| 64 | 0 | 138 |
| 下载次数 | 被引频次 | 阅读次数 |
金融文本分析逐渐从基于情感极性的粗粒度分类任务，转向关注句对之间逻辑关系的细粒度推理建模。针对传统方法无法捕捉“支持”“反驳”“无关”等句间语义关系、推理不可解释及类别不均衡严重等问题，提出一种结构增强型金融Prompt推理网络（financial prompt-based argumentation network,FinPromptNet）。首先，该方法基于LLaMA3-8B-Instruct大语言模型，融合任务显式化的结构化Prompt模板、链式推理提示语（chain-of-thought prompt）、部分参数微调机制与加权采样-代价敏感联合优化策略，从输入编码与参数优化2个层面提升模型对句间逻辑结构的理解能力。然后，在NTCIR-17 FinArg-1金融句对关系分类数据集上开展系统实验。实验结果显示：FinPromptNet在准确率（accuracy）、微F1(micro-F1)和宏F1(macro-F1)等多项指标上整体优于FinBERT、T5等主流模型，其中宏F1达到67.1%，较T5提升7.3个百分点，且在“反驳”类标签的F1得分上提升超过10个百分点。该研究验证了结构化Prompt与类别调节机制在金融逻辑推理任务中的有效性，为金融文档的智能分析与语义建模提供可解释、高性能的解决方案。
Abstract：In recent years, financial text analysis has been shifting from coarse-grained sentiment polarity classification to fine-grained inference tasks focusing on logical relations between sentence pairs. To address limitations in existing approaches—such as the inability to capture semantic relations like ″Support″ ″Attack″ and ″Unrelated″ the lack of reasoning interpretability, and severe class imbalance—this study proposes a 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, and a joint strategy combining weighted sampling and cost-sensitive optimization. Experiments conducted on the NTCIR-17 FinArg-1 financial sentence-pair classification dataset demonstrate that FinPromptNet outperforms 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 10 percentage points of improvement in F1 for the underrepresented ″Attack″ class. These results highlight the effectiveness of structureaware prompt design and imbalance-aware learning in improving both model performance and interpretability for financial logical reasoning tasks.
[1] RAFFEL C, SHAZEER N, ROBERTS A, et al. Exploring the limits of transfer learning with a unified text-to-text transformer[J]. Journal of Machine Learning Research ,2020, 21(1):5485-5551.
[2] CUI Y, JIA M L, LIN T Y, et al. Class-balanced loss based on effective number of samples[C]//Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), June 15-20, 2019, Long Beach, CA, USA. New York:IEEE Xplore, 2019:9260-9269.
[3] CHEN C C, LIN C Y, CHIU C J, et al. Overview of the NTCIR-17 FinArg-1 task:fine-grained argument understanding in financial analysis[C]//Proceedings of the NTCIR Conference on Evaluation of Information Access Technologies, December 12-15, 2023, Tokyo, Japan. Tokyo:NTCIR, 2023:16-20.
[4] DU K, ZHAO Y Z, MAO R, et al. Natural language processing in finance:a survey[J]. Information Fusion, 2025,115:102755.
[5] LI P B, YU H, LUO X F, et al. LGM-GNN:a local and global aware memory-based graph neural network for fraud detection[J]. IEEE Transactions on Big Data, 2023, 9(4):1116-1127.
[6] LEWIS M, LIU Y H, GOYAL N, et al. BART:denoising sequence-to-sequence pre-training for natural language generation , translation , and comprehension[C]//Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , July 5-10 , 2020 , Online.Stroudsburg:Association for Computational Linguistics,2020:7871-7880.
[7] LUUKKONEN R , KOMULAINEN V , LUOMA J , et al.FinGPT:large generative models for a small language[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, December 6-10, 2023,Singapore. Stroudsburg:Association for Computational Linguistics, 2023:2710-2726.
[8] ZHANG X Y, YANG Q. FinMoE:a MoE-based large Chinese financial language model[C]//Proceedings of the Joint Workshop of the 9th Financial Technology and Natural Language Processing(FinNLP), the 6th Financial Narrative Processing(FNP), and the 1st Workshop on Large Language Models for Finance and Legal(LLMFinLegal).Abu Dhabi, UAE:Association for Computational Linguistics, 2025:48-53.
[9] RODRIGUEZ INSERTE P, NAKHL魪M, QADER R, et al.Large language model adaptation for financial sentiment analysis[C]//Proceedings of the Sixth Workshop on Financial Technology and Natural Language Processing, Bali,Indonesia. Stroudsburg:Association for Computational Linguistics, 2023:1-10.
[10] LIU P F, YUAN W Z, FU J L, et al. Pre-train, prompt,and predict:a systematic survey of prompting methods in natural language processing[J]. ACM Computing Surveys,2023, 55(9):1-35.
[11] WEI J, WANG X Z, SCHUURMANS D, et al. Chain-ofthought prompting elicits reasoning in large language models[C]//Proceedings of the 36th International Conference on Neural Information Processing Systems, November 28-December 9, 2022, New Orleans, LA, USA. New York:Association for Computing Machinery, 2022:24824-24837.
[12] WANG X Z, WEI J, SCHUURMANS D, et al. Self-consistency improves chain of thought reasoning in language models[EB/OL].(2022-03-21)[2025-05-30]. http://arxiv.org/abs/2203.11171.
