Large language model-driven multi-topic and multi-level sentiment analysis framework for futures market news

To address the need for systematic and in-depth mining of complex sentiment signals from futures market news, in this paper a large language model-driven multi-topic and multi-level sentiment analysis framework is proposed． A “macro-topic-specific aspect/event” hierarchical strategy is applied to this framework to construct a topic system that covers multi-dimensional market elements, to achieve accurate discrimination of topic-level sentiments． Aspect-sentiment-opinion triplet extraction and event sentiment analysis techniques are incorporated to identify sentiment impact of key market elements and unexpected events． Low-rank adaptation is employed to achieve efficient domain adaptation of a localized large language model, to validate applicability of large language models in financial text analysis． The proposed framework is found to perform rather well across multiple sentiment analysis tasks, effectively distinguishing the sentiment tendencies of different topics and deeply mining complex sentiment signals within news texts． This work provides a systematic solution for in-depth sentiment analysis of futures market news and offers reliable technical support for the quantitative analysis of the impact of news sentiment on the futures market．