Source-linked AI summary

Federated Learning for Emoji Prediction in a Mobile Keyboard

Swaroop Ramaswamy, Rajiv Mathews, Kanishka Rao, Françoise Beaufays

arXiv:1906.04329v1cs.CLcs.LG

TL;DR

The paper asks whether emoji prediction can be trained for mobile keyboards without centrally collecting users’ text. It combines a recurrent emoji predictor with language-model pretraining, triggering and diversification mechanisms, and federated on-device training. The federated model achieves better performance than a server-trained model, supporting the feasibility of production-quality natural-language models trained while data remains on users’ devices.

  • Problem

    Mobile keyboards need accessible emoji suggestions, but training such models must account for constrained devices and users’ data remaining on-device.

  • Method

    The paper uses a recurrent emoji prediction model with language-model pretraining, separate triggering and diversification mechanisms, and federated learning.

  • Results

    The federated model achieves better Accuracy@1 than a server-trained model in federated evaluation, although its AUC is lower.

  • Takeaways & Limitations

    Federated learning can train an emoji prediction model with sparse, poorly balanced data while keeping users’ data on their devices.

  • Takeaways & Limitations

    AUC is computed only on server logs from short snippets typed in selected apps, which are less representative than client-cache data.

Abstract

from arXiv · show

We show that a word-level recurrent neural network can predict emoji from text typed on a mobile keyboard. We demonstrate the usefulness of transfer learning for predicting emoji by pretraining the model using a language modeling task. We also propose mechanisms to trigger emoji and tune the diversity of candidates. The model is trained using a distributed on-device learning framework called federated learning. The federated model is shown to achieve better performance than a server-trained model. This work demonstrates the feasibility of using federated learning to train production-quality models for natural language understanding tasks while keeping users' data on their devices.

1 Introduction

The paper adds text-based emoji predictions to Gboard, while addressing mobile constraints that require low-latency inference.

  • Gboard offers emoji predictions based on text previously typed, alongside word predictions.The paper illustrates this behavior with the context “This party is lit.”
  • Mobile devices constrain memory and CPU, while keyboard responses are typically expected within 20 ms of an input event.
  • A unidirectional RNN supports cached model states at each timestep because it uses only backward-in-time dependencies.Caching is used to reduce inference latency.

2 Federated Learning

Federated learning keeps user data on devices and sends focused model updates to a server, which aggregates local training into a global model.

  • Federated Learning keeps users’ data on their devices rather than collecting it centrally, while transmitting minimal and focused model updates.The framework can also be combined with secure multi-party computation and differential privacy.
  • FederatedAveraging sends a global model to selected devices, where local SGD updates are computed on client batches.
  • The server averages client weights after each training round to produce the next global model parameters.

3 Method

The method combines a compact CIFG recurrent model with language-model pretraining, separate emoji triggering, and probability scaling to diversify candidates.

  • Model architecture: CIFG couples the input and forget gates through f = 1−i, reducing parameters per cell by 25% compared with an LSTM.
  • Model architecture: The model uses a two-layer CIFG with 256 units per layer and predicts probabilities over 100 emoji from a 10,000-word vocabulary.Inputs use 96-dimensional embeddings.
  • Pretraining: Language-model pretraining of all layers except the output projection leads to much faster convergence for the emoji model.The language model achieves Accuracy@1 of 13.7% on the same vocabulary.
  • Triggering: A separate triggering decision is treated as a product choice, allowing control over how often emoji predictions are offered.
  • Triggering: The model uses an additional <UNK> class and shows emoji predictions only when its probability falls below a threshold.Training truncates sentences without emoji labels to random lengths so the model learns not to predict emoji after mid-sentence tokens.
  • Diversification: Because frequent emoji dominate predictions across contexts, the method scales each emoji probability using its empirical training frequency and a live-traffic-tuned factor α.Setting α to 0 removes diversification.

4 Server-based Training

Server-based models are trained on anonymized, opt-in Gboard logs filtered to high-confidence English sentences.

  • Server-based training uses anonymized snippets shared by opted-in Gboard users from selected applications.Personally identifiable information is stripped from the logs before further filtering.
  • The training logs are filtered to sentences labeled English with high confidence by a language detection model.

5 Federated Training

Federated training uses local client caches under eligibility constraints, with one local epoch per round and convergence typically after 2000 rounds.

  • Participating devices must have at least 2 GB of RAM, be in the United States or Canada, and use English (US).Devices must also be idle, charging, and connected to un-metered networks.
  • Each client has approximately 400 sentences on average and trains for one epoch in each round.
  • The model typically converges after 2000 training rounds.Federated evaluation uses a separate task on a different subset of client devices rather than an explicit train–evaluation split.

6 Evaluation

Model quality is measured with Accuracy@1, while AUC evaluates the triggering mechanism using server-collected logs; all metrics are computed before diversification.

  • Accuracy@1 is the ratio of accurate top-1 emoji predictions to examples containing emoji.
  • AUC evaluates the quality of the triggering mechanism.
  • AUC is reported only on server-collected logs because numerical integration is not straightforward in the federated-learning setting.
  • All evaluation metrics are computed before diversification.

7 Federated Experiments

Federated experiments examine client batch size, devices per round, and server optimization before comparing the best federated model with a server-trained model.

  • Client caches change continuously, so results from these non-concurrent experiments cannot be compared across experiments.The study nevertheless compares training configurations and then evaluates the best federated model against a server-trained model.
  • Large client batch sizes substantially improve model quality because emoji-containing sentences are sparse in client caches.Larger batches produce more accurate gradient updates, especially with heavily imbalanced target classes.
  • Accuracy increases with the number of devices per round, but returns diminish beyond K = 500.
  • Momentum of 0.9 with Nesterov accelerated gradients improves convergence speed and model performance over SGD.
  • The production federated model uses B = 1000, K = 1000, and momentum, while assigning zero weight to 99% of <UNK> examples to balance losses.
  • The federated model achieves better Accuracy@1 but lower AUC than the server-trained model in federated evaluation.AUC is computed on server logs, whose short snippets from selected apps are considered less representative of users’ typed text than client caches.

8 Live experiment

Live traffic experiments evaluated emoji prediction models for English-US users in the USA and Canada, with both models improving engagement metrics and the federated model performing better across metrics.

  • Average inference latency is around 1 ms using a quantized TensorFlow Lite model.
  • The live-traffic experiment covered users in the USA and Canada typing in English (US).
  • Both federated and server-trained models significantly increase prediction CTR, total emoji shares, and daily active emoji users over a no-prediction baseline.Table 3 reports relative metric changes with 95% confidence intervals derived using user-bucket jackknife estimates.
  • The federated model performs better than the server-trained model on all reported live-traffic metrics.
  • The increase in CTR is quite large for both models because emoji are triggered rarely.

9 Conclusions

The paper trains an emoji prediction model with a CIFG-LSTM using federated learning, achieving better performance than a server-trained model despite sparse data and poorly balanced classes.

  • Federated learning trains a CIFG-LSTM emoji prediction model without centrally collecting users’ data.
  • The federated model achieves better performance than a server-trained model.
  • The approach works with sparse data and poorly balanced classes.
Loading 1906.04329v1…