conversion_script.py

#!/usr/bin/env python3

"""
Run this script as ./conversion_script.py to convert the Causal News Corpus
(CNC) "V2" -- published as RECESS -- DIRECTLY from its original repository.
``UniCausal2HF`` already reads a CSV given any path/URL pandas' own
``read_csv`` accepts, so no manual download/caching step is needed; point
it straight at the raw GitHub URLs.

Citation / original source
---------------------------
Tan, F. A., Hettiarachchi, H., Hürriyetoğlu, A., Oostdijk, N., Caselli,
T., Nomoto, T., Uca, O., Liza, F. F., & Ng, S.-K. (2023). "RECESS:
Resource for Extracting Cause, Effect, and Signal Spans." IJCNLP-AACL
2023. https://aclanthology.org/2023.ijcnlp-main.6/
Repo (verified live, public, no login): github.com/tanfiona/CausalNewsCorpus
License: CC0-1.0 (verified via GitHub API) -- public domain, no restrictions.

This is the actively-maintained release -- the maintainers themselves
recommend using it over the original 2022 "V1" release ("For 2023 Shared
Task, please use V2" -- repo README): far richer span annotations (subtask
2, Cause/Effect/Signal), 2257 causal relations vs. V1's 183 (verified:
`train_subtask2_grouped.csv` + `dev_subtask2_grouped.csv` row counts). See
../CNC/conversion_script.py for V1, kept as its own separate dataset for
comparison rather than silently overwritten by this newer version.

Format: the ``_grouped`` CSVs already match causalatee's ``UniCausal2HF``
grouped format exactly (one row per sentence; ``causal_text_w_pairs`` is a
Python-repr'd list of 0+ independently <ARG0>/<ARG1>/<SIGn>-tagged copies
of that row's ``text``, one per causal relation -- ARG0=cause, ARG1=effect,
confirmed against real "because"/"due to" examples) -- see
causalatee/data/conversion/_unicausal2hf.py for the shared parsing logic
(also used by BECauSEv2, AltLex).

No usable test split: the real held-out test set (``test_subtask2_text.csv``)
has NO gold labels at all and never will -- confirmed directly from
data/V2/README.md: "We will not release the test labels this year as we
might intend to rerun this Shared Task next year." Mapped here as:
upstream train -> causalatee train, upstream dev -> causalatee dev.parquet
-- written as ``dev.parquet``, NOT ``test.parquet`` (fixed 2026-07-20;
previously written as test.parquet, which claimed a real held-out test
set exists when it's actually the upstream dev split). `with_validation_split`
in the evaluation harness treats a dataset with dev but no test as:
promote this dev to serve as the final eval target, and carve a FRESH
validation split out of train for early stopping instead, so the
promoted dev is never touched during training either way -- see its
docstring.
"""

from pathlib import Path

from causalatee.data.constants import Task
from causalatee.data.conversion import UniCausal2HF

_BASE_URL = "https://raw.githubusercontent.com/tanfiona/CausalNewsCorpus/master/data/V2"

converter = UniCausal2HF(
    {
        "train": f"{_BASE_URL}/train_subtask2_grouped.csv",
        "dev": f"{_BASE_URL}/dev_subtask2_grouped.csv",
    },
    Path.cwd(),
)

for split in ["train", "dev"]:
    converter.convert(Task.CausalityDetection, split)
    converter.convert(Task.CausalCandidateExtraction, split)
    converter.convert(Task.CausalityIdentification, split)