Abstractive Summarization of Historical Documents: A New Dataset and Novel Method Using a Domain-Specific Pretrained Model

Automatic Text Summarization (ATS) systems aim to generate concise summaries of documents while preserving their essential aspects using either extractive or abstractive approaches. Transformer-based ATS methods have achieved success in various domains; however, there is a lack of research in the historical domain. In this paper, we introduce HistBERTSum-Abs, a novel method for abstractive historical single-document summarization. A major challenge in this task is the lack of annotated datasets for historical text summarization. To address this issue, we create a new dataset using archived documents obtained from the Centre Virtuel de la Connaissance sur l’Europe group at the University of Luxembourg. Furthermore, we leverage the potential of HistBERT, a domain-specific bidirectional language model trained on the balanced Corpus of Historical American English, (https://www.english-corpora.org/coha/) to capture the semantics of the input documents. Specifically, our method adopts an encoder-decoder architecture, combining the pre-trained HistBERT encoder with a randomly initialized Transformer decoder. To address the mismatch between the pre-trained encoder and the non-pre-trained decoder, we employ a novel fine-tuning schedule that uses different optimizers for each component. Experimental results on our constructed dataset demonstrate that our HistBERTSum-Abs method outperforms recent state-of-the-art deep learning-based methods and achieves results comparable to state-of-the-art LLMs in zero-shot settings in terms of ROUGE-1, ROUGE-2, and ROUGE-L F1 scores. To the best of our knowledge, this is the first work on abstractive historical text summarization.