Cited text span identification for scientific summarisation using pre-trained encoders

We present our approach for the identification of cited text spans in scientific literature, using pre-trained encoders (BERT) in combination with different neural networks. We further experiment to assess the impact of using these cited text spans as input in BERT-based extractive summarisation methods. Inspired and motivated by the CL-SciSumm shared tasks, we explore different methods to adapt pre-trained models which are tuned for generic domain to scientific literature. For the identification of cited text spans, we assess the impact of different configurations in terms of learning from augmented data and using different features and network architectures (BERT, XLNET, CNN, and BiMPM) for training. We show that identifying and fine-tuning the language models on unlabelled or augmented domain specific data can improve the performance of cited text span identification models. For the scientific summarisation we implement an extractive summarisation model adapted from BERT. With respect to the input sentences taken from the cited paper, we explore two different scenarios: (1) consider all the sentences (full-text) of the referenced article as input and (2) consider only the text spans that have been identified to be cited by other publications. We observe that in certain experiments, by using only the cited text-spans we can achieve better performance, while minimising the input size needed.