Nofing/EventStoryLine-0.9-standard-eci-workere

EventStoryLine v0.9 — standard ECI (direction-corrected)

The EventStoryLine Corpus (ESC) v0.9 for event causality identification, in the standard-eci format, built from the ESC and ECB+ sources. It supersedes Nofing/EventStoryLine-0.9-standard-eci, which is identical on every shared column except that 2,893 of its 5,602 causal pairs (51.6%) point the wrong way: it read every gold line a b FALLING_ACTION as "a causes b".

Direction

The gold is ESC's coreference-extended evaluation format (evaluation_format/full_corpus/v0.9/event_mentions_extended), whose builder writes each pair in token order and flips the label when it swaps the ends:

gold line	meaning
a b PRECONDITION	a causes b
a b FALLING_ACTION	b causes a (a is a consequence of b)

As in Caselli & Vossen (2017), Gao et al. (2019), HOTECI, CPATT, Dr.ECI and iLIF. relations.causes holds [cause, effect] and relations.causedby the mirror [effect, cause]; annots states the same.

Splits

split	documents	topics
train	233	the 20 topics used for 5-fold cross-validation
dev	25	37, 41 (the last two topics: the standard development set)

The gold has the counts every paper reports: 1,770 intra- and 3,855 inter-sentence lines over the 258 documents (23 of them null PLOT_LINKs, in unlabelled_pairs), 5,334 event mentions. On the 233 train documents, 1,592 intra-sentence causal pairs — the gold of Dr.ECI and SERE.

Columns

column	content
id, doc_idx, topic_id, split	document identity; doc_idx is stable, sorted by (topic, file name)
tokens, sentences	the document's tokens; [start, end) token range per sentence
mentions, spans	event mentions (ESC m_id, all ACTION_* / NEG_ACTION_*) and their token indices
relations	causes: [[cause, effect]], causedby: [[effect, cause]], indices into mentions
text, annots	tokens with <m_id surface> markup; the relations as <a ...> causes <b ...> items
plot_links	the PLOT_LINKs exactly as annotated, before propagation: source, target, rel_type (PRECONDITION / FALLING_ACTION / empty), the explicit-causality flags causes, caused_by, and signal
causal_origin	per relations.causes entry: annotated (a PLOT_LINK joins these two mentions) or coreference (a propagated copy)
causal_frame	per relations.causes entry: how the originating link was drawn — PRECONDITION (cause -> effect), FALLING_ACTION (effect -> cause), or both
coref_clusters	ECB+ event coreference within the document (clusters of >= 2 mentions)
coref_instance	per mention: the ECB+ cross-document event instance (empty if none)
mention_type	per mention: the ESC markable type (ACTION_OCCURRENCE, NEG_ACTION_REPORTING, ...)
unlabelled_pairs	the 23 gold lines whose PLOT_LINK has no relType (ESC's null), as [first, second]: not in relations (no direction), but counted as causal in the papers' 1,770 / 3,855

Candidate pairs: the literature's two conventions

	mentions	intra-sentence pairs	used by
all event mentions	5,334	10,347	ERGO, PPAT, HOTECI, DiffusECI, Dr.ECI, SERE
Gao et al. (2019): without aspectual, causative, perception, reporting mentions	4,695	7,805	Gao 2019, ICCL, DAPrompt, GLM4ECI, SemDI

Gao's set is mention_type not ending in ASPECTUAL, CAUSATIVE, PERCEPTION or REPORTING (46,251 cross-sentence pairs; the paper prints 46,521). Gold pairs touching a dropped mention stay in the gold there and cannot be found.

Statistics

causal pairs (each in both directions)	5,602
annotated / coreference copies	2,229 / 3,373
frame PRECONDITION / FALLING_ACTION / both	2,283 / 2,413 / 906
annotated PLOT_LINKs	2,265
coreference clusters / mentions with a cross-document instance	646 / 2,963

Caveats inherited from ESC's release, kept for comparability: 276 pairs are reached in both directions by propagation and ESC keeps one; the release differs slightly from re-running its own propagation script (63 pairs, 12 overwritten by null links).

Build

python dataprep/esc_standard_eci.py --esl-root <EventStoryLine clone> --out <dir> (workere).

Citation

Caselli, T. and Vossen, P. (2017). The Event StoryLine Corpus: A New Benchmark for Causal and Temporal Relation Extraction. Events and Stories in the News Workshop. — Cybulska, A. and Vossen, P. (2014). Using a sledgehammer to crack a nut? Lexical diversity and event coreference resolution. LREC. Both corpora: CC BY 3.0.