Datasets:

Hamyrappy
/

emnlp-2020-2025-atomic-claims

language:
  - en
license: cc-by-4.0
pretty_name: EMNLP 2020-2025 Atomic Contribution Claims (ACC) with Drift Clusters
tags:
  - synthetic
  - scientometrics
  - science-of-science
  - acl-anthology
  - emnlp
  - atomic-claims
  - claim-extraction
  - clustering
  - scientific-text
  - nlp
annotations_creators:
  - machine-generated
language_creators:
  - machine-generated
multilinguality:
  - monolingual
size_categories:
  - 10K<n<100K
source_datasets:
  - original
task_categories:
  - feature-extraction
  - text-classification
configs:
  - config_name: claims
    data_files:
      - split: train
        path: claims.parquet
    default: true
  - config_name: papers
    data_files:
      - split: train
        path: papers.parquet
  - config_name: clusters
    data_files:
      - split: train
        path: clusters.parquet
  - config_name: cluster_summary
    data_files:
      - split: train
        path: cluster_summary.parquet

EMNLP 2020–2025 Atomic Contribution Claims (ACC), with drift clusters

18,293 atomic contribution claims extracted from the abstracts of the full EMNLP main track 2020–2025, plus the canonical 80-cluster drift clustering and per-cluster drift statistics used in the Drift Inspector paper.

An atomic contribution claim (ACC) is a single self-contained sentence stating one concrete contribution of a paper: atomic (one contribution-bearing proposition), decontextualized (pronouns resolved, meta-language removed), and falsifiable (a verifiable assertion).

TheoremLlama uses curriculum learning and block training techniques to train large language models for formal theorem proving. — from 2024.emnlp-main.667

This is the single-venue case-study corpus of the Drift Inspector project — the corpus every number in the paper's evaluation is computed on. For breadth, see the full-anthology release (346k claims, 423 venues):

🤗 Full-anthology dataset: acl-anthology-atomic-claims
🔭 *Live demo (extended, 6 ACL venues): https://hamyrappy.github.io/drift-inspector-acl/
📄 This corpus explored interactively: https://hamyrappy.github.io/drift-inspector-emnlp/
🛠 Code (extraction + clustering + inspector): https://github.com/Hamyrappy/drift-inspector

What's inside

Config	Rows	Contents
`claims` (default)	18,293	one row per extracted claim (4,937 papers yielded claims)
`papers`	6,517	all EMNLP main-track papers 2020–2025, with abstracts
`clusters`	16,576	the paper's canonical clustering: claim → cluster, on a year-balanced subsample
`cluster_summary`	81	per-cluster names, sizes, per-year document frequency, drift statistics

`claims` fields

Column	Type	Description
`claim_id`	string	unique id (`clm_` + hash)
`paper_id`	string	ACL Anthology id, e.g. `2021.emnlp-main.602`
`year`	int32	2020–2025
`atomic_claim`	string	the claim text
`claim_index`	int32	position within the paper's extraction (0-based)
`extractor_model`	string	`qwen/qwen3-235b-a22b-thinking-2507`
`prompt_hash`	string	`f9a9b434c72c` — exact prompt ships with the code
`run_id`, `created_at`	string	extraction provenance

`papers` fields

paper_id, year, title, venue (EMNLP), track (main), url, authors, author_ids, n_authors, abstract (all abstracts are 2020+, hence CC BY 4.0).

`clusters` fields — the canonical drift clustering

The paper's analysis runs on a year-balanced subsample: 748 papers per year (seed 42) → 4,488 papers, 16,576 claims. Claims were embedded with SPECTER2, reduced with UMAP, clustered with HDBSCAN (BERTopic-style, but the input unit is the claim, not the abstract) → 80 clusters + noise (36% of claims; rising 31% → 42% across years as the newest contributions have had the least time to consolidate).

Column	Type	Description
`claim_id`	string	joins to `claims`
`paper_id`, `year`, `atomic_claim`		copied for standalone use
`cluster`	int32	cluster id; `-1` = noise
`cluster_descriptor`	string	extractive c-TF-IDF descriptor, e.g. `attention, parameters, transformer`
`cluster_short_name`	string	LLM-generated, author-reviewed name, e.g. `Transformers & Efficiency`
`cluster_full_name`	string	longer readable name

`cluster_summary` fields

One row per cluster (incl. noise): cluster, cluster_short_name, cluster_full_name, cluster_descriptor, cluster_description (one-paragraph scope note), tags, n_claims, n_papers, df_pct_2020 … df_pct_2025 (share of that year's 748 papers with ≥1 claim in the cluster, %), drift_log2_2025_2020 (log2 ratio of 2025 vs 2020 share, ε = 0.1pp — the Drift Inspector's drift color).

Example drift readings: Syntactic Parsing 5.08% → 0.13% of papers; Math & Logic Reasoning 0.94% → 7.62%.

from datasets import load_dataset
clusters = load_dataset("Hamyrappy/emnlp-2020-2025-atomic-claims", "clusters", split="train")

Dataset creation & validation

Corpus. Full EMNLP main track 2020–2025 from the ACL Anthology (751 / 847 / 826 / 1,047 / 1,237 / 1,809 papers per year).
Extraction. qwen3-235b-a22b-thinking-2507 (via OpenRouter, temperature 0.2), few-shot prompt enforcing atomicity / decontextualization / falsifiability, excluding background, motivation and raw metric claims. Abstracts with no extractable contribution are excluded (e.g. 3 of 751 in 2020).
Human validation. Three annotators labeled 180 items (136 stratified claims + 44 curated negative controls) as Good/Bad/Unsure. Inter-annotator agreement Fleiss' κ = 0.844 (dropping Unsure; 0.76/0.73 under stricter conventions); 97.8/91.8/90.5% Good on sampled claims; 44/44 negatives flagged by 2-of-3 majority. The full protocol and annotations ship with the code repository.
LLM judge. An independent judge from a different vendor (Gemini vs. Qwen), blind to human labels, matches the human majority with 94.5% accuracy (κ = 0.863) and flags 43/44 negatives.
Clustering. SPECTER2 → UMAP → HDBSCAN on the balanced subsample; 80 clusters, c-TF-IDF descriptors, LLM-generated author-reviewed names. Headline drift directions are stable under bootstrap resampling, eight alternative encoders, and a lexical representation (see paper).

Limitations

Claims are LLM-extracted; residual errors (unsupported details, background leakage) exist — the validation above quantifies, not eliminates, them.
Abstract-level only: body-only contributions are invisible.
Cluster boundaries and names depend on UMAP/HDBSCAN hyperparameters and the encoder; headline drift directions are robustness-checked, per-cluster numbers are descriptive.
The clustering covers the balanced subsample (16,576 of 18,293 claims); the remaining claims are extracted but unclustered.

Licensing

Released under CC BY 4.0. Claim texts are LLM-generated derived statements; source abstracts (included in papers) are ACL Anthology materials published 2020–2025, licensed CC BY 4.0. Please attribute the ACL Anthology and cite the underlying papers where appropriate.

Citation

The system paper introducing ACCs and Drift Inspector is under review; until then, please cite the dataset directly:

@misc{acc-emnlp-2026,
  title        = {EMNLP 2020--2025 Atomic Contribution Claims (ACC) with Drift Clusters},
  author       = {Karimov, Vsevolod and Ostarkov, Stepan and Poroshina, Anastasia and Frolov, Anatoly and Panchenko, Alexander},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Hamyrappy/emnlp-2020-2025-atomic-claims}}
}