Microsoft XiaoIce: Architecture and Techniques for Persona‑Driven Conversational AI

Introduction – XiaoIce is a leading cross‑platform AI system; the talk introduces its latest dialogue framework and how to incrementally build a persona‑driven conversational system for social and practical scenarios.

1. Basic Dialogue System Classification – Describes the evolution from simple human‑machine interaction to modern agents that retrieve information from massive, long‑tail knowledge bases, motivating the need for intelligent agents.

Conversation AI Directions – Three main directions: (1) Task Completion (e.g., customer‑service bots), (2) Information/Answer retrieval (knowledge‑base style), and (3) General Conversation (open‑domain chat), with emphasis on the challenges of long‑term context.

Long‑Term Conversation Experience – Discusses session‑oriented design, the importance of extracting latent intents from seemingly irrelevant turns, and using conversation turns (CPS) as a metric for engagement.

2. Dialogue System Architecture – Outlines the macro‑architecture similar to search/recommendation pipelines: Query Understanding (context, entity, intent, emotion signals), passive policy (template, retrieval, generation with aggregation ranking), proactive policy (topic initiation), and downstream user‑profile mining.

3. Query Semantic Analysis & Rewriting – Presents two rewriting strategies: keyword‑based combination and sequence‑modeling with hierarchical encoders; introduces an unsupervised context‑rewriting model (EMNLP 2019) using a 2‑Encoder‑1‑Decoder architecture and data‑generation pipeline (PMI extraction, language‑model insertion, copy‑net, reinforcement learning).

4. Few‑Shot Learning – Explains prototypical networks for intent classification, hierarchical attention prototypical networks, and how they enable rapid addition of new intents without full retraining.

5. Retrieval Model Structure – Three‑layer retrieval: L1 inverted index/ANN recall, L2 ranking with word, sentence, context features and BERT‑based re‑ranking, L3 post‑processing (filtering, profile entailment).

6. Generation Model – Describes open‑domain generation models that incorporate topic words via attention, with examples in beauty‑care chatbots.

7. Empathy Model – Introduces dialogue‑act‑based controllable generation to manage conversation rhythm and user engagement.

AI Beings Concept – Defines AI beings as persona‑equipped agents with profile, capabilities, opinions, and cross‑platform deployment; discusses challenges of tool‑vs‑entity perception.

Character Studio & Bot Characterization – Manages AI‑being assets (conversation style, domain knowledge, opinion, relation) and uses a dual‑bot simulator (character bot + policy bot) for rapid evaluation.

Case Studies – Highlights deployed IP chatbots (e.g., “Daji” for gaming, Japanese food recommendation) and the overall impact of persona‑driven dialogue systems.

Conclusion – Summarizes the industrial perspective on building XiaoIce, the remaining challenges, and calls for community participation.