summarization

To address performance bottlenecks in automatic text summarization (ATS) arising from cross-domain, multi-platform textual proliferation—exacerbated by linguistic stylistic diversity and technical complexity—this paper presents a systematic survey and methodological innovations. We propose the first taxonomy of ATS methods explicitly designed for linguistic style variability; develop a hybrid evaluation matrix integrating linguistic features with deep learning to unify assessment criteria across extractive, abstractive, and hybrid paradigms; and empirically delineate the performance boundaries and applicability scopes of mainstream models via BLEU/ROUGE metrics and controlled comparative experiments. The work delivers a reproducible evaluation benchmark, a principled framework for model selection, and a roadmap for architectural evolution—thereby substantially enhancing ATS generalizability and interpretability over heterogeneous texts.

To address information overload in online news, this paper proposes a deep learning–based extractive text summarization method. We employ BERT to obtain sentence-level semantic embeddings and formulate summarization as a binary classification task, systematically comparing logistic regression, feedforward neural networks, and LSTM models. The LSTM model explicitly captures inter-sentence sequential dependencies, yielding significant improvements over baselines—including Lede-3—on the Cornell Newsroom dataset (1.3 million samples). Experimental results demonstrate that the proposed LSTM model achieves state-of-the-art performance in both F1 score and ROUGE-1, delivering substantial gains in summary quality. Its high accuracy, combined with computational efficiency and architectural simplicity, ensures practical deployability. This work provides a scalable, technically robust solution for enhancing news content accessibility and improving user engagement.

Large language models (LLMs) exhibit unstable summarization quality and limited controllability over abstraction levels. Method: This paper proposes a controllable abstractive summarization framework based on multi-stage prompt engineering, integrating semantic analysis, topic modeling, and noise-aware control to enable adjustable abstraction granularity. We systematically investigate the impact of prompt length, data noise, and text genre on summarization performance using the CNN/Daily Mail benchmark. Contribution/Results: Experiments demonstrate that medium-length prompts yield statistically significant improvements in ROUGE-L scores; increased input noise degrades performance consistently; and LLMs generalize best on news-domain texts. The framework provides an interpretable, configurable pathway to enhance accuracy, consistency, and abstraction-level control in LLM-generated summaries.

To address the challenges of lengthy, low-information-density Bangla news texts and the scarcity of high-quality abstractive summarization resources for low-resource languages, this paper introduces the first end-to-end neural abstractive summarization model tailored for Bangla. Methodologically, it adopts an encoder–decoder architecture enhanced with self-attention mechanisms and Bangla-specific pretrained language representations, jointly optimized via supervised sequence learning and a ROUGE-guided objective. The work systematically investigates neural modeling strategies for Bangla abstractive summarization, achieving improved generation stability without compromising inference efficiency. Evaluated on the newly constructed BanglaNews dataset, the model achieves a ROUGE-L score of 32.7—substantially outperforming extractive and conventional statistical baselines. This study bridges a critical gap in abstractive summarization research for low-resource languages and establishes a foundational framework for future work in under-resourced linguistic settings.

Low-resource languages like Hebrew lack benchmark datasets and rigorous evaluation protocols for abstractive text summarization. Method: We introduce HeSum, the first high-quality Hebrew news summarization benchmark, comprising 10,000 professionally written news articles with human-annotated abstract summaries. Distinct from prior work, we conduct the first systematic linguistic analysis of how Hebrew’s rich morphology—particularly lexical ambiguity and derivational flexibility—affects generative summarization, and design evaluation metrics grounded in linguistic validation that jointly capture abstraction and language-specific properties. Contribution/Results: Experiments reveal substantial performance degradation of state-of-the-art large language models on HeSum, underscoring its difficulty and diagnostic value. HeSum bridges critical gaps in resources and evaluation for low-resource generative NLP, establishing foundational infrastructure to advance multilingual abstractive summarization research.

This study addresses the challenges of information redundancy and coherence in Vietnamese multi-document abstractive summarization by proposing a hierarchical summarization framework. The approach first compresses individual documents using a reference-summary-guided strategy and then aggregates the compressed representations to generate the final summary, thereby enhancing content relevance at each stage and improving overall consistency. Built upon the BART architecture and augmented with external data, the system achieves a ROUGE-2 F1 score of 0.2468 on the VLSP 2022 test set. Additionally, the authors release an expanded Vietnamese multi-document summarization dataset to support further research on summarization for low-resource languages.

Addressing the challenge of generating sentiment-sensitive summaries from unstructured, short social media texts, this paper proposes the first sentiment-aware dual-path summarization framework. The framework integrates a TextRank-based extraction path enhanced with sentiment lexicon augmentation and fine-grained sentiment embeddings, alongside a UniLM-based generation path that jointly models emotional polarity and topical semantics. Unlike conventional summarization models—designed primarily for formal, structured documents—this approach explicitly incorporates sentiment signals to strengthen decision-support capabilities in brand monitoring and market analysis. Experimental results on user-generated content demonstrate substantial improvements: +18.7% in sentiment accuracy and +12.3% in ROUGE-L score (measuring information fidelity), while maintaining real-time processing capability.

This work aims to enhance the quality and accuracy of abstractive text summarization on the English subset of the XL-Sum corpus. Building upon the Transformer-based PEGASUS model, we fine-tune it on the English XL-Sum data and evaluate its performance using ROUGE metrics. Experimental results demonstrate that our approach substantially outperforms the mT5 baseline, achieving relative improvements of 4.04%, 15.25%, and 3.39% in ROUGE-1, ROUGE-2, and ROUGE-L scores, respectively. To the best of our knowledge, this constitutes the state-of-the-art performance for abstractive summarization on this dataset.

To address factual inconsistency, information loss, and poor traceability in long-document summarization, this paper proposes a sentence-level highlighting-guided self-planning generation framework. First, it identifies salient sentences via importance modeling and generates a traceable content plan; subsequently, summary generation is conditioned on this plan, effectively decoupling content selection from surface realization. This novel paradigm significantly enhances summary faithfulness and fine-grained detail retention. On the GovReport benchmark, our approach achieves a +4.1-point improvement in ROUGE-L and a 35% gain in SummaC score. Qualitative analysis confirms more complete preservation of critical details, as well as improved cross-domain accuracy and analytical depth in generated summaries.

To address the challenges of limited context windows and information redundancy in long-document summarization, this paper proposes a BART-based abstract summarization framework incorporating page-level alignment and importance-weighted scoring. The method first partitions documents into pages and introduces a page-level target text alignment mechanism to enable fine-grained supervision. Second, it designs a dynamic page importance weighting module that explicitly prioritizes semantically critical content. Third, it integrates page-wise encoding, an importance scoring network, and a partial-summary generation strategy to jointly optimize coherence and fidelity. Experiments on standard benchmarks demonstrate substantial improvements: ROUGE-1 and ROUGE-2 scores increase by 6.32% and 8.08%, respectively, surpassing state-of-the-art approaches. The framework effectively balances contextual coverage and salient information preservation, offering a principled solution for long-document abstraction under constrained transformer input lengths.