Using a neural network-based feature extraction method to facilitate citation screening for systematic reviews

Citation screening is a labour-intensive part of the process of a systematic literature review that identi- fies citations eligible for inclusion in the review. In this paper, we present an automatic text classification approach that aims to prioritise eligible citations earlier than ineligible ones and thus reduces the man- ual labelling effort that is involved in the screening process. e.g. by automatically excluding lower ranked citations. To improve the performance of the text classifier, we develop a novel neural network-based feature extraction method. Unlike previous approaches to citation screening that employ unsupervised feature extraction methods to address a supervised classification task, our proposed method extracts doc- ument features in a supervised setting. In particular, our method generates a feature representation for documents, which is explicitly optimised to discriminate between eligible and ineligible citations. The generated document representation is subsequently used to train a text classifier. Experiments show that our feature extraction method obtains average workload savings of 56% when evaluated across 23 medical systematic reviews. The proposed method outperforms 10 baseline feature extraction methods by approximately 6% in terms of the WSS @95% metric.