isaac-relevance-ability

ISAAC relevance classifier: Ability (abled vs. disabled)

Binary classifier that decides whether an English text is relevant to the ability social group distinction, that is, whether it actually discusses people in terms of that attribute, as opposed to merely containing a keyword that can also mean something else.

This is stage 3 of the four-stage filtering pipeline that built the Illinois Social Attitudes Aggregate Corpus (ISAAC), a corpus of 527,060,919 Reddit posts (2007–2023) covering six social group distinctions.

Labels: 0 = not relevant, 1 = relevant.

What problem it solves

Keyword filters for social-group discourse are notoriously imprecise: "black" in a chess match, "disabled" in a software changelog, "gay" as a Filipino or French word. Such posts carry no signal about social groups, so they inflate standard errors, depress reliability, and suppress real effects, and most published work never reports how many of them survive. Because this model reads each keyword in the context of the full post, it separates the intended sense from the unintended one in a way a static list cannot.

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

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

texts = ["My grandmother is using a wheelchair for the first time this year."]
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 = "relevant" if int(prob.argmax()) == 1 else "not relevant"
    print(f"{label}  (P(relevant)={prob[1]:.3f})  {text}")

Training data

~1,500 posts drawn from the keyword-filtered ISAAC corpus, sampled at random with stratification by year and by per-post keyword-match count, so the training set spans the full 2007–2023 range and the full range of keyword density rather than over-representing heavily matched posts.

Each post was rated for relevance to the ability distinction by two independent trained annotators. Annotator training involved discussing the rating instructions (Appendix B of the manuscript) and working through examples together. Labels marked unclear were treated as irrelevant for training, and residual disagreements were resolved in favor of relevance.

Interrater reliability on this distinction: k = 1,476 double-rated posts, 98.2% raw agreement, Cohen's κ = .960.

The annotated validation data are released alongside the corpus.

Evaluation

Base model FacebookAI/roberta-large.

Human agreement on the training data. Two trained annotators rated k = 1,476 posts: 98.2% raw agreement, Cohen's κ = .960.

What the full filter delivers. Held-out precision describes this model in isolation. In the released corpus this classifier is one of four filtering stages, and a double-rated audit of the finished ability data (k = 100) found 6.0% of comments irrelevant under the stringent rule and 3.0% under the lenient rule (4.0% for submissions). That end-to-end figure, not the F1 above, is the number to quote when describing corpus quality.

Intended use

Screening English social-media text for relevance to the ability distinction, as a preprocessing step before substantive analysis, the role it plays in ISAAC. It is a topical relevance filter.

Out-of-scope use

This model does not measure attitudes, sentiment, toxicity, or bias. A post labeled relevant may be supportive, hostile, or neutral; relevance says only that the post is about the distinction. Do not use its output as a proxy for prejudice, and do not use it to make decisions about individuals.

It was trained on Reddit text from 2007–2023 and on the ISAAC keyword lists. Performance on other platforms, other registers, other languages, or on text that was not first keyword-filtered is unknown and likely lower. It is not a general-purpose "is this about ability" classifier for arbitrary input.

Limitations and bias

• Reddit-shaped. Reddit's user base is not representative of any general population, and its norms shifted across the 17 years covered. The model inherits that distribution.
• Keyword-conditioned. It was trained on posts that already matched an ISAAC keyword. It has not been evaluated on text that contains no such keyword.
• Trained on human judgment, which varied. Cohen's κ of .960 on the training sample means trained annotators themselves disagreed on some posts; no classifier trained on those labels can be cleaner than they were.
• Annotation conventions push toward recall at training time. Unclear posts were coded irrelevant and disagreements resolved toward relevance, both deliberate choices documented in the manuscript.

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},
  journal = {arXiv},
  eprint  = {2609.27059},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi     = {10.48550/arXiv.2609.27059},
  url     = {https://arxiv.org/abs/2609.27059}
}

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.