Introduction

摘要

The growth in computational power and the rise of Deep Neural Networks (DNNs) have revolutionized the field of Natural Language Processing (NLP). The ability to collect massive datasets with the capacity to train big models on powerful CPUs, has yielded NLP-based technology that was beyond imagination only a few years ago. Unfortunately, this technology is still limited to a handful of resource rich languages and domains. This is because most NLP algorithms rely on the fundamental assumption that the training and the test sets are drawn from the same underlying distribution. When the train and test distributions do not match, a phenomenon known as domain shift, such models are likely to encounter performance drops. Despite the growing availability of heterogeneous data, many NLP domains still lack the amounts of labeled data required to feed data-hungry neural models, and in some domains and languages even unlabeled data is scarce. As a result, the problem of domain adaptation, training an algorithm on annotated data from one or more source domains, and applying it to other target domains, is a fundamental challenge that has to be solved in order to make NLP technology available for most world languages and textual domains.