Category-Aware Indonesian Abstractive Text Summarization with Augmentation

The explosive growth of global digital content, expected to reach 181 zettabytes by 2025, demands efficient strategies for processing vast amounts of text. Automatic Text Summarization helps streamline information consumption by producing concise versions of lengthy content. However, existing models for the Indonesian language often overlook topical or categorical context, leading to less informative summaries. To address this, this study proposes a category-aware summarization framework that integrates IndoBERT-based classification with class based TF IDF to identify category representative keywords. These keywords are embedded into the summarization input using special tokens to enhance contextual understanding. Additionally, backtranslation is used as a data augmentation technique to increase lexical diversity and improve generalization. Experiments conducted using the IndoSum dataset and IndoBART summarization model show that the proposed method yields higher ROUGE scores, particularly in underrepresented categories. However, when combined with backtranslation, the model exhibits a trade-off between specificity and generalization, slightly lowering overall performance. The results highlight the effectiveness of incorporating categorical context for improving summarization quality in low-resource languages like Indonesian.