Hanzi De Go

A browser-based simplified Chinese character writing sprint.

The game gives a pinyin and a concise Chinese definition from word.json. Round-parenthetical content is removed, clues are capped at 100 characters with ..., and every occurrence of the answer glyph or its listed traditional form is shown as ~ until resolution. Draw the matching character on the canvas, then press Check writing. The completed ink is normalized into an image and recognized by PP-OCRv5; stroke order and stroke direction are not passed to the model.

Undo, redo, clear, OCR candidates, seven frequency levels, campaign and Chengyu modes, lives, streaks, scoring, 60-second questions, post-round self-correction, and local best scores are included. The five most recent results and the game-over summary retain small in-memory thumbnails of the player's writing; they are never uploaded or persisted.

Run locally

Serve the public directory over HTTP so the model and WebAssembly runtime can load:

python -m http.server 8000 --directory public

Then open http://localhost:8000.

The 16 MB recognition model is stored locally in public/vendor/paddleocr. ONNX Runtime Web is loaded from jsDelivr.

Character data

characters.csv is the source. Regenerate public/character_data.js with:

powershell -NoProfile -ExecutionPolicy Bypass -File scripts/build_character_data.ps1

The build intersects the CSV with public/vendor/paddleocr/supported_characters.txt, derived from PaddleOCR's official PP-OCRv5 dictionary. Characters absent from the model are excluded automatically.

chengyu.txt is built separately into public/chengyu_data.js:

powershell -NoProfile -ExecutionPolicy Bypass -File scripts/build_chengyu_data.ps1

The Chengyu build keeps each playable character once, prefers a randomly selected four-character expression when available, reads the character's pinyin from word.json, and masks every occurrence of the answer with □.

Current generated source:

• 2,791 source rows in characters.csv
• 2,789 ranked characters with Chinese dictionary definitions
• 2 excluded because word.json has no matching single-character entry
• Level 1 (ranks 1–100): 100
• Level 2 (ranks 101–300): 200
• Level 3 (ranks 301–600): 300
• Level 4 (ranks 601–1,000): 400
• Level 5 (ranks 1,001–2,000): 1,000
• Level 6 (ranks 2,001+): 789
• Level 7: 11,059 unique characters in both word.json and the OCR vocabulary

Campaign modes

Every campaign contains 16 questions: five questions from each of its first three levels, followed by one final-boss question from the fourth level.

Only correct answers advance the campaign. A miss, timeout, or skip costs a life and replaces the prompt with a different character from the same level without moving the stage or route marker.

• Normal: levels 1, 2, 3, then a level 4 boss
• Hard: levels 2, 3, 4, then a level 5 boss
• Very hard: levels 3, 4, 5, then a level 6 boss
• Extreme: levels 4, 5, 6, then a level 7 dictionary boss
• Chengyu: 16 questions drawn from 4,361 unique OCR- and dictionary-supported characters

Constant practice modes are also available for each individual level from 1 through 7; all 16 questions stay in the selected level.

Handwriting recognition

The game uses the official PP-OCRv5 mobile recognition model from PaddlePaddle/PaddleOCR. Its dictionary contains 15,700 unique Han characters, covering all 11,059 Level 7 dictionary characters.

Only the recognition network is used. The document text-region detector is intentionally bypassed because it can reject sparse isolated characters such as 十. The final drawing is cropped, centered, resized to the model's 3 × 48 × 320 input, and evaluated locally in the browser with ONNX Runtime Web.

The CTC output is constrained to exactly one playable Han character across the complete image. The row beneath the canvas contains ranked alternative single-character hypotheses, not a multi-character transcription. This prevents the decoder itself from splitting a left and right component into separate output characters, although a component-shaped character can still be an incorrect alternative from the underlying text-recognition model.

Third-party notices and the Apache 2.0 license are in public/vendor/paddleocr.