Maskrio/nya-ir-miracl-id-preprocessed

MIRACL-id with five -nya preprocessing strategies

This dataset is a derivative work of MIRACL (Zhang et al., 2023) restricted to the Indonesian (id) subset, preprocessed five different ways to study how Indonesian -nya clitic handling affects retrieval quality. Licensed under Apache-2.0, matching MIRACL.

Preprocessing strategies

keep

Baseline pass-through. Text is preserved exactly as MIRACL ships it.

naive_strip

Every word ending in -nya has the suffix stripped unconditionally. Includes false positives like punya, tanya, hanya, Kenya — this strategy exists as a deliberately bad baseline.

sastrawi_clitic

PySastrawi's RemoveInflectionalPossessivePronoun visitor, guarded by the bundled 29,931-root Indonesian dictionary so that punya/tanya/hanya/Kenya are recognised as roots and preserved.

sentinel

Like sastrawi_clitic, but the stripped suffix is replaced with the literal token <NYA>. Lets a retriever learn that something clitic-shaped was here without committing to a specific reading.

rule_resolved

Rule-based resolver: when a -nya form is preceded within ~2 sentences by a capitalised antecedent that is itself not a -nya form, the suffix is replaced by that antecedent. Falls through to keep when no antecedent is found. This is an in-house heuristic, NOT the IndoCoref system (Artari et al., 2021); see the paper's Limitations section.

Layout

README.md                       (this file)
root_dict.txt                   Sastrawi root dictionary (29,931 entries, Path A)
sensitivity_annotations.csv     Per-query H5 sensitivity (alpha=1*total + beta=2*anaphoric + gamma=3*entity_ref) + tertile
qrels/qrels_dev.txt             TREC 4-column qrels extracted from MIRACL dev positives
keep/                     queries_dev.jsonl (960 rows), corpus_train.jsonl (1446315 rows)
naive_strip/              queries_dev.jsonl (960 rows), corpus_train.jsonl (1446315 rows)
sastrawi_clitic/          queries_dev.jsonl (960 rows), corpus_train.jsonl (1446315 rows)
sentinel/                 queries_dev.jsonl (960 rows), corpus_train.jsonl (1446315 rows)
rule_resolved/            queries_dev.jsonl (960 rows), corpus_train.jsonl (1446315 rows)

All JSONL files have two fields: id (str) and contents (str). Loadable with datasets.load_dataset("maskrio/nya-ir-miracl-id-preprocessed", data_files=...) or with the Pyserini / sentence-transformers ingestion paths.

Citation

If you use this dataset, please cite both MIRACL and this work:

@article{zhang2023miracl,
  title={MIRACL: A Multilingual Retrieval Dataset Covering 18 Diverse Languages},
  author={Zhang, Xinyu and others},
  journal={Transactions of the ACL},
  year={2023}
}

Reproducibility

Everything here is reproducible from the code repo via scripts/run_preprocessing_sweep.py followed by scripts/upload_to_hf.py. The Sastrawi dictionary (root_dict.txt) is shipped alongside the data so that the dictionary-guarded strategies (sastrawi_clitic, rule_resolved) can be reproduced bit-for-bit even if PySastrawi ships a different kata-dasar.txt in a future release.