yáng táo (杨桃) is an augmented reality app for learning Chinese characters. It was developed for the 2019 International Collegiate Competition for Brain-inspired Computing（ICCBC 2019). We were ranked in the top 16 and participated in the finals in Tsinghua University, Beijing, China. The idea is to combine many different cues to better remember Hanzi, e.g. etymology, radicals, pronounciation, mnemonics. For more information, please consult our Technical Document or watch our video!
 |
|---|
| Watch me on Youtube |
 |
 |
|---|---|
| The collection shows characters that can be scanned or already scanned (marked by a star) | The detail view shows information regarding pronunciation, meaning, decompositon, etymology and user-defined mnemonic. |
 |
 |
| In the AR view, characters can be scanned so that they appear in 3D. | After scanning a character, the user has the choices to either accept it as correct, wrong, or choose one of the 10 predictions instead. |
 |
|
| The augmented reality view allows users to scan and then explore characters in 3D. Clicking on the info sign leads the user to the details view. |
Chinese characters are an essential part of the Chinese language and culture. They are an integral part of becoming proficient in Mandarin. Learning even only the most common 2000-3000 characters is an arduous task that requires many hours of study. To support students, we implemented an augmented reality smart phone app for interacting with Chinese characters. When pointing the smart phone camera to a single character, we identify the character with machine learning and show its 3D representation above it. Its etymology, decomposition and mnemonic can be looked up. We use the knowledge about how the brain works and learns so that our users can better and more efficiently remember all the Hanzi. This is done by e.g. teaching character decompositions 六書, showing the decomposition in the 3D view or helping users coming up with mnemonics to connect Hanzi parts to a story.
The learning process implemented in Yangtao relies on the Dual Coding Theory of Reading and Writing which states that the brain best remembers if the item to remember, in this case Hanzi, is associated with as many different cues as possible. Therefore, we present the users with visual cues (the 3D Hanzi), audio cues (the pronunciation) and the logic behind the Hanzi (its decomposition, etymology and kind of formation).
A mnemonic for 休 could be e.g. a man leaning on a tree to rest, which is arguably easier to remember than learning its meaning by rote. The first app prototype is available for Android and uses Google ARCore, Sceneform and tflite. The character recognition itself uses a feed-forward neural network that is fed with HOG (Histogram of oriented gradients) features from filtered camera images.
Our prototype supports 250 selected characters. In the next iteration, we want to add online learning/lifelong learning capabilities so that the character recognition is improved by user feedback if it was wrong, adapting the app to the users’ environment. This is similar to how the brain continually learns and adjusts.
The app has three screens. When opening, it starts in the collection screen where all included Hanzi can be viewed. Clicking on a Hanzi opens its detail screen, showing information about meaning, mnemonic, decomposition and more. The mnemonic can be changed and saved with the buttin in the top right corner.
From the main screen, when clicking on the camera+ button in the bottom left, the AR view can be openend. At first, planes have to be detected, e.g. a table top. For this, the phone needs to be moved around left and right until plane indicator in form of dots appear. Then you are ready to scan Hanzi. The character should be aligned to the focus in the middle of the screen and scanned from the top. When pressing the scanning button, it then opens a menu. If the character is correct, it can be directly accepted. If it can be found in the dropdown list, it can be selected and then accepted. After accepting, the hanzi in 3D appears. One can move around and investigate its composition and shape as if it was a real object. It can be removed by clicking on the X or one can switch to the details view by clicking on the (i).
If it was totally wrong, then it needs to be scanned again. Typical problems we encountered are shadows, noisy background or only partially capturing the Hanzi. In the main view, the settings menue in the top right can be used to activate the debug view. It shows more information during scanning about how the app processes the image.
We use the 250 most common characters mostly based on Cai and Brysbaert 2010. They are
一三上下不与世业东两个中为主么义之也了事二于些产人什从他代以们件任会但位体何作
你使信做儿先入全公关其内再军几出分利别到制前力加务动化十原去又及反发受变口只可
各合同名后向员和四回因国在地场声处外多大天太头女她好如子学它安定实家对将小少尔
就工己已常平年并应度建开当很得心必性总情想意感成我或战所手才打把报接提政教数文
斯新方无日时明是更最月有本机条来果样次正此比民气水没法活海点然物特现理生用由电
的目相看真着知神种立第等系经结给美老者而能自行表被西要见解认论话说走起身过还这
进通道那部都里重量长门问间面题高
- We use SVG and dictionary from MakeMeAHanzi
- We use the SCUT-SPCCI Hanzi images created by the Group of Lianwen Jin of South China University of Technology as training data for the character classifier
- We use The dictionary Etymological Dictionary of Han/Chinese Characters by Lawrence J. Howell