain data (Ganin and Lem- pitsky, 2015; Saito et al., 2018). However, many a times it is not possible to release the source domain dataset because of privacy concerns. This prob- lem becomes particularly relevant when working with clinical Natural Language Processing (NLP) datasets because they contain highly sensitive in- formation …

Content truncated to 500 words