A Semi-supervised Transfer Learning Framework for Low Resource Entity and Relation Extraction in Scientific Domain

With the development of scientific communities, the amount of papers increases quickly. It's important to convert the unstructured scientific papers into structured knowledge base, which relies on Information Extraction (IE) to extract entities and their relationships. Most existing IE methods require abundant annotated data, which is time-consuming and expensive to obtain, especially in scientific domain because it requires annotators with domain knowledge. Recently, several works have been proposed to solve the problem by semi-supervised learning. However, these methods require the input sentence to contain only two entities and simply classify the relationship between these two entities. Obviously, it is far from the realistic application scenarios that both entities and relations need to be extracted from raw text. In this paper, we propose a Semi-supervised Transfer Learning (STL) framework to tackle joint entity and relation extraction problem in a low resource situation. Specifically, STL adopts two main strategies: a rebalancing strategy for alleviating the bias to the majority class during semi-supervised learning, and a transfer learning strategy for transferring knowledge from domains with relatively rich annotation to domains that lack annotated data. Experiment results on two public scientific IE datasets show the effectiveness of the proposed method.