Understanding Chat Messages for Sticker Recommendation in Messaging Apps

Abhishek Laddha, Mohamed Hanoosh, Debdoot Mukherjee, Parth Patwa, Ankur NarangAAAI 2020 · DOI

Introduction

Stickers have become a cornerstone of modern digital communication, providing a level of expressivity that standard emojis cannot match. However, as sticker libraries expand into the hundreds of thousands, the cognitive load on users to find the appropriate sticker in real-time increases significantly. This paper explores a system developed for Hike Messenger that recommends stickers as users type, specifically addressing the challenges of linguistic diversity and hardware constraints in markets like India.

Overview of the sticker recommendation system showing the UI and high-level process flow.

The primary technical hurdles include:

  1. Linguistic Complexity: Users often transliterate native languages into Roman script, leading to extreme orthographic variation (e.g., over 300 ways to write "What are you doing?" in Hindi).
  2. Hardware Constraints: A significant portion of the user base operates low-end devices with limited memory and CPU, making complex on-device neural networks impractical.
  3. Latency Requirements: To ensure a seamless user experience, recommendations must be generated within tens of milliseconds.

Methodology

The proposed Sticker Recommendation (SR) system is decomposed into two main stages: Message Prediction and Sticker Substitution.

Unsupervised Message Clustering

To address the vast variation in slang and transliteration, the system focuses on predicting "intent clusters" rather than specific strings. The core of this approach is an encoder that combines a Character-level Convolutional Neural Network (CharCNN) with a Transformer or Gated Recurrent Unit (GRU). The CharCNN is vital for capturing sub-word information, allowing the model to recognize semantic similarities between variations like "gud" and "good" based on their character structure. Dense vector embeddings generated by this encoder are then grouped using HDBSCAN, mapping millions of variations into a manageable set of intent clusters without requiring manual labeling.

Hybrid Prediction Architecture

To balance accuracy with the performance limits of budget smartphones, a hybrid server-client model was implemented:

  • Server-side (Reply Model): Upon receiving a message, a server-side neural network predicts likely response clusters based on conversation context and sends these scores to the client.
  • Client-side (Typed Model): As the user types, a lightweight Trie-based search runs locally on the device, matching the current prefix against known phrases within the clusters.
  • Combiner: A final algorithm merges the context-based scores from the server with the real-time typing scores from the client to produce the final recommendation.

Evaluation and Results

The system was evaluated using the Hike Chat Corpus, consisting of 27 million anonymized message tuples. To assess embedding quality, a manually curated Phrase Similarity Dataset of over 5,000 phrase pairs was used.

ROC curves comparing different model architectures, showing the superior performance of the Trans+CharCNN model.

The results demonstrate that the Trans+CharCNN architecture outperforms standard GRU baselines, achieving an Area Under the Curve (AUC) of 0.863—a relative improvement of approximately 6%. Furthermore, the hybrid model proved more efficient than quantized 8-bit on-device neural networks, maintaining a memory footprint of less than 1MB.

In real-world deployment across eight Indian languages, the system led to an 8% relative increase in the number of users sending stickers and a significant boost in overall sticker volume.

Conclusion and Future Work

By prioritizing intent over identity and leveraging a hybrid architecture, the researchers successfully implemented a high-performance sticker recommendation system for resource-constrained environments. Future work will focus on automating the mapping of stickers to intent clusters and exploring more efficient on-device convolutions to further reduce server dependency. Additionally, incorporating individual user preferences could further personalize the recommendation experience.

Summary updated

BibTeX
@inproceedings{laddha2020understanding,
  author = {Abhishek Laddha and Mohamed Hanoosh and Debdoot Mukherjee and Parth Patwa and Ankur Narang},
  title = {{Understanding Chat Messages for Sticker Recommendation in Messaging Apps}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i08.7019},
}
Figure