This dataset contains 9,737 scientific papers from the ACL Anthology, paired with their human-written abstracts, curated for abstractive scientific summarization.

Unlike full-paper datasets, this dataset emphasizes signal quality over raw scale, focusing on sections that contribute most to summary content. The dataset is intended for research on faithful abstraction, compression, and long-document summarization under realistic constraints.

Curated by: Independent research project
Language(s): English
License: Apache 2.0
Total Examples: 9,737

Dataset Sources

Repository: ACL Anthology
Original Papers: https://aclanthology.org/

Dataset Structure

Data Instances

Each example contains:

content: Full or partial scientific paper text (Source).
abstract: Gold standard human-written abstract (Target).

Data Statistics

Metric	Value
Mean Source Length	6,480.5 tokens
Mean Summary Length	164.7 tokens
Mean Compression Ratio	0.028x
Total Vocabulary Size	134,817

Dataset Quality Evaluation

A comprehensive automated evaluation was performed to assess the relationship between source documents and summaries.

1. Alignment & Faithfulness

Mean Alignment: 0.4578 (Indicates moderate semantic overlap between source and summary).
Extractive Density: 0.0608 (Very low, suggesting highly abstractive summaries with a high risk of "hallucination" if not constrained).
Novel Entity Ratio: 0.4258 (High number of entities in abstracts not explicitly found in the source text).

2. Coverage & Salience

Mean Coverage: 0.0073 (Summaries cover a very specific, narrow slice of the source content).
Lead Bias: 0.1262 (Low lead bias; information is distributed rather than just taken from the introduction).

3. Diversity & Difficulty

Diversity Quality: High (Summary Type-Token Ratio: 0.0211).
Difficulty Distribution: Mean difficulty score of 6,314.
Example Split: 2,430 easy examples vs. 2,432 hard examples based on length and alignment.

4. Overall Assessment

Grade: Needs Improvement (Score: 0.36) The dataset shows high abstractive difficulty. Key areas for caution:

Low content alignment: Summaries may contain external knowledge not present in the provided source text.

Low extractive density: Models may struggle to remain faithful to the source.

Low coverage: Summaries are highly compressed, missing significant portions of the paper.

Uses

Direct Use

Fine-tuning encoder–decoder models (e.g., BART, T5, LongT5) for abstractive summarization.
Research on hallucination detection and faithfulness.
Benchmarking long-document compression.

Out-of-Scope Use

Extractive summarization (the summaries are too abstractive).
Applications requiring guaranteed factual correctness without external grounding.

Downloads last month: 21