structured_paper_summarization

A 151 k‑example dataset of chat‐style prompt → structured abstract pairs, built from ~19 000 research papers across business, management, information‑systems and social‑science domains. Each example shows the full paper (body text) being summarised into a five‑section Emerald‑style structured abstract (Purpose, Design/methodology/approach, Findings, Practical implications, Originality/value).

---

Why this dataset?

Large‑language models (LLMs) frequently struggle to:

1. Condense long scientific prose into factual, concise summaries.
2. Follow rigid output structures (e.g. subsection headings).

This dataset targets both challenges simultaneously, enabling fine‑tuning or instruction‑tuning of LLMs that must output structured scholarly abstracts.

---

At a glance

Split	Rows	Size (compressed)
train	145 067	626 MB
test	6 650	29 MB
Total	151 717	≈655 MB

^{Counts taken from the Hugging Face viewer on 2025‑04‑29.}

---

Data schema

{
  title:    string                    # Paper title
  keywords: list[string]             # Author‑supplied keywords (0‑23)
  messages: list[dict] length ≥ 2    # ChatML‑style conversation
}

messages format

Each list contains alternating dictionaries with:

• role: either "user" or "assistant".
• content: UTF‑8 text.

Typical pattern (2 items):

[
  {
    "role": "user",
    "content": "Summarize the following paper into structured abstract.\n\n<full paper text>"
  },
  {
    "role": "assistant",
    "content": "Purpose: …\nDesign/methodology/approach: …\nFindings: …\nPractical implications: …\nOriginality/value: …"
  }
]

Some papers are longer and may be truncated to ~8 k tokens.

---

Loading the data

from datasets import load_dataset

ds_train = load_dataset(
    "Neooooo/structured_paper_summarization", split="train"
)
print(ds_train[0]["messages"][1]["content"][:500])

The dataset is stored as Apache Parquet with streaming support; the example above requires ~5 s to start iterating with no local download.

---

Suggested use‑cases

• Instruction‑tuning chat LLMs for long‑document summarisation.
• Research on controlled text generation and output formatting.
• Training retrieval‑augmented systems that must cite sections of the source paper.

---

Source & construction

1. Full‑text articles were collected via institutional access to the Emerald Insight corpus (open‑access + subscription).
2. The canonical structured abstract supplied by each journal was extracted as ground truth.
3. The article’s main body was embedded into a prompt of the form shown above.
4. Data were converted to Hugging Face datasets ➜ auto‑parquet.

No additional manual cleaning was performed; typos and OCR artefacts may persist.

---

Licensing & acceptable use

The article texts are copyright their original publishers/authors and are redistributed here solely for non‑commercial research. By using this dataset you agree to:

• Not redistribute the raw paper texts.
• Cite the original articles in any derivative work.
• Abide by Emerald’s usage policy and your local copyright laws.

The metadata & structured abstracts are released under CC BY‑NC 4.0. For commercial licensing, please contact the original rights‑holders.

---

Citation

If you use this dataset, please cite:

@dataset{hu_2025_structured_prompts,
  author       = {Xingyu Hu},
  title        = {structured_paper_summarization},
  year         = 2025,
  url          = {https://huggingface.co/datasets/Neooooo/structured_paper_summarization},
  note         = {Version 1.0}
}

---

Contributions

Feel free to open PRs to:

• Fix metadata errors.
• Provide additional splits (validation, domain‑specific subsets).
• Add scripts for evaluation or preprocessing.

---

Happy summarising!