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

Financial news recommender systems are crucial for helping investors and financial analysts access essential market information. Recently, the incorporation of external knowledge into recommendation algorithms serves as supplementary data, aiming to mitigate the challenges associated with cold-start and data scarcity. However, the current research faces challenges in assimilating knowledge from diverse sources into the prompt-based framework. Additionally, multi-task learning methods for recommendation tasks and related tasks enhance recommendation effectiveness. However, the comprehensive consideration of the attributes of the news itself, such as sentiment, topic, and popularity, is neglected. To tackle these challenges, we present a novel multi-task prompt large language model (LLM) approach for financial news recommendation that effectively unifies external knowledge. In this study, we first develop a dual knowledge enhancement strategy to integrate structured and unstructured financial knowledge into the recommendation process to enrich the semantic understanding of news. Second, we design hierarchical knowledge prompt templates to enable LLM to learn diverse knowledge for specific tasks. Finally, we implement a multi-task prompt integration mechanism that jointly optimizes recommendation, sentiment analysis, topic classification, and popularity prediction tasks, leveraging inter-task dependencies. Experimental results show that our approach improves performance significantly on real financial news datasets, particularly in few-shot scenarios.