APPLICATIONS OF DOMAIN DIVERGENCES FOR DOMAIN ADAPTATION IN NLP

Abstract

The robustness of machine learning models is important for their deployment in various conditions, as they may not work well under different input distributions. Domain divergence is a mathematical tool used to quantify the difference between two input distributions, and its understanding and application can improve the usefulness of machine learning models. This thesis explores the different applications of divergence measures, particularly in Natural Language Processing (NLP). First, it provides a taxonomy of divergence measures. The thesis makes contributions in three areas: predicting model performance drop under new distributions, aligning source and target domain representations for novel applications, and understanding the robustness of models under different domains. Additionally, the thesis proposes a method for applying divergence measures in a parameter-efficient manner for domain adaptation in NLP. The work concludes with limitations and future research avenues.