Contextual Typeahead Sticker Suggestions on Hike Messenger

Mohamed Hanoosh, Abhishek Laddha, Debdoot MukherjeeIJCAI 2019 · DOI

TL;DR

This paper introduces a novel sticker recommendation system for Hike Messenger that suggests relevant stickers in real-time as users type. By decomposing the problem into message intent prediction and then mapping intents to stickers, they achieve low-latency, on-device suggestions, significantly improving sticker usage.

What Problem Does This Paper Address?

Messaging apps like WhatsApp and Hike heavily rely on visual elements like stickers to convey emotions and thoughts. However, discovering the perfect sticker to match a user's intent can be a cumbersome experience, especially with a vast library of tens of thousands of stickers. The challenge lies in providing timely and accurate sticker suggestions that feel like a natural extension of the conversation, without introducing noticeable delays or requiring heavy computational resources on mobile devices. This is further complicated by the sheer volume of orthographic variations in user input, making it difficult to map directly from typed text to specific sticker intents.

Methods

The authors tackle this by decomposing the sticker recommendation problem into two main steps, moving away from a single, end-to-end model.

  1. Message Intent Prediction: Instead of predicting the exact message a user might send, they first predict the message cluster or intent. This is achieved through a hybrid approach:
  2. Server-side Response Prediction: A neural network trained on historical chat data predicts message clusters based on the previous message in the conversation.
  3. Client-side Typeahead Prediction: A trie-based language model processes user input in real-time, suggesting frequent messages and their associated cluster IDs. This component is crucial for low-latency, on-device operation.
  4. The scores from these two components are combined, with the typeahead model's influence increasing as the user types more.
  5. Message Cluster to Sticker Mapping: Once a message cluster (intent) is predicted, the system maps this intent to relevant stickers. This mapping is initially done via manual tagging of stickers with chat phrases and is then refined using user feedback (clicks vs. views) via a Multi-Armed Bandit algorithm. This approach allows for efficient updates to sticker relevance and the inclusion of new stickers without retraining the entire system.

Datasets

The paper mentions using a "historical chat corpus" for training their models. They also refer to a dataset of messages and their corresponding responses mapped to message clusters. The exact size and composition of these datasets are not detailed, but they are noted to be specific to the Hike Messenger platform and its user base, which includes a significant amount of transliterated text from native languages.

Key Insights and Ideas

  • Decomposition is Key: Breaking down the complex problem into message intent prediction and then intent-to-sticker mapping significantly simplifies the task and enables efficient, on-device processing.
  • Intent over Exact Message: Focusing on message intents rather than exact message strings is crucial for handling the vast number of orthographic variations and reducing redundancy.
  • Hybrid Server-Client Approach: Leveraging both server-side intelligence for context and client-side processing for real-time typing feedback is an effective strategy for balancing accuracy and latency on mobile devices.
  • User Feedback Loop: Using user interaction data (clicks vs. views) to refine sticker recommendations is a practical way to keep the system relevant and adapt to evolving user preferences and new sticker additions.

Key Results

The paper highlights that their sticker recommendation system is used by millions of users daily on Hike Messenger. A/B tests showed an 8% improvement in the fraction of messaging users who use stickers. The system operates under 20 milliseconds with a low CPU and memory footprint on devices, making it suitable for a wide range of mobile hardware.

Limitations and Future Work

The authors acknowledge that the current system is trained separately for each geographical region and language. Future work includes:

  • Deploying quantized neural network models for even more efficient on-device prediction on high-end devices.
  • Expanding the conversational context considered, potentially including emotions detected in the chat.
  • Personalizing suggestions based on individual user preferences, such as art style and gender.

Summary updated

BibTeX
@inproceedings{hanoosh2019contextual,
  author = {Mohamed Hanoosh and Abhishek Laddha and Debdoot Mukherjee},
  title = {{Contextual Typeahead Sticker Suggestions on Hike Messenger}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/945},
}
Figure