Datasets:

Rychanfox
/

semantic-highlighting-abstracts

license: other
license_name: arxiv-abstracts-various
license_link: https://info.arxiv.org/help/api/tou.html
task_categories:
  - token-classification
language:
  - en
tags:
  - accessibility
  - rhetorical-structure
  - scientific-papers
  - weak-supervision
pretty_name: Semantic Highlighting Rhetorical Roles

Semantic Highlighting: Rhetorical Roles in arXiv Abstracts

Rhetorical-role span annotations for ML/NLP arXiv abstracts, built for a semantic highlighting accessibility tool (dyslexia/ADHD reading support) -- see the project repo for the full research writeup (docs/findings.md).

Labels

Each span is tagged with one of 7 roles: Contribution, Method, Result, Evidence, Limitation, FutureWork, Safety. Full definitions, cue phrases, and boundary rules in docs/annotation_guideline.md.

Schema

{
  "arxiv_id": str,
  "title": str,
  "text": str,               # the abstract
  "spans": [
    {"start": int, "end": int, "label": str, "text": str},
    ...
  ]
}

start/end are character offsets into text. Spans within a record do not overlap.

Splits

train: silver-labeled abstracts (LLM weak-labeled, see Methodology), excluding every abstract that appears in test.
validation: a small held-out slice of the same silver pool, used only to watch for overfitting during training -- also silver-labeled, so treat it as noisy, not as a second gold set.
test: hand-corrected gold examples. The only split with human review.

Methodology

Abstracts fetched from the arXiv API (cat:cs.LG, cat:cs.CL, plus cat:cs.RO AND abs:safety and abs:"future work"-biased queries added specifically because Safety/FutureWork are rare under random sampling).
Weak-labeled by a local LLM (Ollama, ministral-3:latest) prompted with the annotation guideline, constrained to valid JSON matching the label schema. Every proposed span is checked against the source text character-for-character; anything the model didn't quote verbatim is dropped rather than guessed at (~9% drop rate).
A subset hand-corrected against the guideline to build test. Systematic error patterns found during correction (Evidence tagged without any number, FutureWork tagged on plain findings with no forward-looking language, overlapping/duplicate spans) were then propagated across the entire silver corpus via automated rules in scripts/clean_silver.py -- not just the hand-corrected subset.

Full detail, including exact error rates and what didn't work, in the project's findings log.

Known limitations

Silver labels are LLM-generated and imperfect even after automated cleaning; only test has human review.
Skewed label distribution: Method/Result/Contribution/Evidence are common, Safety/FutureWork/Limitation are rare even after deliberately biasing collection toward them.
test is small (21 abstracts) -- any model evaluated against it should be treated as a directional signal, not a precise number.

Licensing

Abstract text is sourced from arXiv under arXiv's API Terms of Use; individual papers retain their own licenses (arXiv itself does not claim copyright over submitted abstracts). The span annotations (labels, offsets) added by this project are original work.