conversion_script.py

#!/usr/bin/env python3

"""
Run this script as ./conversion_script.py to convert the Cause-Effect
relation subset of SemEval-2007 Task 4 ("Classification of Semantic
Relations between Nominals") into HF-compatible parquet files.

Citation / original source
---------------------------
Girju, R., Hearst, M., Nakov, P., Nastase, V., Szpakowicz, S., Turney, P.,
& Yuret, D. (2007). "SemEval-2007 Task 04: Classification of Semantic
Relations between Nominals." Proceedings of the 4th International
Workshop on Semantic Evaluations (SemEval-2007), pages 13-18.
https://aclanthology.org/S07-1003/
License: CC BY-SA 2.5 (verified: the dataset's own bundled
`copyright.txt` states "The Complete Dataset and the Trial Dataset for
Task 4 are released under the Creative Commons Attribution-Share Alike
2.5 License").

The task covers SEVEN semantic relations between nominal pairs
(Cause-Effect, Instrument-Agency, Product-Producer, Origin-Entity,
Theme-Tool, Part-Whole, Content-Container), one file per relation
(``relation-1`` .. ``relation-7``); only relation-1 (verified directly
by reading its own definition PDF: "Cause-Effect") is used here.

Repo (verified live, public, no login, byte-identical to the original
archive -- diffed relation-1's train/test/key files directly against a
copy fetched from the official SemEval-2007 distribution before trusting
it): github.com/davidsbatista/Annotated-Semantic-Relationships-Datasets,
``datasets/SemEval2007-Task4.tar.gz`` (a plain, uncompressed POSIX tar
despite the ``.tar.gz`` name -- verified via ``file``), containing
nested ``train.tar.gz``/``test.tar.gz``/``key.tar.gz``.

Format: each relation file is a sequence of blank-line-separated records:
  ``NNN "sentence with <e1>...</e1> and <e2>...</e2> markers"``
  ``WordNet(e1) = "...", WordNet(e2) = "...", Cause-Effect(eX,eY) = "true"/"false"/"?", Query = "..."``
  optionally followed by a ``Comment:`` line.
The ``<e1>``/``<e2>`` markers are ALREADY in this project's own marker
format (verified: `causalatee.data.utils.parse_entity_markers` parses
them directly with no preprocessing). The ``Cause-Effect(eX,eY)``
argument order gives the (cause, effect) role assignment for THAT
record -- verified this is NOT fixed: 130/140 train records use
``(e2,e1)`` but 10/140 use ``(e1,e2)`` (e.g. record 011:
"<e1>Zinc</e1> is essential for <e2>growth</e2>", Cause-Effect(e1,e2) --
zinc causes growth, e1 IS the cause here), so the parser reads the
argument order per record rather than assuming a fixed e2->e1 direction.

Train (140 records, real true/false labels inline) and test (80 records,
label hidden as "?" in ``test/relation-1-test.txt``) are SEPARATE files;
test's real gold labels live in a third file, ``key/relation-1-score.txt``
-- verified: matched 1:1 by record index against test's own text/entity-
direction with 0 mismatches across all 80 test records. Total 220
records (114 true / 106 false) -- note Hagen et al. 2026 (arXiv:2510.08224)
Table 2 cites this as "220 causal + 114 noncausal": 220 is actually the
TOTAL record count here, and 114 matches this project's own CAUSAL count,
not noncausal (106) -- looks like a mislabeling in that table (total
mistaken for the causal count, or causal/noncausal swapped), not a
mismatch in this conversion; flagged rather than silently reproduced.

causality-detection uses all 220 records (label = causal iff true).
causal-candidate-extraction/causality-identification use only the 114
causal (true) records for their entity/relation content -- matching this
project's established convention elsewhere (e.g. BioCause, FinCausal) of
only extracting spans that back an actual relation; the 106 false records
still appear in causality-identification (entities marked, empty
relations list) so the evaluation harness's own pair-flattening can
derive negative pairs from them, exactly as for every other dataset here.
"""

import io
import re
import tarfile
import urllib.request
from pathlib import Path

import pandas as pd

from causalatee.data.constants import ClassLabel, Relation, Task
from causalatee.data.utils import parse_entity_markers, verify_dataset

_TAR_URL = (
    "https://raw.githubusercontent.com/davidsbatista/Annotated-Semantic-Relationships-Datasets"
    "/master/datasets/SemEval2007-Task4.tar.gz"
)
_CACHE_DIR = Path(__file__).parent / ".cache"

