mainlp/word-level-translations

Word-Level Translations (WLT)

A word-level translation dataset covering 9 languages (English, German, French, Spanish, Italian, Portuguese, Indonesian, Japanese, Mandarin Chinese), designed for evaluating word/concept-level translation accuracy across many language pairs. Concepts are aligned across languages, with 100 query concepts (includes synonyms for evaluation) and 25 example concepts (intended for few-shot examples).

For more information on the construction process, please refer to https://arxiv.org/abs/2604.17633.

We welcome any corrections or suggestions for improvement.

Configs

This dataset ships three configs, all derived from the same underlying word list but serving different roles in an evaluation pipeline: query_set and word_alternatives are linked (the concepts in query_set are listed with synonym data in word_alternatives), while example_pool is a disjoint set of words reserved for building few-shot examples.

query_set

The evaluation queries: one row per concept, with its rank (frequency rank), part-of-speech info, and a single canonical translation per language. These are the concepts that have full synonym data in word_alternatives, joined on word_original.

column	description
rank	frequency rank of the word
pos_set	list of parts of speech the word can take, e.g. ["noun", "verb"]
primary_pos	the dominant part of speech
word_original	the English source word (concept key)
en, de, fr, es, it, pt, id, ja, zh	canonical translation in each language

example_pool

Same schema as query_set. This is a separate, disjoint pool of words used to build few-shot / in-context examples for evaluation prompts, kept apart from query_set so evaluation queries aren't contaminated by example selection. These words do not have synonym data — they don't appear in word_alternatives.

word_alternatives

One row per (word_original, source_lang, target_lang). For every source word and every target language, gives the set of valid translation alternatives (synonyms).

column	description
word_original	the English concept key, used to align this row with the same concept in query_set, example_pool, and other rows of word_alternatives
source_lang	the source language of this row (de, en, es, fr, it, id, ja, pt, or zh)
prompt	the actual source-language surface form used to query for translations (e.g. "Materialien" for the German concept "materials"). Equal to word_original only when source_lang == "en".
target_lang	the target language for this row
translations	list of valid translations (synonyms) from source_lang/prompt into target_lang

There is no row for source_lang == target_lang.

from datasets import load_dataset

# each config has a single split named "data"
query_set = load_dataset("mainlp/word-level-translations", "query_set", split="data")
example_pool = load_dataset("mainlp/word-level-translations", "example_pool", split="data")
word_alternatives = load_dataset("mainlp/word-level-translations", "word_alternatives", split="data")

# filtering word_alternatives to a specific language pair
de_fr = word_alternatives.filter(lambda r: r["source_lang"] == "de" and r["target_lang"] == "fr")

If you use this dataset in your work, please cite (accepted to EMNLP 2026, updated bibtex pending)

@misc{körner2026copyfirsttranslatelater, title={Copy First, Translate Later: Interpreting Translation Dynamics in Multilingual Pretraining}, author={Felicia Körner and Maria Matveev and Florian Eichin and Gitta Kutyniok and Barbara Plank and Michael A. Hedderich}, year={2026}, eprint={2604.17633}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2604.17633}, }