Abstract:
Extreme Abstractive Summarization of long scientific papers requires domain knowledge and a concise summary maintaining faithfulness to the source and covering novel aspects presented in the paper. Human annotations are indeed expensive for the task, so we propose ExGrapf2, a novel encoder architecture that uses fractality, FFT, and Graph Convolution Network as its strong foundation to address the challenge. We observed that when the model is presented with different views of the source, it extracts more information from the same amount of data. ExGrapf2 successfully accomplishes the objective and beats the state-of-the-art models on SciTLDR dataset without any data augmentation. We also used the contrastive loss to enhance the performance further. The novelty is not only for the modules but also for how we fuse them.
