PLOTS: Paired Plot Outlines for Text-level Narrative Similarity and Retrieval

PLOTS pairs the plot summary of a film in English Wikipedia with the plot summary of the same film written independently by another language edition. The two texts tell the same story but are not translations: each edition chooses its own level of detail, focus and wording, so a model can only match them by what happens in the story. No machine translation and no manual annotation are used to build the pairs.

• 222,542 English–X pairs linking the English plot of 48,883 films to their plots in 89 other languages (90 languages in total, 37 with at least 1,000 pairs)
• Film-level splits: no film contributes to both training and evaluation, in any language
• Decontamination against SemEval-2026 Task 4 and TREC Tip-of-the-Tongue 2023
• Entity-masked variants (type labels and document-local identifiers) for en, fr, es, pt, de, it
• PLOTS-IR, a cross-lingual known-item retrieval benchmark over 32,655 English plots

Quick start

from datasets import load_dataset

pairs = load_dataset("Jgmorenof/PLOTS", "pairs")            # train / validation / test
coll = load_dataset("Jgmorenof/PLOTS", "ir_collection", split="masked")
queries = load_dataset("Jgmorenof/PLOTS", "ir_queries", split="test_masked")
# relevance: each query's only relevant document is the collection entry with doc_id == query qid

Configurations

Config	Content	Rows
pairs (default)	English–X pairs, splits train / validation / test	197,694 / 5,306 / 12,717
plots	every plot section (English and other languages) with its film-level split	299,099
masked	type-masked (text_type) and identifier-masked (text_idx) plots, 6 languages	166,935
ir_collection	PLOTS-IR collection: full, masked, masked_idx	32,655 each
ir_queries	PLOTS-IR query sets (test_*, dev_*)	see below
pt_controlled	4,465 training films: English plot, native Portuguese plot, Portuguese plot machine-translated into English	4,465

Splits

Split	Films	Pairs
train	43,902	197,694
dev	1,000	5,306
test	2,500	12,717
contam (held out: in an evaluation set)	1,521 (1,481 with a pair)	never used for training
en_only (no plot in another language)	27,634	distractors only

Dev and test films are sampled among films with a plot in at least one of eight evaluation languages (fr, es, pt, de, it, ru, ja, zh).

PLOTS-IR

The collection holds the English plots of all films never used for training (test, dev, held-out and English-only films). A query is the plot of a test film in another language; its only relevant document is the English plot of the same film (doc_id == qid).

Query set	Test queries	Content
full	6,965	whole plot (fr, es, pt, de, it, ru, ja, zh)
masked	5,359	type-masked plot, against the type-masked collection
masked_idx	5,359	identifier-masked plot, against the identifier-masked collection
first / last	4,517	first or last sentence only
last_masked / last_masked_idx	4,517	last sentence, masked

Fields

• plots: qid (Wikidata id of the film), lang, split, title, url, page_id (source Wikipedia page), heading (plot section heading), year, wikidata_description, labse_score (similarity used by the quality filter), text
• pairs: qid, lang, en_title, en_url, en_page_id, en_text, x_title, x_url, x_page_id, x_text, year
• masked: qid, lang, split, text_type, text_idx, n_entities
• ir_collection: doc_id (= film qid), title, text; ir_queries: query_id, qid, lang, text

Construction

Plot sections are extracted from the wikimedia/wikipedia dump (2023-11-01) and linked across editions through Wikidata sitelinks (2026-03-30). A page is a film when its English Wikidata description has film as head noun. The plot heading of each language is learned from the data. A section is kept if it has at least 300 characters, is written in the language of its edition, has a length in line with the English section, and is more similar to the other plots of the same film than 99% of plots of unrelated films (LaBSE, calibrated per language). Entities are masked with spaCy NER (PERSON, LOCATION, ORG).

Decontamination

Films of SemEval-2026 Task 4 and the targets of the TREC ToT 2023 train and dev queries are matched by Wikidata id and excluded from training (split contam); a second pass by normalised title and release year found no other overlap. metadata/contamination.json lists these ids. It also lists 38 stories that occur only in the SemEval track-B test labels and fall in the train split; the splits are kept as in the paper for reproducibility, and metadata/ns_track_b_leak_free_mask.json marks the 355 of the 400 track-B test triples that involve none of them. TREC ToT 2023 test labels are not public, so its targets cannot be excluded.

Limitations and bias

PLOTS covers films only and every pair involves English. Pairs rely on automatic steps (heading-based section extraction, a LaBSE filter that may drop valid but divergent summaries, and spaCy's small NER models, which miss entities more often in non-Western texts). The data inherits Wikipedia's coverage and framing: recent films (39% from 2000–2019) and European productions and languages are over-represented, and summaries may carry the stereotypes of their editors. All texts are public encyclopaedic summaries of fictional works; the only personal data are names of public figures already in Wikipedia.

License and attribution

The texts come from Wikipedia and are distributed under CC BY-SA 4.0. Every record keeps the title, URL and page id of its source page for attribution.

Citation

@misc{moreno2026plots,
  title        = {{PLOTS} -- Same Story, Different Words: Paired Plot Outlines for Narrative Similarity and Retrieval},
  author       = {Moreno, Jose G. and Hamdi, Ahmed and Boros, Emanuela and Jatowt, Adam and Bordea, Georgeta and
                  Gonz{\'a}lez-Gallardo, Carlos-Emiliano and Doucet, Antoine},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Jgmorenof/PLOTS}}
}