Abstract
Microsoft India’s "Ruuh" is a deep-learning-powered social chatbot designed to move beyond simple task-oriented utility toward genuine social interaction. By combining retrieval-based response ranking, emotion detection, and human-like image commenting, Ruuh successfully engaged over 2 million real-world users in more than 150 million conversations. This paper explores the architecture and strategies employed to navigate the nuances of human interaction, including slang, emotional subtext, and offensive content.
1. Introduction
Most conversational agents are "utilitarian"—designed for specific tasks like booking flights or checking the weather. However, building a social agent requires a different set of skills to handle the messy reality of human interaction. A social bot must express empathy, share excitement, and navigate the "dark side" of the internet, such as trolling and abusive language. Ruuh was developed to address these challenges, holding extended, multi-turn conversations while understanding nuances like slang and code-mixed language (e.g., mixing Hindi and English).
2. Architecture and Methods
Ruuh’s architecture is a pipeline of several specialized deep learning components designed for robustness and social relevance.
2.1 Response Retrieval
Instead of purely generative models, which can often produce "safe" but boring or nonsensical text, Ruuh uses an Information Retrieval (IR) approach. It indexes millions of paired tweets and responses. At runtime, it uses a three-step process: a TF-IDF fetch for relevant candidates, feature extraction via a Convolutional Deep Structured Semantic Network (CDSSN), and a final ranking based on 3-turn conversation context.
2.2 Emotion and Offensive Content Handling
The system uses a novel deep learning approach that combines semantic representations with sentiment-based representations to classify inputs as happy, sad, or angry. To handle trolls, the team deployed a Neural Bi-directional LSTM model specifically tuned to catch "toxic" inputs, even when they contain intentional misspellings or ambiguous entity names.
2.3 Image Commenting
Rather than standard image captioning (e.g., "A white kitten"), Ruuh uses a model trained on millions of image-comment pairs from social media to generate "human-like" opinions. This shift from captioning to commenting is a key differentiator for social AI.
3. Key Insights and Results
Ruuh's deployment provided significant "in-the-wild" evidence of success, reaching 2 million users and facilitating 150 million conversations. Several key insights emerged from this large-scale interaction:
- Relevance is Not Enough: For a social agent, a response can be "relevant" but socially inappropriate. Agents must meet "social needs," such as sharing a user's enthusiasm or offering comfort.
- The "Dodge" Strategy: When faced with controversial topics or toxic users, the authors found it more effective to use "clever dodging techniques" rather than engaging directly.
- Sentiment + Semantics: Detecting emotion in text is difficult without facial expressions. Combining semantic meaning with sentiment-specific features significantly improved detection accuracy.
- Opinionated Vision: In a social context, users don't want a bot to tell them what is in a photo; they want the bot to react to it.
4. Limitations and Future Work
Despite its success, text-based emotion detection remains a major challenge due to inherent ambiguity (e.g., interpreting a message as sad vs. angry). Additionally, while Ruuh addresses code-mixing in the Indian context, it remains an ongoing area of research. Future iterations may look at hybrid models that combine the safety of retrieval with the flexibility of generative AI.
Summary updated