Multi-Encoder Transformer for Korean Abstractive Text Summarization

In this paper, we propose a Korean abstractive text summarization approach that uses a multi -encoder transformer. Recently, in many natural language processing (NLP) tasks, the use of the pre-trained language models (PLMs) for transfer learning has achieved remarkable performance. In particular, transformer-based models such as Bidirectional Encoder Representations from Transformers (BERT) are used for pre-training and applied to downstream tasks, showing state-of-the-art performance including abstractive text summarization. However, existing text summarization models usually use one pre-trained model per model architecture, meaning that it becomes necessary to choose one PLM at a time. For PLMs applicable to Korean abstractive text summarization, there are publicly available BERT-based pre-trained Korean models that offer different advantages such as Multilingual BERT, KoBERT, HanBERT, and KorBERT. We assume that if these PLMs could be leveraged simultaneously, better performance would be obtained. We propose a model that uses multiple encoders which are capable of leveraging multiple pre-trained models to create an abstractive summary. We evaluate our method using three benchmark Korean abstractive summarization datasets, each named Law (AI-Hub), News (AI-Hub), and News (NIKL) datasets. Experimental results show that the proposed multi-encoder model variations outperform single -encoder models. We find the empirically best summarization model by determining the optimal input combination when leveraging multiple PLMs with the multi-encoder method.