基于大型语言模型的金融领域议论挖掘方法

打开文本图片集

中图分类号：TP181 文献标志码：A 文章编号：1673-2340（2025）03-0034-10

Argument mining method in financial domain based on large language models

DING Fei'，KANGXin²*

L.School of Artificial Inteligence and Computer Science，Nantong University，Nantong 226ol9，China;

2.Department of Science and Technology，Tokushima University， Tokushima 77O85O6，Japan）

Abstract：Inrecentyears，financialtextanalysis hasbeenshiftingfromcoarse-grainedsentimentpolarityclasification tofine-grained inference tasks focusing on logical relations between sentence pairs.Toaddress limitations inexisting approaches—suchasthe inabilityto capture semantic relations like "Support""Attack"and "Unrelated"the lackof reasoning interpretability，and severe class imbalance—this studyproposesa structure-enhanced Financial Prompt-based Argumentation Network （FinPromptNet）．Built upon the LLaMA3-8B-Instruct large language model，FinPromptNet integrates structurally explicit prompt templates，chain-of-thought （CoT）reasoning guidance，partial parameter fine-tuning，andajoint strategycombining weighted samplingand cost-sensitive optimization.Experiments conducted ontheNTCIR-l7FinArg-lfinancial sentence-pair classification dataset demonstrate that FinPromptNetoutperforms state-of-the-art baselines including FinBERT and T5 in terms of accuracy，micro- ⋅F1 ，and macro- ∇⋅F1∇ .Specifically，it achieves a macro- .F1 score of 67.1% ，outperforming T5 by 7.3 percentage points，and yields over lO percentage points of improvement in F1 for the underrepresented "Attack" class.These results highlight the effectiveness of structureawareprompt designand imbalance-aware learning inimprovingboth model performanceand interpretability for financial logical reasoning tasks.

KeyWords： financial text analysis; inter-sentence logical relation;large language model; prompt-based fine-tuning; chain-of-thought reasoning;imbalanced learning

近年来，随着人工智能技术的迅猛发展，尤其是自然语言处理在金融领域的深入应用，文本分析由静态的信息抽取逐步演化为面向动态推理建模的深层次方法。（剩余19510字）