Korean News Summarization with Contrasts by Augmenting Counterfactual Data

We introduce a method utilizing counterfactual data to increase performance for an abstractive summarization with contrasts. Previous summarization studies concentrate on summary generation given only input text; however, our method incorporates counterfactual data that includes hypothetical scenarios different from the original text. By augmenting the model with these counterfactual examples, it learns to identify key information and its contrasts more effectively, leading to more refined and accurate summaries. Experimental results demonstrate that our proposed method outperforms traditional methods and general LLM prompting methods, showcasing its potential in generating contrastive summaries.