Knowledge-enhanced multi-task learning via prompt LLM for financial news recommendation

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Financial Technology

Background:

• Financial news recommender systems are vital for investors and analysts.
• Integrating external knowledge helps address cold-start and data scarcity issues.
• Current methods struggle with diverse knowledge assimilation and neglect news attributes like sentiment and topic.

Purpose of the Study:

• To develop a novel multi-task prompt large language model (LLM) for financial news recommendation.
• To effectively unify external knowledge and enhance news semantic understanding.
• To improve recommendation accuracy by considering news attributes and leveraging multi-task learning.

Main Methods:

• A dual knowledge enhancement strategy was developed to integrate structured and unstructured financial knowledge.
• Hierarchical knowledge prompt templates were designed for LLM to learn diverse task-specific knowledge.
• A multi-task prompt integration mechanism was implemented, jointly optimizing recommendation, sentiment analysis, topic classification, and popularity prediction.

Main Results:

• The proposed approach significantly improved performance on real financial news datasets.
• The method demonstrated particular effectiveness in few-shot learning scenarios.
• Joint optimization of related tasks leveraged inter-task dependencies for enhanced recommendations.

Conclusions:

• The novel multi-task prompt LLM approach effectively unifies external knowledge for financial news recommendation.
• Integrating news attributes and leveraging multi-task learning enhances recommendation effectiveness.
• The approach shows promise for improving financial information access for investors and analysts.