Learning to select data for transfer learning with Bayesian Optimization

TL;DR

This paper introduces a method to learn data selection measures for transfer learning using Bayesian Optimization, improving performance across multiple NLP tasks by considering domain similarity and diversity.

Contribution

It proposes a novel approach to learn data selection measures with Bayesian Optimization, outperforming existing domain similarity metrics across various tasks.

Findings

Learned measures outperform existing similarity measures.

Incorporating diversity improves data selection.

Some transferability of learned measures across models and tasks.

Abstract

Domain similarity measures can be used to gauge adaptability and select suitable data for transfer learning, but existing approaches define ad hoc measures that are deemed suitable for respective tasks. Inspired by work on curriculum learning, we propose to \emph{learn} data selection measures using Bayesian Optimization and evaluate them across models, domains and tasks. Our learned measures outperform existing domain similarity measures significantly on three tasks: sentiment analysis, part-of-speech tagging, and parsing. We show the importance of complementing similarity with diversity, and that learned measures are -- to some degree -- transferable across models, domains, and even tasks.