Alibaba XiaoMi: Intelligent Service Architecture and Emotion Response Capabilities

The talk, presented by Alibaba’s XiaoMi emotion dialogue algorithm lead Song Shuangyong, outlines the development of Alibaba’s intelligent customer service platform, from early phone and text‑based support to the current AI‑driven XiaoMi system that routes complex queries to human consultants while providing smart assistance.

In the intelligent service mode, problems are divided into inquiry and transaction dispute categories, each served by dedicated consultants; XiaoMi handles routine queries and escalates unresolved issues to human agents, while human agents receive AI‑assisted suggestions such as similar historical answers.

The ecosystem has expanded to three service layers—platform (Alibaba internal platforms like Taobao and Tmall), merchant (store‑level bots), and enterprise (external businesses)—each supported by specialized XiaoMi variants (Alibaba XiaoMi, Store XiaoMi, Enterprise XiaoMi).

Emotion reply capabilities are discussed along two research directions: human‑like emotional expression and multimodal emotion delivery (text, voice, facial cues). Robots are classified into three types: (1) no emotion processing, (2) full emotion recognition (e.g., handling abusive language with polite deflection), and (3) limited‑emotion customer‑service bots that can express positive emotions but must avoid negative ones.

Customer emotion soothing is implemented via offline and online pipelines. The offline pipeline includes an emotion classification model (38 emotion classes, with a focused 7‑class model for higher accuracy), a topic classification model, and knowledge‑base construction for precise replies. The online pipeline matches user utterances to existing knowledge or, if none matches, provides broader emotional soothing.

The emotion classification model combines sentence‑level semantic features extracted by SWEM, n‑gram features via CNN, and emotion embeddings that weight words by their relevance to specific emotions, enabling fine‑grained emotion detection for short user inputs.

Service quality detection targets two problem types—negative attitude and poor attitude—using features such as sentence length, speaker role, and semantic content to evaluate both robot and human agent interactions, especially in multi‑turn dialogues.

Generative emotional dialogue aims to move beyond generic “safe responses” by incorporating emotion and topic information into generation models, producing context‑aware replies like “I’m so happy for you!” when a user expresses joy.

Future work includes building a session‑level satisfaction estimator, currently based on manual surveys, to automate satisfaction scoring and reduce labor and statistical volatility.

Contact information and speaker biography are provided, inviting interested researchers and practitioners in NLP, machine learning, and AI to reach out for collaboration or recruitment.