Page 325 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
The use case is deployed in the Hubei Customer Service Center of China
Testbeds or Pilot Telecom Corporation Limited. The Hubei Telecom 10000-system currently 4.3 - 5G
Deployments
supports over 200 business scenarios and nearly 800 business interfaces.
Code repositories Not Available

2 Use Case Description

2�1 Description

This case involves the development of the "Diting" Intelligent Customer Service Agent, which
uses AI to empower China Telecom's 10000 customer service hotline. It provides users with
warm, expert-level customer service while reducing operational costs for the enterprise.

The current goal is to leverage "Diting" to quickly learn and adapt to new businesses,
accommodate the complexity of operations and the uncertainty of human-computer
interactions, and continuously reduce the volume of cases requiring human customer service
intervention.

Innovative Technological Approaches Being Implemented at the algorithm level, knowledge
base level, and business level are described below:

Algorithm Level: The "Diting" Intelligent Customer Service Agent adopts a system architecture
that leverages collaboration between large and small models to adapt to different scenarios. It
encompasses the entire lifecycle of AI models, including model training, fine-tuning, inference
acceleration, model evaluation, and data feedback loops. In terms of model training, it requires
obtaining dialog text between users and customer service representatives, which is then
manually labeled by an annotation team. To date, over 300,000 pieces of data have been
labeled. During application, the collaboration between large and small models enables rapid
recognition of user intent and automatic extraction of key information such as ID numbers,
phone numbers, and addresses provided by users. For intent recognition, the small model
uses the Tiny-BERT architecture and applies the standard BERT embedding techniques such
as Token Embedding, Segment Embeddings, and Position Embeddings [3]. The training data
consists of approximately 300,000 dialogues between users and customer service agents.
The large model used for intent recognition is the Telechat32B or Qwen8B, which shares the
same dataset as the small model. It is fine-tuned using full-parameter fine-tuning, trained with
the LLaMA Factory tool, and deployed using the VLLM framework to support inference. Both
models are hosted on servers and made available for use via API calls. Human customer service
agents can communicate with users based on recommended scripts provided by the "Diting"
agent, resolving complex customer issues.

Knowledge Base Level: Telecom services and the latest promotional activities are input into
a knowledge base, which is accessed via API calls for scenarios such as Q&A and product
recommendations. This knowledge base architecture is based on a graph-structured
text indexing framework. Using large language models and natural language processing
technologies, entities and relationships are identified from text data. Extracted entities are
used as graph nodes, and relationships between entities serve as connecting lines between

289