ISAAC-corpus/isaac-moralization

ISAAC moralization classifier

Binary classifier that decides whether an English text frames its subject in moral terms — judgments of right and wrong, harm, fairness, loyalty, authority, purity — as opposed to describing it non-morally.

This is the document-level moralization labeler used to annotate all 527,060,919 posts in the Illinois Social Attitudes Aggregate Corpus (ISAAC).

Labels: 0 = non-moralized, 1 = moralized.

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

REPO = "ISAAC-corpus/isaac-moralization"
tokenizer = AutoTokenizer.from_pretrained(REPO)
model = AutoModelForSequenceClassification.from_pretrained(REPO).eval()

texts = [
    "People who cut in line are selfish and should be ashamed of themselves.",
    "The bus arrives at the corner of 5th and Main every fifteen minutes.",
]
enc = tokenizer(texts, padding=True, truncation=True, max_length=512,
                return_tensors="pt")
with torch.no_grad():
    probs = torch.softmax(model(**enc).logits, dim=1)

for text, prob in zip(texts, probs):
    label = "moralized" if int(prob.argmax()) == 1 else "non-moralized"
    print(f"{label}  (P(moralized)={prob[1]:.3f})  {text}")

No thresholding: the deployed rule is plain argmax.

Training data

The Moral Foundations Reddit Corpus (MFRC; Trager et al., 2022), with its foundation-level annotations reduced to a binary moralized / non-moralized target. Disagreements between MFRC annotators were resolved by majority vote; residual ties were broken toward the moralized label, to maximize sensitivity to moral content.

MFRC was chosen over Twitter-trained alternatives such as MFTC because it samples the same platform as ISAAC, and platform-specific language style and conversational context matter for this construct.

Why binary rather than per-foundation

Two reasons, both deliberate. First, published per-foundation classification performance is highly uneven; applying unevenly performing labels across ISAAC's six social group distinctions would produce corresponding unevenness in downstream construct validity. Second, a binary moralization construct commands broader theoretical agreement than the individual foundations, around which debate continues.

If you need foundation-level labels, this is not the model.

Evaluation

Base model google-bert/bert-base-uncased, fine-tuned on the Moral Foundations Reddit Corpus (MFRC; Trager et al., 2022) reduced to a binary moralized / non-moralized target.

The evaluation slice is 49.1% moralized, so accuracy is interpretable against a ~50% baseline.

Binary by design. Per-foundation classification performance in the published literature is highly uneven, and a binary moralization construct commands broader theoretical agreement than the individual foundations. This model therefore does not return moral-foundation labels.

Face validity in the corpus

Applied across ISAAC, the classifier reproduces the expected ordering: highly contested domains moralize more than less contested ones (race 68.1%, ability 72.2%, sexuality 69.5%, versus body weight 60.2%).

Note that 49–74% moralized is far above the 2–5% reported for unselected everyday speech and donated personal social media (Atari et al., 2023). That gap is expected — ISAAC is pre-filtered for relevance to social distinctions that attract intense normative scrutiny, and a binary operationalization is more inclusive than foundation-specific coding — but it means the base rate here should not be read as a population estimate.

Intended use

Document-level annotation of English social-media text for the presence of moral framing, at scale, in aggregate research designs.

Out-of-scope use

• —Not a moral judgment. The model detects that moral language is being used, not whether the position taken is right, and not whether the author is moral.
• —Not per-foundation. See above.
• —Not for individual-level decisions. Accuracy of .755 on a balanced held-out set is useful for aggregate estimates over hundreds of thousands of posts; it is not adequate for consequential judgments about a single author or a single post.
• —Reddit-shaped, English-only. Trained and evaluated on Reddit comments. Performance on other platforms, registers, or languages is unknown.

Limitations and bias

• —Moralization is a contested construct with genuine annotator disagreement in the source corpus; the ceiling for any model trained on it is well below 1.0.
• —The tie-breaking rule (ties → moralized) means the model is tuned to be sensitive rather than conservative, and will over-call ambiguous cases.
• —MFRC samples a limited set of subreddits, so topical coverage of moral language is narrower than ISAAC's.

Links

Citation

Please cite the ISAAC paper. One citation covers the whole project — the corpus, the pipeline, and every model. Please do not cite this model repository separately; keeping references in one place is what allows the project's citations to be found together.

@article{hemmatian2026isaac,
  author = {Hemmatian, Babak and Hadjarab, Sarah and Chen, Jessica and Kurdi, Benedek},
  title  = {The {Illinois} Social Attitudes Aggregate Corpus ({ISAAC}): An Open Tool and Reproducible Pipeline for Analyzing Social Group Discourse at Scale},
  year   = {2026},
  note   = {Manuscript submitted for publication}
}

License

Released under a Creative Commons Attribution 4.0 International License. You may use, share, and adapt these weights, including commercially, provided you give appropriate credit — see Citation above.

The ISAAC corpus itself is governed separately by the project Data Use Agreement.