[13] SANH V, WEBSON A, RAFFEL C, et al. Multitask prompted training enables zero-shot task generalization[EB/OL].(2021-10-15)[2025-05-30]. http://arxiv.org/abs/2110.08207.
[14] LI X L, LIANG P. Prefix-tuning:optimizing continuous prompts for generation[C]//Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing(Volume 1:Long Papers), August 1-6, 2021, Online. Stroudsburg:Association for Computational Linguistics, 2021:4582-4597.
[15] LIU X, JI K X, FU Y C, et al. P-tuning:prompt tuning can be comparable to fine-tuning across scales and tasks[C]//Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics(Volume 2:Short Papers), May 22-27, 2022, Dublin, Ireland. Stroudsburg:Association for Computational Linguistics, 2022:61-68.
[16] SHEN Y L, TAN Z Q, WU S H, et al. PromptNER:prompt locating and typing for named entity recognition[C]//Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers),July 9-14, 2023, Toronto, Canada. Stroudsburg:Association for Computational Linguistics, 2023:12492-12507.
[17] DING F, KANG X, WANG L H, et al. Causal inference and prefix prompt engineering based on text generation models for financial argument analysis[J]. Electronics ,2024, 13(9):1746.
[18] HUANG H, ZHANG B W, LI Y Y, et al. Knowledge-enhanced prompt-tuning for stance detection[J]. ACM Transactions on Asian and Low-Resource Language Information Processing, 2023, 22(6):1-20.
[19] OUYANG L, WU J, JIANG X, et al. Training language models to follow instructions with human feedback[C]//Proceedings of the 36th International Conference on Neural Information Processing Systems, November 28-December9, 2022, New Orleans, LA, USA. New York:Association for Computing Machinery, 2022:27730-27744.
[20] GRATTAFIORI A, DUBEY A, JAUHRI A, et al. The LLaMA 3 herd of models[EB/OL].(2024-07-31)[2025-05-30]. https://arxiv.org/abs/2407.21783.
[21] HENDRYCKS D, BURNS C, BASART S, et al. Measuring massive multitask language understanding[EB/OL].(2020-09-07)[2025-05-30]. http://arxiv.org/abs/2009.03300.
[22] TOUVRON H, LAVRIL T, IZACARD G, et al. LLaMA:open and efficient foundation lan guage models[EB/OL].(2023-02-27)[2025-05-30]. https://arxiv.org/abs/2302.13971.
[23] LI J W, GALLEY M, BROCKETT C, et al. A diversitypromoting objective function for neural conversation models[C]//Proceedings of the 2016 Conference of the North American Chapter of the Associa tion for Computational Linguistics:Human Language Technologies, June 12-17,2016, San Diego, California. Stroudsburg:Association for Computational Linguistics, 2016:110-119.
[24] DETTMERS T, PAGNONI A, HOLTZMAN A, et al.QLoRA:efficient finetuning of quantized LLMs[C]//Proceedings of the 37th International Conference on Neural Information Processing Systems, December 10-16, 2023,New Orleans, USA. New York:Curran Associates, Inc,2023:10088-10115.
[25] BINI M, GIRRBACH L, AKATA Z. DeLoRA:decoupling angles and strength in low-rank adaptation[EB/OL].(2025-03-23)[2025-05-30]. http://arxiv.org/abs/2503.18225.
[26] LIU Z, HUANG D G, HUANG K Y, et al. FinBERT:a pre-trained financial language representation model for financial text mining[C]//Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence,July 11-17, 2020, Yokohama, Japan. Yokohama:International Joint Conferences on Artificial Intelligence Organization, 2020:4513-4519.
[27] TAO Z W, JIN Z, BAI X Y, et al. SEAG:structure-aware event causality generation[C]//Proceedings of the Findings of the Association for Computational Linguistics:ACL2023, July 9-14, 2023, Toronto, Canada. Stroudsburg:Association for Computational Linguistics, 2023:4631-4644.
[28] QIN L B, WEI F X, XIE T B, et al. GL-GIN:fast and accurate non-autoregressive model for joint multiple intent detection and slot filling[C]//Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing(Volume 1:Long Papers), August 1-6, 2021. Online. Stroudsburg:Association for Computational Linguistics, 2021:178-188.
[29] XIE Q Q, HUANG J M, LI D, et al. FinNLP-AgentScen-2024 shared task:financial chal lenges in large language models-FinLLMs[C]//Proceedings of the Joint Workshop of the 8th Financial Technology and Natural Language Processing and the 1st Agent AI for Scenario Planning, August 3, 2024, Jeju, South Korea. Stroudsburg:Association for Computational Linguistics, 2024:119-126.
[30] DU J C, XU R F, HE Y L, et al. Stance classification with target-specific neural attention networks[C]//Proceedings of the 26th International Joint Conference on Artificial Intelligence, August 19-25, 2017, Melbourne, Australia. New York:Association for Computing Machinery,2017:3988-3994.
基本信息:
DOI：10.12194/j.ntu.20250601001
中图分类号:F830;TP18;TP391.1
引用信息:
[1]丁飞,康鑫.基于大型语言模型的金融领域议论挖掘方法[J].南通大学学报(自然科学版),2025,24(03):34-43.DOI:10.12194/j.ntu.20250601001.
基金信息:
日本公益财团法人JKA项目（2024M-502）