Domain adaptation is particularly important in natural language processing tasks, where manually creating labeled data for a specific domain is costly. Thus leveraging labeled data of available domains has become essential. In this study, we focus on the transferability of knowledge in sentence-level classification, and specifically, on classifying the functional categories of sentences in a scientific abstract, such as the background, objective, method, result, and conclusion. Considering the significant differences in both the semantics and the syntactics of sentences between the source (e.g., biological science) and target domains (e.g., computational linguistics), we present a framework to automatically construct mappings in semantics and syntactics across domains, and embed these transformation blocks into a context-aware convolution neural network-based classifier.
Our results show that our best performing model improves the best baseline by 40%, which demonstrates the effectiveness of the proposed framework.
Keywords: Domain Adaptation, Sentence Classification, Structured Abstract, Context-aware, Convolutional Neural Network