MSQA: An Unsupervised Domain-Adapted Question Answering Method Based on Multiple Source Domains

In recent years, with the development of knowledge bases and crowdsourcing,researchers have successfully proposed several high-quality, large-scaledatasets. The rise of pre-trained language models has extensively promoted theprogress of machine reading comprehension tasks. Even so, the need for large-scaletraining data is one of the critical challenges in this field. Existingmethods: one is to improve the transferability of downstream tasks byoptimizing and improving pre-trained language models; the other is to usemultiple existing datasets to build a more extensive dataset or generatesynthetic data. However, the effect is not ideal and lacks specific logicalgeneralization abilities. In this paper, we propose a new domain adaptationframework based on BERT, called Multi-Source Domain Adaptive QA (MSQA), whichutilizes multiple existing large datasets and uses adversarial domaintechniques to narrow the distribution of multiple sources and target domains,thereby improving performance on target domain datasets. Extensive experimentson multiple widely used English datasets demonstrate that our method exhibitssuperior adaptive performance.