mandododo/greek-dictionary-futo

Greek dictionary for FUTO Keyboard (Greek-main_el)

A Greek word list with word frequencies for FUTO Keyboard: 908,211 words, including everyday slang, insults and swearing, weighted so that chat vocabulary ranks where it would in real typing, plus valid rare inflected forms (added at low frequency) so that they are not "corrected" into something else.

It gives FUTO its Greek word list and word frequencies: suggestions, autocorrect targets and swipe candidates all come from it. 608 vulgar words are flagged possibly_offensive, so FUTO's Block offensive words setting decides whether they are suggested.

Part of FUTO Keyboard Polyglot, a patched FUTO Keyboard with Greek and Albanian models for voice input, typing, dictionaries and dictation cleanup. The dictionary builder is in training/dictionary and the full model catalogue in models/.

Files

File	What it is
Greek-main_el.dict	For FUTO Keyboard. Compiled AOSP binary dictionary, locale el: 908,211 words, 608 flagged offensive.
Greek-main_el.combined	The source word list of the .dict, in AOSP .combined text format.
LICENSE, NOTICE	GPL-3.0 licence and the GPL notice.

The .combined format is plain text, one word per line with its frequency (AOSP scale, 30…250 here):

dictionary=main:el,locale=el,description=Ελληνικά (with slang),date=1790768875,version=1
 word=...,f=230
 word=...,f=187,possibly_offensive=true

Rebuild the .dict from it with AOSP's dicttool: java -jar dicttool_aosp.jar makedict -s Greek-main_el.combined -d main_el.dict (see training/dictionary for where to get the jar and its checksum).

Install in the app

Works in FUTO Keyboard Polyglot and in stock FUTO Keyboard.

1. Copy Greek-main_el.dict to the phone.
2. Open it from a file manager or share it to the keyboard, and import it as the dictionary for Greek. Alternatively, go to Settings → Languages & Models, open Greek, and import the dictionary there.
3. The keyboard language must match the dictionary's locale (el).
4. Optional: FUTO's Block offensive words setting (text prediction settings) hides the flagged words from suggestions. They stay valid words either way, so they are never autocorrected away. Mild insults and everyday slang are not flagged.

How it was built

build_dict.py from training/dictionary, config languages/el.json:

1. Counts words in each corpus. Every source has a boost: informal text is scarce, so subtitles, chat, offensive-comment datasets and slang lists are weighted up until chat vocabulary ranks where it would in real typing.
2. Picks one casing per word: a word is stored capitalised only if at least 90% of its uses are capitalised (names, places); otherwise lowercase, and FUTO capitalises it when needed.
3. Maps counts to AOSP frequencies, f = 255 + log(p) / log(1.15), clipped to 30…250. A kept word never gets f=0.
4. Flags vulgar words possibly_offensive=true from hand-made stem lists, keeping their normal frequency. Capitalised forms are treated as names and never flagged.
5. Merges Helium314's main_el2 word list at low frequency, so valid rare inflected forms are not "corrected".

Sources

Source	Used for	Licence
FineWeb-2 ell_Grek	Word frequencies	ODC-By 1.0; also subject to Common Crawl's Terms of Use
Wikipedia 20231101.el	Word frequencies	CC BY-SA 3.0 and GFDL
OpenSubtitles (OPUS), monolingual el, v2024, from subtitles on opensubtitles.org	Word frequencies (informal speech, boosted)	No licence of its own for the text: OPUS states it does not own it and honours takedown requests
GreekReddit (IMISLab)	Word frequencies (chat)	Apache-2.0 (dataset); the posts come from Reddit, whose terms restrict ML training
OffensEval 2020 Greek / OGTD (Pitenis, Zampieri, Ranasinghe), config gr	Word frequencies (swearing and insults)	CC-BY-4.0
Helium314/aosp-dictionaries main_el2, from dim-geo/greekdictionary	Rare Greek word forms added at low frequency	GPL-3.0, which is why this dictionary is GPL-3.0
Our own informal-sentence list (slang, insults, swearing)	Word frequencies	Our own work

Limitations

• The offensive-word flags need native review. They come from hand-made stem lists (608 words flagged). Some vulgar words are surely missing, and a few innocent look-alikes may be flagged.
• Frequencies reflect web, Wikipedia, subtitle and chat text, not a balanced corpus.

Checksums (SHA-256)

File	Bytes	SHA-256
Greek-main_el.dict	9,297,026	19f4f8ab74a3a19713d4d650b541cec11767ad4a4ae88baf75cb6514e3e9626e
Greek-main_el.combined	31,992,712	360525fc32330d74690a373560af39831800759810d2b2551068bf967e0c718b

The same checksums are in SHA256SUMS (sha256sum -c SHA256SUMS). Both files are byte-identical to the files in the GitHub release models-v1.

Licence and attribution

GPL-3.0. Greek-main_el.dict includes word forms from the Greek wordlist main_el2 of Helium314/aosp-dictionaries, which is derived from dim-geo/greekdictionary and licensed under the GNU General Public License v3.0. The dictionary is therefore distributed under GPL-3.0 (see LICENSE). Its corresponding source is the text word list Greek-main_el.combined in this repository, together with the build script training/dictionary/build_dict.py and its config languages/el.json. If you redistribute the .dict, keep the .combined, LICENSE and NOTICE with it.

Word frequencies were counted from the sources listed above; thank you to everyone involved. Required attributions: FineWeb-2 (Hugging Face), ODC-By 1.0, also subject to Common Crawl's Terms of Use. OpenSubtitles corpus from OPUS (University of Helsinki), from subtitles published on http://www.opensubtitles.org/ (P. Lison and J. Tiedemann, OpenSubtitles2016, LREC 2016). Wikipedia (Wikimedia Foundation and Wikipedia contributors), CC BY-SA. OffensEval 2020 Greek / OGTD (Pitenis, Zampieri, Ranasinghe), CC-BY-4.0. GreekReddit (IMISLab), Apache-2.0. The dictionary is compiled with AOSP's dicttool (Apache-2.0), in the prebuilt jar shipped by Helium314/aosp-dictionaries.

Takedown or licence questions: open an issue in the GitHub repository.