Summarization of COVID-19 news documents deep learning-based using transformer architecture

Abstract

Facing the news on the internet about the spreading of Corona virus disease 2019 (COVID-19) is challenging because it is required a long time to get valuable information from the news. Deep learning has a significant impact on NLP research. However, the deep learning models used in several studies, especially in document summary, still have a deficiency. For example, the maximum output of long text provides incorrectly. The other results are redundant, or the characters repeatedly appeared so that the resulting sentences were less organized, and the recall value obtained was low. This study aims to summarize using a deep learning model implemented to COVID-19 news documents. We proposed transformer as base language models with architectural modification as the basis for designing the model to improve results significantly in document summarization. We make a transformer-based architecture model with encoder and decoder that can be done several times repeatedly and make a comparison of layer modifications based on scoring. From the resulting experiment used, ROUGE-1 and ROUGE-2 show the good performance for the proposed model with scores 0.58 and 0.42, respectively, with a training time of 11438 seconds. The model proposed was evidently effective in improving result performance in abstractive document summarization.