apertium-keyboard

Language model

Currently supports a morpheme-aware character language model. Sample Guarani model (converted_model.tflite) lives in the assets directory (./app/src/main/assets), along with the set of symbols the LM is aware of (labels.txt), in order of their numerical index as used by the neural net.

Currently, the keyboard uses a simplified prediction logic. It presents continuations of the current input, up to the next predicted boundary (morpheme, word, or sentence). In the keyboard view, word continuations are presented with a "_" at the end, morpheme continuations with a "@", and sentences with a "<eos>".

This is achieved via a greedy unroll of the top 3 character predictions given the character history (i.e., running the language model with it's top choice given some starting point).

Building

Building on the command line:

$ gradle assembleDebug

The .apk will be found in ./app/build/outputs/apk/debug/app-debug.apk.

Make sure you have the following environment variables set:

ANDROID_HOME=/path/to/android-sdk/
NDK_HOME=/path/to/android-ndk-r19c
JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64

Building in android studio.

Note that when you load the keyboard from Android studio, you may need to activate it as an input method:

Settings -> Languages & Input -> Virtual Keyboard -> Manage Keyboard

TODO: Add additional logic for rewriting the current unit, rather than just completing it.

Files

./app/src/main/res/xml/qwerty.xml: This is the main keyboard layout file
./src/main/assets
./src/main/java
./src/main/res

Key classes and methods:

LanguageModel/LanguageModelMobileNet: These classes in the tflite directory store the code for loading a Tensorflow Lite model, and using it to generate keyboard predictions. See LanguageModel::generate() for the inference function.

SmartKeyboard: This kotlin class in the service directory contains all the relevant keyboard code. Here are the some important functions:

SmartKeyboard::updateCandidates() //Get a new set of predicted continuations based on the current input history.

SmartKeyboard::getPredictions() //Call the language model to get some number of prediction candidates.

SmartKeyboard::pickSuggestion() //What should the keyboard do when the user picks a predicted continuation?

SmartKeyboard::handleBackspace() //What should happen if the user made a mistake and backs up?

Acknowledgements

The original fork this code is based on was written by Christo Kirov.

apertium / apertium-keyboard Goto Github PK