Exploring AI Applications, Cathay Financial Holdings Demonstrates Open-Source Small Language Model for Customer Intent Recognition
Cathay Financial Holdings announced on June 4, 2025, a proof-of-concept study using an open-source Small Language Model (SLM) to accurately determine customer intent. Initial results show this approach can reduce reliance on complex prompt engineering or vector retrieval modules, simplifying system architecture. The fine-tuned SLM achieved performance close to mainstream closed-source Large Language Models (LLMs) in customer intent tasks. The study used fully synthetic data, avoiding real customer information, to enhance future financial service operational efficiency.
📋 Article Processing Timeline
- 📰 Published: June 4, 2026 at 22:09
- 🔍 Collected: June 4, 2026 at 22:18 (9 min after Published)
- 🤖 AI Analyzed: June 6, 2026 at 15:28 (41h 9m after Collected)
Cathay Financial Holdings continues to explore the application of generative AI in financial scenarios. Following last year's sharing of empirical results on financial Large Language Models (LLMs), Cathay Financial Holdings today published a forward-looking proof-of-concept research result. It utilizes an open-source Small Language Model (SLM) to accurately determine customer intent, which is expected to improve the operational efficiency of future financial services upon deployment.
In recent years, Cathay Financial Holdings has been promoting innovative AI applications in financial scenarios, gradually building a scalable technical foundation covering internal process optimization, customer service experience enhancement, financial knowledge understanding, and model governance.
Cathay Financial Holdings issued a press release today stating that this proof-of-concept used an open-source model as the training target. Preliminary verification results indicate that under test conditions, it can reduce dependence on complex Prompt Engineering (designing instructions to guide the AI model to correctly understand problems and generate answers) or vector retrieval modules, helping to simplify system architecture.
Cathay Financial Holdings pointed out that the results also show that combining appropriate financial scenario data design with model fine-tuning could potentially improve model stability, inference efficiency, and application controllability. The fine-tuned SLM can achieve performance close to mainstream closed-source LLMs on customer intent judgment tasks, serving as a reference for enterprises when selecting AI language models for training in the future.
Regarding data governance and privacy protection, this proof-of-concept adopted 'fully synthetic data' to avoid using real customer data for training. By optimizing for the Taiwanese context and using methods like keyword expansion, the model's understanding of Taiwan's local financial service context, proper nouns, and ambiguous queries was strengthened. Application scenarios include common needs such as mortgage balance inquiries, credit card payments, and branch service guidance, laying the technical foundation for future intelligent search and service routing.
In recent years, Cathay Financial Holdings has been promoting innovative AI applications in financial scenarios, gradually building a scalable technical foundation covering internal process optimization, customer service experience enhancement, financial knowledge understanding, and model governance.
Cathay Financial Holdings issued a press release today stating that this proof-of-concept used an open-source model as the training target. Preliminary verification results indicate that under test conditions, it can reduce dependence on complex Prompt Engineering (designing instructions to guide the AI model to correctly understand problems and generate answers) or vector retrieval modules, helping to simplify system architecture.
Cathay Financial Holdings pointed out that the results also show that combining appropriate financial scenario data design with model fine-tuning could potentially improve model stability, inference efficiency, and application controllability. The fine-tuned SLM can achieve performance close to mainstream closed-source LLMs on customer intent judgment tasks, serving as a reference for enterprises when selecting AI language models for training in the future.
Regarding data governance and privacy protection, this proof-of-concept adopted 'fully synthetic data' to avoid using real customer data for training. By optimizing for the Taiwanese context and using methods like keyword expansion, the model's understanding of Taiwan's local financial service context, proper nouns, and ambiguous queries was strengthened. Application scenarios include common needs such as mortgage balance inquiries, credit card payments, and branch service guidance, laying the technical foundation for future intelligent search and service routing.