Abstractive English Document Summarization Using BART Model with Chunk Method
Pascal Wilman, Talia Atara, Derwin Suhartono · Procedia Computer Science · 2024
Particularly for the summarization task and the summary of relatively brief documents, abstractive document summarization has become a favored research topic in the field of natural language processing as it continues to advance. We recommend employing the enhanced abstractive summarization model, which integrates a pre-trained BART model from the CNN/Daily Mail dataset with chunk method processing. This model is capable of processing both brief and lengthy text input, including news articles and research articles, which are categorized as long documents. Our approach involves chunking the string into a list of sentences and encoding the text of a lengthy document written in English. Subsequently, we employ a BART model decoder to generate a summary by combining the results of each chunk text summarization, which then serves as the result summary. The performance of our chunking method is better when utilizing BART, which has been pre-trained on the CNN/Daily Mail dataset, compared to the one pre-trained on the XSum dataset.