Two-Stage Synthesis Networks for Transfer Learning in Machine Comprehension
This addresses the problem of domain adaptation in machine comprehension for researchers, offering a method that is incremental but shows strong gains in a specific setting.
The paper tackles transfer learning for machine comprehension across domains without labeled data, achieving an F1 of 44.3% on NewsQA using a model pretrained on SQuAD, which outperforms a baseline of 7.6% and approaches in-domain performance of 50.0%.
We develop a technique for transfer learning in machine comprehension (MC) using a novel two-stage synthesis network (SynNet). Given a high-performing MC model in one domain, our technique aims to answer questions about documents in another domain, where we use no labeled data of question-answer pairs. Using the proposed SynNet with a pretrained model from the SQuAD dataset on the challenging NewsQA dataset, we achieve an F1 measure of 44.3% with a single model and 46.6% with an ensemble, approaching performance of in-domain models (F1 measure of 50.0%) and outperforming the out-of-domain baseline of 7.6%, without use of provided annotations.