_RECORD_RE = re.compile(r'^(\d+)\s+"(.*)"\s*$')
_LABEL_RE = re.compile(r'Cause-Effect\((e\d),(e\d)\)\s*=\s*"([^"]*)"')


def _fetch_files() -> dict[str, str]:
    """Download+extract once (cached); return {"train": ..., "test": ..., "key": ...} raw text."""
    _CACHE_DIR.mkdir(parents=True, exist_ok=True)
    out = {}
    for name in ["train", "test", "key"]:
        cache_path = _CACHE_DIR / f"relation-1-{name}.txt"
        if cache_path.exists():
            out[name] = cache_path.read_text(encoding="latin-1")
            continue
        with urllib.request.urlopen(_TAR_URL) as resp:
            outer = tarfile.open(fileobj=io.BytesIO(resp.read()))
            inner_name = {"train": "train.tar.gz", "test": "test.tar.gz", "key": "key.tar.gz"}[name]
            inner_bytes = outer.extractfile(f"SemEval2007-Task4/{inner_name}").read()
            inner = tarfile.open(fileobj=io.BytesIO(inner_bytes))
            fname = {"train": "train/relation-1-train.txt", "test": "test/relation-1-test.txt",
                      "key": "key/relation-1-score.txt"}[name]
            text = inner.extractfile(fname).read().decode("latin-1")
        cache_path.write_text(text, encoding="latin-1")
        out[name] = text
    return out


def _parse_records(text: str) -> dict[str, dict]:
    """idx -> {"text": marked sentence, "cause_ref": "e1"|"e2", "effect_ref": ..., "label": "true"/"false"/"?"}."""
    records = {}
    for block in re.split(r"\n\s*\n", text.strip()):
        lines = block.strip().splitlines()
        if not lines:
            continue
        m = _RECORD_RE.match(lines[0])
        if not m:
            continue
        idx, sent = m.groups()
        rel_line = lines[1] if len(lines) > 1 else ""
        lm = _LABEL_RE.search(rel_line)
        if not lm:
            continue
        cause_ref, effect_ref, label = lm.groups()
        records[idx] = {"text": sent, "cause_ref": cause_ref, "effect_ref": effect_ref, "label": label}
    return records


def _load_split(split: str) -> dict[str, dict]:
    """causalatee split name -> {idx: record} with resolved true/false labels."""
    files = _fetch_files()
    if split == "train":
        return _parse_records(files["train"])
    test_records = _parse_records(files["test"])
    key_records = _parse_records(files["key"])
    for idx, rec in test_records.items():
        rec["label"] = key_records[idx]["label"]
    return test_records


def convert_for_causality_detection(split: str) -> None:
    records = _load_split(split)
    rows = []
    for idx, rec in records.items():
        clean_text, _ = parse_entity_markers(rec["text"])
        label = ClassLabel.Causal if rec["label"] == "true" else ClassLabel.Uncausal
        rows.append({"index": f"semeval2007t4_{split}_{idx}", "text": clean_text, "label": label})
    df = pd.DataFrame(rows).set_index("index")
    for error in verify_dataset(df, Task.CausalityDetection):
        print(f"WARNING [SemEval2007T4 causality detection/{split}]: {error}")
    df.to_parquet(f"./causality-detection/{split}.parquet", engine="pyarrow")


def convert_for_causal_candidate_extraction(split: str) -> None:
    records = _load_split(split)
    out = []
    for idx, rec in records.items():
        if rec["label"] != "true":
            continue
        clean_text, segments = parse_entity_markers(rec["text"])
        entity = sorted(seg for segs in segments.values() for seg in segs)
        out.append({"index": f"semeval2007t4_{split}_{idx}", "text": clean_text, "entity": [list(s) for s in entity]})
    df = pd.DataFrame(out).set_index("index")
    for error in verify_dataset(df, Task.CausalCandidateExtraction):
        print(f"WARNING [SemEval2007T4 causal candidate extraction/{split}]: {error}")
    df.to_parquet(
        f"./causal-candidate-extraction/{split}.parquet", engine="pyarrow"
    )


def convert_for_causality_identification(split: str) -> None:
    records = _load_split(split)
    out = []
    for idx, rec in records.items():
        clean_text, segments = parse_entity_markers(rec["text"])
        relations = []
        if rec["label"] == "true":
            relations.append({
                "relationship": Relation.Procausal,
                "first": rec["cause_ref"],
                "second": rec["effect_ref"],
            })
        out.append({"index": f"semeval2007t4_{split}_{idx}", "text": rec["text"], "relations": relations})
    df = pd.DataFrame(out).set_index("index")
    for error in verify_dataset(df, Task.CausalityIdentification):
        print(f"WARNING [SemEval2007T4 causality identification/{split}]: {error}")
    df.to_parquet(
        f"./causality-identification/{split}.parquet", engine="pyarrow"
    )


if __name__ == "__main__":
    for split in ["train", "test"]:
        convert_for_causality_detection(split)
        convert_for_causal_candidate_extraction(split)
        convert_for_causality_identification(split)