Elamur Research⌕
☰
home / papers / 2026 / Normalizing Section Names and Structure of Scientific Articles

Normalizing Section Names and Structure of Scientific Articles

Paper · 2026 · Nicolau Duran-Silva, Julian Moreno-Schneider, César A. Parra-Rojas et al.

Summary

Parsers for scientific PDFs often lose or garble section boundaries and hierarchy, and authors title sections in many different ways. This makes it hard to align papers for RAG, question answering, or scientometric tasks such as citation-context analysis. The authors treat section name normalization as hierarchical classification over a two-level taxonomy: coarse Level-1 roles such as Method, each with fine-grained Level-2 subtypes such as data preprocessing. Section depth is predicted as main section versus subsection. They compare three paradigms on the SASC dataset:

  • SciBERT classifiers that label each section on its own;
  • a sequential model in which a Transformer encoder runs over the ordered section embeddings of a whole document;
  • zero-shot prompting of instruction-tuned LLaMA models.

The discriminative models beat the generative ones in every setting. Document-level context gives the sequential model a consistent edge, and adding section text is crucial for fine-grained labels. The work argues that structure-aware modelling is needed for large-scale scholarly document processing and research-assessment tools.

Key points

  • Dataset: SASC (released at github.com/sirisacademic/sasc) contains 4,896 full-text papers with about 117k sections.
    • Its papers come from four domains (Energy, Cancer, Neuroscience, Transport) plus a sample drawn from general science.
    • Each section has its raw title, content, explicit number (when available), Level-1/Level-2 labels, and hierarchical level.
  • Taxonomy: 11 Level-1 categories, from Abstract through Metadata, References, and Other, each with multiple Level-2 subtypes. Hierarchy levels deeper than subsections are merged into their parent.
  • Independent SciBERT setup: the input is the header plus the first 50 content tokens. Adding content raised Level-1 accuracy from 0.650 (header only) to 0.777, and joint Level-1+Level-2 accuracy from 0.481 to 0.579.
  • Generative models: LLaMA 1B and 8B were tested zero-shot with document-level and section-level prompts.
    • Both lagged behind the classifiers, especially on fine-grained labels.
    • The 8B model was clearly better in the section-level setting, but the 1B model did better at document level.
    • The authors attribute the weak results to output parsing problems, poor retention of the taxonomy, and hallucinations.
  • Per-category results: frequent, clearly defined sections were classified best, such as Method, Conclusion, Metadata, and References. Heterogeneous subtypes, such as several Introduction and Appendix subcategories, scored much lower, with some F1 values at 0.00.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🔬German Research Centre for Artificial IntelligenceKaiserslautern · Germany77 works · 29,479 citations🎓Universitat Pompeu FabraBarcelona · Spain47 works · 11,723 citations🏢SIRIS AcademicBarcelona · Spain2 works

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Nicolau Duran-Silva, Julian Moreno-Schneider, César A. Parra-Rojas and Georg Rehm. 2026. Normalizing Section Names and Structure of Scientific Articles. Proceedings of Natural Scientific Language Processing (NSLP) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/5ftk2fxxf7jd

@inproceedings{duransilva2026normalizing,
  title = {{Normalizing Section Names and Structure of Scientific Articles}},
  author = {Nicolau Duran-Silva, Julian Moreno-Schneider, César A. Parra-Rojas and Georg Rehm},
  booktitle = {Proceedings of Natural Scientific Language Processing (NSLP) @ LREC 2026},
  year = {2026},
  doi = {10.63317/5ftk2fxxf7jd},
  url = {https://doi.org/10.63317/5ftk2fxxf7jd}
}

