ISAAC generalization: clause segmenter

Same weights as the public DiSCo release. This repository is the ISAAC-facing copy of BabakScrapes/disco-clause-segmenter; the checkpoints are identical. That card carries the full model description and the DiSCo corpus details. This one documents how the model is used inside the ISAAC pipeline and restates the performance figures reported in the ISAAC manuscript. Cite whichever matches your use; if you are using ISAAC labels, cite both.

Token classifier that splits English text into clauses, the unit the situation-entity classifier then labels. Together the two models produce the clause-level generalization columns of the Illinois Social Attitudes Aggregate Corpus (ISAAC).

Decoding

The model emits one tag per word (majority-voted across sub-word tokens). The decoder in code/label_generalization.py reads them as follows:

• tag 2 marks a clause-final word; it closes the clause it appears in;
• tags 0 and 1 mark clause-internal words;
• a new clause opens on the word after a 2;
• when a word receives no aligned prediction, it defaults to 1.

Reference implementation, including the sub-word majority vote:

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

REPO = "ISAAC-corpus/isaac-generalization-segmentation"
tokenizer = AutoTokenizer.from_pretrained("roberta-base", use_fast=True,
                                          add_prefix_space=True)
model = AutoModelForTokenClassification.from_pretrained(REPO).eval()

text = "My gay neighbor watered my plants while I was traveling"
words = text.split()
enc = tokenizer(words, is_split_into_words=True, return_tensors="pt",
                truncation=True, max_length=512, padding="max_length")
word_ids = enc.word_ids(batch_index=0)
with torch.no_grad():
    token_preds = model(**enc).logits[0].argmax(dim=-1).tolist()

per_word = [[] for _ in words]
for token_idx, word_id in enumerate(word_ids):
    if word_id is not None:
        per_word[word_id].append(token_preds[token_idx])
tags = [max(set(p), key=p.count) if p else 1 for p in per_word]

clauses, current, prev = [], [], 2
for word, tag in zip(words, tags):
    if prev == 2:
        current = []
    current.append(word)
    if tag == 2 and prev in (0, 1):
        clauses.append(" ".join(current))
        current = []
    prev = tag
if current:
    clauses.append(" ".join(current))

print(clauses)

Texts longer than 200 words are split on sentence boundaries before segmentation in the ISAAC pipeline; see generalization.py in the Space.

Training data

The DiSCo corpus of opinionated, mixed-register English text (Hemmatian, 2022), with human-verified clause boundaries. See the DiSCo card.

Evaluation

Two FacebookAI/roberta-base models run in sequence: a clause segmenter, then a 18-way situation-entity classifier. Both are the same weights as the public DiSCo release; see the clause segmenter and situation-entity classifier cards for the full model description and the corpus they were trained on.

Segmentation. The segmenter covered most of the target clause span for 95.5% of human-verified clauses.

Classification, on the held-out 10% of the disco gold corpus (k ≈ 2,357):

The 18-way macro F1 of .514 is held down by heavy label imbalance across the rarer situation-entity types. The three collapsed features are what ISAAC actually reports, and they are the numbers to rely on.

Intended use

Clause segmentation as the first stage of the ISAAC generalization pipeline.

Out-of-scope use

Not a general-purpose syntactic parser, constituency parser, or sentence splitter. It targets the specific clause unit the situation-entity framework requires, which does not always coincide with a syntactic clause. English only.

Limitations and bias

• Coverage is reported as the share of human-verified clauses whose target span the model mostly recovered; it is not an exact-match boundary F1, and exact boundaries are frequently off by a word.
• Segmentation errors propagate into every downstream generalization label.
• Long inputs are truncated at 512 sub-word tokens, which is why the ISAAC pipeline pre-splits at ~200 words.
• Reddit text is noisy (missing punctuation, run-on constructions, markup), and segmentation quality degrades accordingly.

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.