A New Framework for Domain Adaptation without Model Retraining

We propose a principled and effective domain adaptation framework that pursues the goal of Open Domain NLP (train once, test anywhere). Most domain adaptation frameworks adapt the models trained on the source domain data by retraining it on target domains (with a mix of labeled and unlabeled data). However, it is time consuming to retrain big models or pipeline systems, and may not even be feasible if you consider a streaming data that may not be coherent (e.g., web data). We propose an adaptation framework that does not require retraining the original model. Instead, our approach adapts the target domain input so that it is more similar to the source domain, while preserving the labeling, thus increasing the accuracy of the original model when evaluated on target data. Our experiments on the named entity recognition task in scientific domains show an absolute F1 improvement of 13% over a state-of-the-art named entity recognizer. We also show that without any retraining, the proposed method outperforms the bootstrapping based adaptation method of (Jiang and Zhai, 2007b) that requires multiple rounds of retraining on the target domain data.