MultiIdiom: A Multilingual Idiom Detection Dataset with Joint Idiomaticity and Span Annotation

MultiIdiom is a seven-language idiom dataset (English, Spanish, Hindi, Telugu, Indonesian, French, German) for joint idiom detection and span localization — predicting whether an idiom in a sentence is used idiomatically or literally, and marking its exact character-level span. 14,505 idioms / 34,184 senses / 112,317 examples, constructed via a six-stage LLM-assisted pipeline from Wiktionary-seeded idiom entries. English/Spanish/Hindi/Telugu are split 80/10/10 at the idiom-type level; Indonesian/French/German are held out entirely as zero-shot test languages. English and Telugu were evaluated and found to have an idiomaticity label agreement of 203/221 = 91.86% for English and 90/102 = 88.24% for Telugu. Data Splits Split EN ES HI TE ID FR DE Total Train 2,030 2,034 508 506 — — — 5,078 Dev 252 254 62 61 — — — 629 Test 251 254 62 62 325 800 800 2,554 Total 2,533 2,542 632 629 325 800 800 8,261 Indonesian, French, and German are zero-shot evaluation languages (test-only, no train/dev). French/German test counts (800 each) match the exact evaluation set used in the companion IdiomBERT paper. corpus_<lang>.jsonl files contain the full per-language pool (all senses/examples, no split) — use these instead of the official test split if you need more data than the paper evaluated on. Full Corpus (corpus_<lang>.jsonl — all senses/examples, no split) Lang Idioms Senses Examples Idiomatic Literal English 9,638 23,140 76,082 57,373 18,709 Spanish 3,185 7,518 24,767 18,480 6,287 Hindi 183 409 1,420 660 760 Telugu 103 210 734 318 416 Indonesian 46 95 328 156 172 French 332 703 2,038 1,218 820 German 1,018 2,109 6,948 3,656 3,292 Total 14,505 34,184 112,317 81,861 30,456