Google is committed to advancing racial equity for Black communities. See how.

Implementing Text Classifier

Text classifier uses machine learning techniques to help developers classify text.

Android 11 release text classifier

Android 11 introduces an updatable default implementation of the text classifier service in the ExtServices module. On devices running Android 11 or higher, the getTextClassifier() method returns this default implementation in the ExtServices module. Device manufacturers are recommended to use this implementation of TextClassifierService as it can be updated through Mainline OTA updates.

Android 11 also removes the local text classifier default implementation previously introduced in Android 8.1. As a result, getLocalTextClassifier() returns a NO_OP text classifier. Instead of the local implementation, you should use the getDefaultTextClassifierImplementation() method.

For device manufacturers who might want to use their own text classifying algorithms, they can implement a custom text classifier service by specifying config_defaultTextClassifierPackage in the config.xml file. If this configuration isn't specified, the default system implementation is used. Custom implementations can get an instance of the default implementation by calling TextClassifierService.getDefaultTextClassifierImplementation(Context). For more information, see Implementing a custom text classifier service.

Testing

To validate your implementation of the text classifier service, use the Compatibility Test Suite (CTS) tests in platform/cts/tests/tests/textclassifier/.

Android 10 release text classifier enhancements

Android 10 introduces two methods to the TextClassifier API: suggestConversationActions and detectLanguage. The suggestConversationActions method generates suggested replies and actions from a given conversation and the detectLanguage method detects the language of the text.

The model files for these methods are shown below and can be found in external/libtextclassifier/models/.

  • suggestionConversationActions: actions_suggestions.universal.model
  • detectLanguage: lang_id.model

To release a device with the latest model files in your factory image, do the following:

  1. Fetch the latest model files.

    external/libtextclassifier/models/update.sh
    
  2. Rename the downloaded files to replace the existing ones.
  3. Verify the setup.
    adb shell dumpsys textclassification
    

    This is a sample of the output of that command.

    TextClassifierImpl:
      Annotator model file(s):
        ModelFile { path=/etc/textclassifier/textclassifier.universal.model name=textclassifier.universal.model version=608 locales=und }
        ModelFile { path=/etc/textclassifier/textclassifier.en.model name=textclassifier.en.model version=608 locales=en }
      LangID model file(s):
        ModelFile { path=/etc/textclassifier/lang_id.model name=lang_id.model version=0 locales=und }
      Actions model file(s):
        ModelFile { path=/etc/textclassifier/actions_suggestions.universal.model name=actions_suggestions.universal.model version=0 locales=und }
    

Android 9 release text classifier enhancements

Android 9 extended the text classification framework introduced in Android 8.1 with the new text classifier service. The text classifier service is the recommended way for OEMs to provide text classification system support. The text classifier service can be part of any system APK and can be updated when necessary.

Android 9 includes a default text classifier service implementation ( TextClassifierImpl) that is used unless you replace it with a custom text classifier service implementation.

Implementing a custom text classifier service

The following sections describe how to implement a custom text classifier service that you develop.

Extend android.service.textclassifier.TextClassifierService

public final class TextClassifierServiceImpl
        extends TextClassifierService {

    // Returns TextClassifierImpl.
    private final TextClassifier tc = getLocalTextClassifier();

    @Override
    public void onSuggestSelection(
            @Nullable TextClassificationSessionId sessionId,
            @NonNull TextSelection.Request request,
            @NonNull CancellationSignal cancellationSignal,
            @NonNull Callback<TextSelection> callback) {
        CompletableFuture.supplyAsync(
                () -> tc.suggestSelection(request))
                .thenAccept(r -> callback.onSuccess(r));
    }

    @Override
    public void onClassifyText(
            @Nullable TextClassificationSessionId sessionId,
		@NonNull TextClassification.Request request,
		@NonNull CancellationSignal cancellationSignal,
		@NonNull Callback<TextClassification> callback) {
        ...
    }

    @Override
    public void onGenerateLinks(
            @Nullable TextClassificationSessionId sessionId,
            @NonNull TextLinks.Request request,
            @NonNull CancellationSignal cancellationSignal,
            @NonNull Callback<TextLinks> callback) {
        ...
    }
    ...
}

Define the service in the Android manifest

[AndroidManifest.xml]

<service android:name=".TextClassifierServiceImpl"
  android:permission="android.permission.BIND_TEXTCLASSIFIER_SERVICE">
    <intent-filter>
      <action android:name=
        "android.service.textclassifier.TextClassifierService"/>
    </intent-filter>
</service>

Note that the service must require the android.permission.BIND_TEXTCLASSIFIER_SERVICE permission and must specify the android.service.textclassifier.TextClassifierService intent action.

