OCC: A Smart Reply System for Efficient In-App Communications

Yue Weng, Huaixiu Zheng, Franziska Bell, Gökhan Tür

Abstract

Smart reply systems have been developed for various messaging platforms. In this paper, we introduce Uber's smart reply system: one-click-chat (OCC), which is a key enhanced feature on top of the Uber in-app chat system. It enables driver-partners to quickly respond to rider messages using smart replies. The smart replies are dynamically selected according to conversation content using machine learning algorithms. Our system consists of two major components: intent detection and reply retrieval, which are very different from standard smart reply systems where the task is to directly predict a reply. It is designed specifically for mobile applications with short and non-canonical messages. Reply retrieval utilizes pairings between intent and reply based on their popularity in chat messages as derived from historical data. For intent detection, a set of embedding and classification techniques are experimented with, and we choose to deploy a solution using unsupervised distributed embedding and nearest-neighbor classifier. It has the advantage of only requiring a small amount of labeled training data, simplicity in developing and deploying to production, and fast inference during serving and hence highly scalable. At the same time, it performs comparably with deep learning architectures such as word-level convolutional neural network. Overall, the system achieves a high accuracy of 76% on intent detection. Currently, the system is deployed in production for English-speaking countries and 71% of in-app communications between riders and driver-partners adopted the smart replies to speedup the communication process.

Introduction

Communication between riders and driver-partners is a critical component of the ride-sharing experience, particularly during the pickup phase. However, safety is a paramount concern; research indicates that texting while driving increases accident risk by over six times. To address this, Uber developed One-Click-Chat (OCC), a smart reply system designed to facilitate safe, efficient communication.

Unlike general-purpose smart reply systems (e.g., Gmail), OCC must navigate the unique challenges of mobile chat in a ride-sharing context. Messages are typically very short (averaging 4–5 words) and frequently contain non-canonical text, including typos, slang, and abbreviations. Furthermore, the intents are highly domain-specific, requiring a system that is both accurate and computationally lean for real-time mobile use.

System Architecture

The OCC system utilizes a two-stage pipeline rather than an end-to-end generative model. This approach provides greater control over the suggested content and ensures the replies are appropriate for the professional context of the app.

High-level architecture of the OCC system showing the two-stage process of intent detection and reply retrieval.

1. Intent Detection

The first stage involves mapping an incoming message to a specific intent. Uber utilizes unsupervised Doc2vec embeddings trained on millions of anonymized historical chat logs. This allows the system to map messages into a dense vector space where semantically similar messages cluster together, regardless of orthographic variations.

To classify these vectors, the system employs a Nearest Neighbor Classifier (NNC). By calculating centroid vectors for pre-defined intents (e.g., "Where are you?") using a small labeled dataset, the system can classify new messages by identifying the closest intent centroid via cosine distance. This non-parametric approach was chosen over complex deep learning models because it requires significantly less labeled data and offers better interpretability for production debugging.

2. Reply Retrieval

Once an intent is identified, the system retrieves the most relevant responses from a pre-defined mapping. This mapping is derived from historical data, identifying the most popular and helpful replies for each specific intent. A final layer of manual augmentation and reordering by a content team ensures that the top four suggestions presented to the driver are clear and actionable.

Results and Impact

The OCC system is deployed on Michelangelo, Uber’s machine learning platform, to ensure low-latency inference. In production, the system has demonstrated significant success:

  • Accuracy: The NNC-based intent detection achieves 76% accuracy. When considering the "Top-4" suggested replies, accuracy exceeds 90%.
  • User Adoption: Currently, 71% of all English-language in-app communications between riders and drivers utilize OCC smart replies, drastically reducing the need for manual typing while driving.
  • Robustness: The system successfully clusters semantically related terms (like "fee" and "refund") without explicit supervision, proving effective at handling the "long tail" of infrequent message types.
The OCC user interface showing how smart replies are presented to the driver-partner.

Conclusion and Future Work

OCC represents a significant step forward in balancing mobile communication efficiency with driver safety. While highly effective, the current system has limitations, such as a lack of conversational context and a reliance on static intent-to-reply mappings. Future iterations aim to incorporate external metadata (like GPS location or ETA) and transition toward dynamic ranking algorithms and active learning to further personalize and improve the smart reply experience.

Summary updated

Abstract

Smart reply systems have been developed for various messaging platforms. In this paper, we introduce Uber's smart reply system: one-click-chat (OCC), which is a key enhanced feature on top of the Uber in-app chat system. It enables driver-partners to quickly respond to rider messages using smart replies. The smart replies are dynamically selected according to conversation content using machine learning algorithms. Our system consists of two major components: intent detection and reply retrieval, which are very different from standard smart reply systems where the task is to directly predict a reply. It is designed specifically for mobile applications with short and non-canonical messages. Reply retrieval utilizes pairings between intent and reply based on their popularity in chat messages as derived from historical data. For intent detection, a set of embedding and classification techniques are experimented with, and we choose to deploy a solution using unsupervised distributed embedding and nearest-neighbor classifier. It has the advantage of only requiring a small amount of labeled training data, simplicity in developing and deploying to production, and fast inference during serving and hence highly scalable. At the same time, it performs comparably with deep learning architectures such as word-level convolutional neural network. Overall, the system achieves a high accuracy of 76% on intent detection. Currently, the system is deployed in production for English-speaking countries and 71% of in-app communications between riders and driver-partners adopted the smart replies to speedup the communication process.
BibTeX
@inproceedings{weng2019occ,
  author = {Yue Weng and Huaixiu Zheng and Franziska Bell and Gökhan Tür},
  title = {{OCC: A Smart Reply System for Efficient In-App Communications}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2019},
  url = {https://openreview.net/forum?id=-HKDG6bF3Nf},
}
Figure