Domain Adaptation via Anaomaly Detection

1 Jan 2021  ·  Vivek Madan, Ashish Khetan, Zohar Karnin ·

Domain shift in finetuning from pre-training can significantly impact the performance of deep neural networks. In NLP, this led to domain specific models such as SciBERT, BioBERT, ClinicalBERT, and FinBERT; each pre-trained on a different, manually curated, domain-specific corpus. In this work, we present a novel domain-adaptation framework to tailor pre-training so as to reap the benefits of domain specific pre-training even if we do not have access to large domain specific pre-training corpus. The need for such a method is clear as it is infeasible to collect a large pre-training corpus for every possible domain. Our method is completely unsupervised and unlike related methods, works well in the setting where the target domain data is limited in size. We draw a connection between the task of adapting a large corpus to a target domain and that of anomaly detection, resulting in a scalable and efficient domain adaptation framework. We evaluate our framework and various baselines on eight tasks across four different domains: Biomedical, Computer Science, News, and Movie reviews. Our framework outperforms all the baseline methods and yields an average gain of $1.07\%$ in performance. We also evaluate it on one of the GLUE task, sentiment analysis and achieve an improvement of $0.4\%$ in accuracy.

PDF Abstract

Results from the Paper


  Submit results from this paper to get state-of-the-art GitHub badges and help the community compare results to other papers.

Methods


No methods listed for this paper. Add relevant methods here