esg-cid-plus 0.1.0

ESG-CID-Plus: A Stratified Benchmark for Disclosure Content Index Table Extraction from Corporate Sustainability Reports

esg-cid-plus

Supporting code for the EMNLP 2026 paper: ESG-CID-Plus: A Stratified Benchmark for Disclosure Content Index Table Extraction from Corporate Sustainability Reports

Install

uv pip install -e .

Modules

Module	Purpose
cid	Data models (CIRow, ContentIndex, FrameworkSpec) and normalization helpers
data	HuggingFace loaders — load_split, load_framework, resolve_pdf_path
pdf_toolkit	PDFSession — text extraction, lunr search, camelot tables, page labels
pdf_toolkit_ocr	OCRPDFSession — extends above with image rendering, rapidocr, img2table
extract	Rule-based pipeline: S1 detect → S2/S3 draft → S4 resolve → ContentIndex
eval	Tuple-level P/R/F1 — evaluate_predictions, format_feedback

Quickstart

from esg_cid_plus.data import load_split, load_framework, resolve_pdf_path
from esg_cid_plus.extract import extract
from esg_cid_plus.eval import evaluate_predictions

pdf_rows, cid_rows = load_split("train_small")
fw = load_framework("gri_2021")

pdf_path = resolve_pdf_path(pdf_rows[0])
ci = extract(pdf_path, fw)

gt = [r for r in cid_rows if r["report_name"] == pdf_rows[0]["report_name"]]
pred = [{"report_name": ci.report_name, "standard_id": r.standard_id,
         "page_location_indices": r.page_location_indices} for r in ci.rows]

_, overall = evaluate_predictions(gt, pred)
print(overall)

Tests

uv run pytest              # fast (unit)
uv run pytest -m slow      # integration — requires HuggingFace access