Xiaomi Intelligent Dialogue Development Platform

Artificial Intelligence
1

Xiaomi Intelligent Dialogue Development Platform

In a recent presentation at the “Artificial Intelligence Technology Salon (Hangzhou Station)” hosted by Alibaba algorithm expert Tangcheng Guang, the birth of Xiaomi’s intelligent dialogue development platform was discussed. The platform is designed to address the growing demand for customer service, particularly in the e-commerce industry.

The Challenge of Customer Service

Statistics from Taobao, a leading e-commerce platform, reveal that customer service teams receive an average of 800-900 inquiries per day during off-season periods, rising to 1500-2000 during peak seasons. This is a significant burden on customer service staff, who often require additional support during major promotions like Double 11 and Double 12. Even China Mobile, with its tens of thousands of customer service agents, struggles to manage the volume of inquiries.

The Solution: Intelligent Customer Service

To address this challenge, Alibaba established a special intelligent services division, which aims to drive data and technology to simplify customer service. The Xiaomi intelligent dialogue development platform was born to solve the problem of Alibaba’s own customer service and later expanded to serve Taobao Lynx and government institutions.

Platform Architecture

The platform follows a common dialogue robot architecture, which consists of the following components:

  1. ASM (Automatic Speech Recognition): Converts user voice input into text.
  2. NLU (Natural Language Understanding): Analyzes the text input to determine user intent.
  3. Dialogue Session Management Tracking Module: Tracks the conversation and identifies the user’s goal.
  4. Policy Dialogue Module: Generates a response based on the user’s input and the platform’s knowledge base.
  5. NLG (Natural Language Generation): Converts the response into natural language.
  6. TTS (Text-to-Speech): Converts the natural language response into speech.

Design Principles

The platform is designed to be flexible, customizable, and robust. The three basic design principles are:

  1. Cold Start Fast: Enables the platform to do more, while reducing the user’s effort.
  2. Flexible Customization: Allows developers to build their own business scenarios based on the platform’s abstract dialogue components.
  3. Robust Evolution: Enables the platform to adapt to changing user needs and improve its performance over time.

Technical Realization

The platform’s technical realization involves the following components:

  1. Entity: A dimension-based entity that provides built-in functions and integrates industry-related functions.
  2. Intent: A node that determines the user’s intent based on natural language semantics.
  3. Function Node: A node that performs a specific function based on the user’s input.
  4. Recovery Node: A node that recovers from errors and inconsistencies.
  5. Fill Slots Node: A node that abstracts out common nodes and fills in missing information.

Example: Weather Query

The weather query scenario is an example of how the platform’s components can be used to achieve a simple task. The process involves the following nodes:

  1. Common Inlet Node: Receives user input.
  2. Check Weather Trigger Node: Determines the user’s intent.
  3. Fill Slots Node: Fills in missing information, such as location and time.
  4. API Function Node: Calls an external API to retrieve weather information.
  5. Reply Node: Conveys the response to the user in the form of natural language.

Functional Programming

China Mobile’s query scenario is an example of how the platform’s functional programming capabilities can be used to achieve a complex task. The process involves the following components:

  1. Determine Current Account: Checks whether the user is currently in a specific account.
  2. Detect Number: Detects whether the user’s number belongs to China Mobile.
  3. Shutdown Broadcast Information: Determines whether the user is eligible for a shutdown broadcast.
  4. Logic Processing: Processes the user’s input and generates a response based on the platform’s knowledge base.

Robustness of Evolution

The platform’s robustness of evolution is achieved through the following components:

  1. Rule-Based Syntax: Enables developers to define rules and patterns to handle complex user queries.
  2. Model: Integrates from the rule to extend the model, which can be converted into a training corpus.
  3. Transfer Learning: Enables the platform to adapt to changing user needs and improve its performance over time.

Dialogue Status Tracking

The platform’s dialogue status tracking capabilities involve the following components:

  1. NBT Solution: Enables the platform to clarify user intent and confirm the user’s intention.
  2. Optimal Policy Dialogue: Proposes optimal policy dialogue options, such as clarification of timing and volume.

Conclusion

The Xiaomi intelligent dialogue development platform is a cutting-edge solution for customer service that addresses the growing demand for intelligent customer service. The platform’s design principles, technical realization, and components enable developers to build their own business scenarios and achieve a more personalized and efficient customer service experience.