aieng-lab/en-sentiment-nrc

GRADIEND English Sentiment (NRC Adjective) Data

Masked tweet contexts where the masked word is a sentiment adjective: top 10 adjectives per valence attested as spaCy ADJ in `cardiffnlp/tweet_eval` (sentiment), with polarity taken from the NRC Emotion Lexicon (Mohammad & Turney, 2013) for target selection.

Frozen training artifact for `gradiend.examples.train_sentiment`.

Not a discrete emotion taxonomy (joy/anger/…). Binary polarity cloze over adjectives.

Companion neutrals: `aieng-lab/en-sentiment-nrc-neutral`.

Configs (subsets)

from datasets import load_dataset

# As generated
ds = load_dataset("aieng-lab/en-sentiment-nrc", "default", split="train")

# Package-paper vocabulary-held-out splits
ds = load_dataset("aieng-lab/en-sentiment-nrc", "split", split="train")
# or: load_dataset("aieng-lab/en-sentiment-nrc", split="train")  # if default is enough

Under config split, each target adjective appears in exactly one of train / validation / test.

Dataset Details

Classes

Ambiguous NRC polarity words (listed as both pos and neg) are dropped before ranking. Classes are balanced (balance="strict").

Size

Fifty rows per adjective is intentional: min_count_per_word=50 matches the lexicon attestation floor, and the rarest top-10 ADJs in tweeteval only yield ~50–60 unique maskable hits under spaCy `ADJ`. Larger `maxsizeperclass` values (e.g. 3000) do not produce a meaningfully larger unique set under strict balance.

Structure

• —masked: context with the target adjective replaced by [MASK]
• —split: train / validation / test
• —label_class: positive or negative
• —label: mask target adjective
• —feature_class_id: equal to label_class

Extra unified GRADIEND columns may be present; treat the fields above as the public API.

Sources

Dataset Creation

gradiend.examples.train_sentiment.generate_data
  require_adjectives=True
  max_words_per_class=10
  min_count_per_word=50
  max_size_per_class=500   # → 50 rows / adjective after strict balance
  balance=strict
  seed=0

HF staging writes both configs: python scripts/upload_sentiment_hf_datasets.py (see scripts/README_sentiment_hf_upload.md).

Bias, Risks, and Limitations

• —Tweet / social-media domain bias.
• —Polysemy (e.g. cool, top, real, crazy).
• —Not a validated psychological emotion instrument.

License

[CC BY 3.0](https://creativecommons.org/licenses/by/3.0/) — same as TweetEval sentiment.

Provide attribution (CC BY). Tweet content remains subject to Twitter / X Terms of Service, as required by TweetEval. The NRC Emotion Lexicon is not redistributed here; only common English adjectives appear as labels, selected using NRC polarity. For the full lexicon, see https://saifmohammad.com/WebPages/AccessResource.htm

Citation

@inproceedings{drechsel2026gradiend,
  title     = {{GRADIEND}: Feature Learning within Neural Networks Exemplified through Biases},
  author    = {Drechsel, Jonathan and Herbold, Steffen},
  booktitle = {Proceedings of the International Conference on Learning Representations},
  year      = {2026},
  url       = {https://arxiv.org/abs/2502.01406}
}

@article{mohammad2013nrc,
  title   = {Crowdsourcing a Word-Emotion Association Lexicon},
  author  = {Mohammad, Saif M. and Turney, Peter D.},
  journal = {Computational Intelligence},
  volume  = {29},
  number  = {3},
  pages   = {436--465},
  year    = {2013}
}

@inproceedings{barbieri2020tweeteval,
  title     = {{TweetEval}: Unified Benchmark and Comparative Evaluation for Tweet Classification},
  author    = {Barbieri, Francesco and Camacho-Collados, Jose and Espinosa-Anke, Luis and Neves, Leonardo},
  booktitle = {Findings of EMNLP},
  year      = {2020}
}

Dataset Card Authors

jdrechsel