References

18 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Pablo Accuosto and Horacio Saggion. 2019. Discourse-driven argument mining in scientific abstracts. In International Conference on Applications of Natural Language to Information Systems, pages 182–194. Springer. find
  2. Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciB-ERT: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3615–3620,… find
  3. Decheng Duan, Jitong Peng, Yingyi Zhang, and Chengzhi Zhang. 2025. SciNLP: A domain-specific benchmark for full-text scientific entity and relation extraction in NLP. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 14473– 14486, Suzhou, China. Associa… find
  4. Ingo Frommholz, Philipp Mayr, Guillaume Cabanac, Suzan Verberne, and Christin Katharina Kreutz. 2025. The first workshop on scholarly information access (scolia). In European Conference on Information Retrieval, pages 326–331. Springer. find
  5. Tirthankar Ghosal, Philipp Mayr, Anita De Waard, Aakanksha Naik, Amanpreet Singh, Dayne Freitag, Georg Rehm, Sonja Schimmler, and Dan Li. 2025. Overview of the fifth workshop on scholarly document processing. In Proceedings of the Fifth Workshop on Scholarly Document Processing (SDP 2025), pages 1–6… find
  6. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, https://www.nfdi4datascience.de ↗
  7. Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  8. Garapati Keerthana and Manik Gupta. 2025. Cli-rag: A retrieval-augmented framework for clinically structured and context aware text generation with llms. arXiv preprint arXiv:2507.06715. normalization under explicit output constraints. Prompts enumerate the allowed section taxonomy and enforce deter… — [arXiv:2507.06715](https://arxiv.org/abs/2507.06715) find
  9. Document-Level Section Classification Prompt Gabriela F Nane, Nicolas Robinson-Garcia, François van Schalkwyk, and Daniel Torres-Salinas. 2023. Covid-19 and the scientific publishing system: growth, open access and scientific fields. Scientometrics, 128(1):345–362. find
  10. You are an expert scientific document section normalizer. TASK: For each section below, assign exactly ONE label from the ALLOWED LABELS list. ALLOWED LABELS: {L1 \| L2 taxonomy listing} Michael D Skarlinski, Sam Cox, Jon M Laurent, James D Braza, Michaela Hinks, Michael J Hammerling, Manvitha Ponnap… find
  11. RULES: - Use "L1\|L2" when a specific subtype applies. - Use only "L1" when no subtype is appropriate. - Output exactly N labels, one per section, in order. - Output ONLY a JSON array of strings. No explanations. find
  12. Luciana B Sollaci and Mauricio G Pereira. 2004. The introduction, methods, results, and discussion (imrad) structure: a fifty-year survey. Journal of the medical library association, 92(3):364. find
  13. DOCUMENT: [1] HEADER: {SECTION_HEADER_1} \| CONTENT: { CONTENT_SNIPPET_1} [2] HEADER: {SECTION_HEADER_2} \| CONTENT: { CONTENT_SNIPPET_2} ... [N] HEADER: {SECTION_HEADER_N} \| CONTENT: { CONTENT_SNIPPET_N} Deep Search Team. 2024. Docling technical report. Technical report. find
  14. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.
  15. Section-Level Prompt with Local Context You are an expert scientific document section normalizer. find
  16. Thanasis Vergoulis, Serafeim Chatzopoulos, Kleanthis Vichos, Ilias Kanellos, Andrea Mannocci, Natalia Manola, and Paolo Manghi. 2022. Bip! scholar: a service to facilitate fair researcher assessment. In Proceedings of the 22nd ACM/IEEE Joint Conference on Digital Libraries, pages 1–5. find
  17. TASK: Classify the CURRENT section into exactly ONE label from the ALLOWED LABELS. ALLOWED LABELS: {L1 \| L2 taxonomy listing} SURROUNDING CONTEXT: PREVIOUS SECTION: {PREV_HEADER} (label: { PREV_LABEL}) NEXT SECTION: {NEXT_HEADER} David Wadden, Kyle Lo, Lucy Lu Wang, Arman Cohan, Iz Beltagy, and Hann… find
  18. CURRENT SECTION: HEADER: {CURRENT_HEADER} CONTENT: {CONTENT_SNIPPET} OUTPUT: Return ONLY a single label string (e.g., " method\|data_collection", "introduction\|background", or "references"). 9. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/silva-2026-normalizing-section-names-structure-scientific/