News is More than a Collection of Facts: Moral Frame Preserving News Summarization - code

Code used to generate summaries that preserve the moral framing of the original news article. We leverage the zero-shot summarization ability of Large Language Models, shown to produce results on par with human-generated summaries. We compare three language models and five prompting methods. Leveraging the intuition that journalists intentionally use or report moral-laden words in the article text, we propose approaches that first identify moral-laden words in the article (e.g., through Chain-of-Thought or supervised classification) and then guide the language model in preserving such words in the summary.