Set a system default text classifier service in the config overlay

[config.xml]

<string name="config_defaultTextClassifierPackage" translatable="false">com.example.textclassifierservice</string>

Build the text classifier service into the system image

Your custom text classifier service can be a standalone APK that is built into the system image or a part of another system APK. The system uses PackageManager.MATCH_SYSTEM_ONLY to resolve the service.

Testing

Run tests in android.view.textclassifier.cts.

Other text classification changes in Android 9

Refer to Inspecting installed language modules.

Android 9 model files are incompatible with Android 8.x model files.

Android 9 model files have the naming pattern: texclassifier.[language-code].model (for example, textclassifier.en.model) instead of textclassifier.smartselection.en.model in Android 8.x.

Obtaining the latest text classification model files

To obtain the most up-to-date models the following script can be run, which updates the TextClassifier models in the source tree:

external/libtextclassifier/native/models/update.sh

Android release 8.1 text classifier

Android 8.1 introduced the TextClassfier API to implement text classification.

TextClassificationManager tcm =
    context.getSystemService(TextClassificationManager.class);
TextClassifier classifier = tcm.getTextClassifier();
TextSelection selection = classifier.suggestSelection(...);
TextClassification classification = classifier.classifyText(...);

Developers can set a custom text classifier:

tcm.setTextClassifier(customTextClassifier);

But if an app developer sets the text classifier to null, a system default text classifier is returned for getTextClassifier().

See android.view.textclassifier.TextClassifierImpl.

TextView and WebView use TextClassifier for smart selection and smart text share features.

Text classifier usage

Figure 1. TextClassifier usage

TextClassifier neural-net models

The Android Open Source Project (AOSP) features a number of neural network models for classifying text. Each model file is trained for a single language. You can install any combination of models. The models are defined in:

external/libtextclassifier/Android.mk

Preinstalling language models on devices

You can specify a bundle of language models and install them on a device:

# -----------------------
# Smart Selection bundles
# -----------------------

include $(CLEAR_VARS)
LOCAL_MODULE           := textclassifier.smartselection.bundle1
LOCAL_REQUIRED_MODULES := textclassifier.smartselection.en.model
LOCAL_REQUIRED_MODULES += textclassifier.smartselection.es.model
LOCAL_REQUIRED_MODULES += textclassifier.smartselection.de.model
LOCAL_REQUIRED_MODULES += textclassifier.smartselection.fr.model
include $(BUILD_STATIC_LIBRARY)

For example, in device/google/marlin/device-common.mk.

# TextClassifier smart selection model files
PRODUCT_PACKAGES += \
    textclassifier.smartselection.bundle1

Inspecting installed language modules

Use ADB to list the files in the directory:

$ adb shell ls -l /etc/textclassifier
-rw-r--r-- 1 root root ... textclassifier.smartselection.de.model
-rw-r--r-- 1 root root ... textclassifier.smartselection.en.model
-rw-r--r-- 1 root root ... textclassifier.smartselection.es.model
-rw-r--r-- 1 root root ... textclassifier.smartselection.fr.model

Model updates

Models can be updated either by having a new model included as part of a system image update, or dynamically by having a system component that triggers an update through the system API ACTION_UPDATE_SMART_SELECTION intent. By broadcasting this system API intent, the framework is able to update the language model of the currently set language. The models themselves contain the supported language and a version number so the latest appropriate model is used.

So you don't need to preload models for all languages because they can be added later. If no model file is found for the specified language, text classification returns no-op values.

Compatibility Test Suite tests

The associated Android Compatibility Test Suite (CTS) tests can be found in:

cts/tests/tests/view/src/android/view/textclassifier/cts/TextClassificationManagerTest.java

cts/tests/tests/widget/src/android/widget/cts/TextViewTest.java

  • testSmartSelection
  • testSmartSelection_dragSelection
  • testSmartSelection_resetSelection