EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

SGATSUM: Semantic Graph Attention Network for Extractive Summarization

Jianqiang Chen, Yi Wang, Yijun Liu

IEEE Access · 2026 · 14 · 50489-50505

DOI: 10.1109/access.2026.3679555 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Jianqiang Chen und 2 weiteren Autor:innen. Die Arbeit nennt 58 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Guangzhou University of Software.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the vigorous development of natural language processing technology, traditional encoder-decoder extractive summarization models have achieved certain effects in the summarization generation task. However, they have limitations in modeling semantic associations between sentences and are prone to missing multi-level semantic information—a limitation confirmed by Zhong et al., which shows that such models miss 23% of inter-sentence semantic relationships in long documents. Given that the semantic information is useful for improving the performance of the summarizers, we apply graph architecture to capture it in the summarization process. In this work, we propose a Semantic Graph Attention network for extractive SUMmarization (SGATSUM) which leverages different levels of semantic information to guide the summarization process. It contains four modules including an embedding module, local-semantics module, global-semantics module, and linear classifier. Specifically, the local semantic information enhances SGATSUM to recognize different types of idiomatic phrases and conjunctions. These means of expression will intuitively improve the coherence and readability of summaries. To improve the coverage of summary, the global-semantics module helps SGATSUM transmit the information within each sentence for mining the hidden relationships. We also propose a pre-training task for training SGATSUM to further recognize the semantic difference between two sentences. In a comparison of the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) performance of summarizers on the CNN/DailyMail and SciSumNet datasets, SGATSUM outperforms baseline models. Furthermore, SGATSUM also behaves better than baseline models in experiments about extraction probabilities of sentences.

Förderung

  • Guangzhou University of Software

Zitieren

APA

Chen, J., Wang, Y., & Liu, Y. (2026). SGATSUM: Semantic Graph Attention Network for Extractive Summarization. IEEE Access, 14, 50489-50505. https://doi.org/10.1109/access.2026.3679555

Vancouver

Chen J, Wang Y, Liu Y. SGATSUM: Semantic Graph Attention Network for Extractive Summarization. IEEE Access. 2026;14:50489-50505. doi:10.1109/access.2026.3679555

BibTeX
@article{Chen2026,
  author = {Chen, Jianqiang and Wang, Yi and Liu, Yijun},
  title = {SGATSUM: Semantic Graph Attention Network for Extractive Summarization},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {50489--50505},
  doi = {10.1109/access.2026.3679555}
}
Kostenlos starten →