AI-Powered Automatic Summarization of Short Documents: A Transformer-Based Approach

Short, information-dense texts such as news headlines, product blurbs, and mobile notifications require concise and factually accurate summaries, yet current abstractive summarization models often hallucinate when context is limited . Existing research largely focuses on long-form summarization or uses decoding-time heuristics, leaving a reproducible, training-based approach tailored to short-document factuality underexplored. We curate a stratified short-document corpus (inputs ≤128 tokens, summaries ≤32 tokens) and propose a transformer-based model that augments cross-entropy training with a coverage penalty and an NLI-derived factuality regularizer, evaluating systems with ROUGE, BERTScore, automated factuality proxies (NLI, QA-consistency ), and human judgments. The proposed model outperforms strong baselines (BART-base, T5-small) with ROUGE-1 = 39.1, ROUGE-2 = 16.9, ROUGE-L = 37.8 and a BERTScore of 0.88, and achieves higher automated factuality (NLI entailment 84%, QA-consistency 82%). Human evaluation corroborates these gains (mean Fluency 4.4, Relevance 4.2, Faithfulness 4.0) and shows smaller performance degradation under domain shift and input-length variation. These results indicate that integrating factuality-aware objectives into transformer fine-tuning yields more reliable short-document summaries and provides a reproducible path toward deployment in applications requiring concise, trustworthy text generation.