About
g-sho is a Japanese–English dictionary inspired by jisho.org. It is a static site: the dictionary is split into thousands of small files at build time, and your browser downloads only the pieces it needs for each search. Nothing you search for is sent anywhere else.
Data sources
The data comes from the projects below, through the jmdict-simplified JSON builds. Their licenses require attribution and, for the EDRDG files, that derived data be shared under the same license.
| Data | Source | License |
|---|---|---|
| Words | JMdict, by Jim Breen and the Electronic Dictionary Research and Development Group | CC BY-SA 4.0 |
| Kanji | KANJIDIC2, EDRDG | CC BY-SA 4.0 |
| Radicals and kanji parts | RADKFILE and KRADFILE, EDRDG | CC BY-SA 4.0 |
| Stroke order | KanjiVG, by Ulrich Apel and contributors | CC BY-SA 3.0 |
| Word frequencies (for ranking, and spelling suggestions) | wordfreq, by Robyn Speer, and the priority tags in JMdict | CC BY-SA 4.0 |
| Example sentences | Tatoeba: the ones JMdict links to word senses, and more from Tatoeba's word index (jpn_indices) | CC BY 2.0 FR |
| Japanese definitions (国語), and English words in Japanese (英和) | 日本語版ウィクショナリー (Wiktionary contributors), extracted by wiktextract / kaikki.org | CC BY-SA 4.0 and GFDL |
| English definitions | Open English WordNet (2025 edition), based on Princeton WordNet | CC BY 4.0 |
| Pronunciation recordings (English) | People's recordings on Wikimedia Commons, as listed by the Japanese Wiktionary; each file's author and license are on its Commons page | Mostly CC BY-SA |
| Spelling suggestions | Learned from Wikipedia's list of common misspellings (Wikipedia contributors) | CC BY-SA 4.0 |
| Handwriting recognition model | LT8/japanese-handwriting-onnx, trained on the ETL Character Database (Electrotechnical Laboratory, Japanese Technical Committee for Optical Character Recognition, ETL Character Database, 1973–1984) | ETL Character Database terms |
| Text recognition models (reading photos) | PaddleOCR PP-OCRv6 (tiny detection, small recognition), as converted to ONNX by RapidOCR | Apache 2.0 |
The processed data files this site serves (under data/) are derived from
these sources and are available under the same licenses.
How search works
- Japanese words are matched exactly and by prefix. Inflected verbs and adjectives (食べました, 高かった) are traced back to their dictionary form with a set of conjugation rules.
- Romaji is converted to kana and searched alongside English.
- A misspelled English word is corrected: beterment finds betterment, with a link to search for what you typed. The likeliest words are worked out on your device, from how common each word is and the mistakes people often make.
- English words can also be shown explained in Japanese (英和) and defined in English, in Settings. Forms lead to their word: went to go, walked to walk.
- Handwriting is recognized from the picture of what you draw, by a neural network running in your browser (WebGPU, or WebAssembly where WebGPU isn't available). Stroke count, order and direction don't matter.
- Text in photos (the camera button) is found and read by two more neural networks in your browser (WebGPU, or WebAssembly where WebGPU isn't available): one finds the lines of text, horizontal or vertical, the other reads each line. The camera's pictures and photos never leave your device.
- Sentences are split into words by finding the longest dictionary word at each position. This is simpler than the morphological analyzers some dictionaries use, so it occasionally splits text in the wrong place.