mk-sq-dictionary-muse

A Macedonian–Albanian word dictionary of 7,388 pairs. It is built by joining the MUSE English–Macedonian and English–Albanian bilingual dictionaries on their shared English words. It is meant for research on Macedonian ↔ Albanian machine translation and bilingual lexicon induction.

No one has checked these pairs by hand. Apart from 17 rows corrected by hand (see below), each pair comes from a shared English word only. A shared English spelling does not always mean a shared meaning, and MUSE has no word-sense information. The en column shows which English word(s) link each pair, so you can see what every row rests on.

Usage

from datasets import load_dataset

ds = load_dataset("EdonFetaji/mk-sq-dictionary-muse", split="train")
print(ds[0])  # {'mk': 'абас', 'sq': 'abaz', 'en': 'abbas'}

Data fields

field	type	description
mk	string	Macedonian word, in Cyrillic script
sq	string	Albanian word
en	string	the English pivot word(s) that link the pair; several are joined with "; "

One row per distinct (mk, sq) pair, ignoring case. There is a single train split.

Examples:

mk	sq	en
аберација	shmangie	aberration
физички	fizikisht	physically
почитува	respekton	respects
џуџе	xhuxh	dwarf; midget

How it was built

1. Sources. These two files were downloaded unchanged: en-mk.txt (50,749 lines) and en-sq.txt (52,111 lines). Both decode as strict UTF-8. No lines were malformed and no mojibake was found.

2. Normalisation. Strings were normalised to Unicode NFC and trimmed. Case, Albanian diacritics and Cyrillic were kept. Nothing was transliterated, accent-stripped, stemmed or lemmatised.

3. Copies of the English word removed. MUSE often "translates" an English word into itself (world → world). Lines whose translation equals the English word, ignoring case, were dropped before the join: 6,754 lines from en-mk and 24,472 from en-sq.

4. Join through English. The two dictionaries were matched on the English word, ignoring case. This gives 13,678 shared English words and 40,797 candidate MK–SQ combinations.

5. Filtering. The final dictionary keeps:

   • unambiguous pairs: the English word has exactly one Macedonian and one Albanian translation;
   • corroborated pairs: for English words with several translations, only MK–SQ pairs that are linked by two or more different English words. The other combinations are left out, because they mostly mix word senses (e.g. bank as a riverbank vs. a financial bank).

   Also left out:

   • entries with script problems, such as Macedonian with no Cyrillic, or Latin and Cyrillic mixed inside a word (oд);
   • markup debris (#за, статија}}, trailing commas);
   • Albanian spellings that lack the ë/ç of a variant listed for the same English word (e.g. mire when mirë is also listed).

6. Hand corrections. The dataset author reviewed the 27 entries blocked for script problems. 17 were corrected and kept, for example aтина → атина (Latin a replaced with Cyrillic а) and grad → сиво ("gray"). The rest were dropped.

No translations were generated or repaired automatically. No machine translation, LLM, Wiktionary data or confidence score was used.

Composition

part	rows
unambiguous pairs only	5,972
corroborated pairs only	1,374
both unambiguous and corroborated	25
hand-corrected entries	17
total	7,388

6,721 distinct Macedonian words and 6,751 distinct Albanian words.

Limitations

• Pivot noise. MUSE was made with an internal translation tool, not by lexicographers. Some pairs are loose or wrong in sense (e.g. цент → qind via cent; qind means "hundred").
• Proper names are included as transliterations (сузи → suzi via susie).
• Inflection. Many entries are inflected forms, and the two sides do not always agree in form (византија → bizantit via byzantium: nominative vs. genitive/definite).
• Coverage. Coverage is limited to English words present in both MUSE dictionaries. Many frequent words have several translations on each side, so they appear only when corroborated.
• No senses or part-of-speech information.

Treat it as a noisy silver-standard lexicon. Before using it as a gold reference, validate it.

License

CC BY-NC 4.0. The MUSE dictionaries are released under CC BY-NC 4.0, and this derived dataset uses the same licence. Non-commercial use only, with attribution to MUSE (Facebook Research) as below.

Citation

If you use this dataset, please cite the MUSE paper:

@inproceedings{conneau2018word,
  title     = {Word Translation Without Parallel Data},
  author    = {Conneau, Alexis and Lample, Guillaume and Ranzato, Marc'Aurelio and Denoyer, Ludovic and J{\'e}gou, Herv{\'e}},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2018},
  url       = {https://arxiv.org/abs/1710.04087}
}

Source repository: https://github.com/facebookresearch/MUSE