Page 124 - AI for Good-Innovate for Impact Final Report 2024

AI for Good-Innovate for Impact

Use Case – 27: Metro Intelligent Customer Service Center Case

Country: China

Organization: Guangzhou GRG Intelligent Technology Solution Co., Ltd.

Contact person: ShuShen, Cheng, chengshushen@ grgbanking .com

27�1� Use case summary table

Domain Rail Transit Domain

Problem being Excavate and enrich the corpus data in the rail transit industry, construct a
addressed Large Language Model of Rail Transit Domain and continuously optimize
it, study the integration of Large Language Model technology with existing
small models to avoid AI hallucination issues in the public service field, and
deploy the large model in a lightweight manner.

Key aspects of Processing of multi-source heterogeneous data, fine-tuning of large models,
the solution text classification and retrieval, integration of large models with industry
knowledge graphs.

Technology Large Language Models, Fine-tuning, Lightweight, Knowledge Graphs.
keywords

Data availabil- The data source is non-public and primarily consists of industry-specific
ity data, including both structured and semi-structured data, such as docu-
ments from metro operating entities regarding regulations, fare parameters,
and route stations, as well as records of conversations between passengers
and staff.

Metadata(type The core algorithms of this project are focused in the field of NLP (Natural
of data) Language Processing), hence the training data is predominantly textual,
although it also involves voice data. However, ultimately, the voice data is
transformed into text data through ASR (Automatic Speech Recognition)
algorithms for NLP model training and optimization.

Model training Instructions Tuning, LoRA Tuning.
and finetuning

Testbed and URL: https:// aimetro .grgbanking .com/ guangzhou -qa/ #/ chat
pilot deploy- The testbed is primarily aimed at passengers of the Guangzhou Metro,
ment with the scope of knowledge encompassing issues related to passengers'
travel, such as route navigation, inquiries about facilities within the station,
fare inquiries, and metro ticketing rules. At present, this environment only
supports the Chinese language. Please ask questions in Chinese.

108