Abstractive text summarization: A comprehensive survey of techniques, systems, and challenges

Abstractive text summarization addresses information overload by generating paraphrased content that mimics human expression, yet it faces significant computational and linguistic challenges. This paper presents a detailed functional taxonomy of abstractive summarization, structured along four dimensions: techniques (including structure-based, semantic, and deep learning approaches, including large language models), system architectures (ranging from single-model to multi-agent and human-in-the-loop interactive systems), evaluation methods (covering lexical, semantic, and human-centered assessments), and datasets. Our taxonomy explicitly distinguishes techniques from architectures to clarify how methodological strategies are operationalized in practice. We examine pressing multilingual challenges such as linguistic complexity, data scarcity, and performance disparities in cross-lingual transfer, particularly for low-resource languages. Additionally, we address persistent issues such as factual inaccuracies, content hallucinations, and biases in widely used evaluation metrics. The paper highlights emerging trends—including cross-lingual summarization, interactive summarization systems, and ethically grounded frameworks—as key directions for future research. This synthesis not only maps the current landscape but also outlines pathways to enhance the accuracy, reliability, and applicability of abstractive summarization in real-world settings.