A bottleneck for training deep learning-based architectures on text is the availability of large enough annotated training corpora. This is especially an issue in highly specialized domains such as those of biomedicine. Transfer Learning (TL) approaches address this problem by leveraging existing labeled data originally designed for related tasks or domains [1]. However, adaptation between dissimilar domains may lead to negative transfer, i.e. transfer that decreases the performance for the target domain. In this article, we apply a TL strategy using the TreeLSTM model for the task of biomedical Relation Extraction (RE). We propose an analysis of the syntactic features of source and target domain corpora to provide elements of interpretation for the improvements we obtained.
展开▼
