基于思维链的问答系统增强推理方法

Chain-of-Thought Based Enhancement Method for Inference in Question-Answering Systems

摘要

针对生产测井数据多源异构、解释多解性强和传统方法依赖专家经验且效率较低等问题, 提出了一种基于思维链(CoT: Chain of Thought)的问答系统增强推理方法。该方法在大语言模型元AI(LLaMA2: Large Language Model Meta AI2)基础上, 引入参数高效微调低秩适配(LoRA: Low-Rank Adaptation)技术, 通过在自注意力和前馈网络层插入低秩矩阵, 实现了模型性能保持与计算资源优化的平衡。同时, 构建了发散思维链(DCoT: Divergent Chain of Thought)推理框架, 使模型能生成多条推理路径, 提升系统对复杂生产测井问题的多步解释能力。实验表明, 该方法在F₁、精确率和召回率上分别提升4.6%、7.0%和2.1%, 验证了其推理性能的优越性。研究证明, LoRA+DCoT方法在提升测井数据智能解释方面具有良好的实用性和可扩展性。

Abstract

To address the issues of multi-source heterogeneity in production logging data, strong multi-solution interpretability, the inefficiency and expert-dependency of traditional methods, a question-answering system enhanced reasoning method based on CoT(Chain of Thought) is proposed. This method is built on the LLaMA2 large language model and the parameter-efficient fine-tuning LoRA(Low-Rank Adaptation) technique is introduced, which inserts low-rank matrices into the self-attention and feed-forward network layers to achieve a balance between model performance maintenance and computational resource optimization. A DCoT(Divergent Chain of Thought) reasoning framework is constructed, enabling the model to generate multiple reasoning paths and enhancing the system's multi-step interpretation ability for complex production logging problems. Experimental results show that this method improves F ₁ score, precision, and recall by 4.6%, 7.0%, and 2.1%, respectively, compared to the baseline model, demonstrating strong reasoning performance and application potential. The research proves that the LoRA+DCoT method has good practicality and scalability in improving the intelligent interpretation effect of logging data.

关键词

大语言模型 / 思维链 / 生产测井 / 增强推理 / 低秩适配 / 发散思维链

Key words

large language model / chain of thought / production logging / enhanced reasoning / low-rank adaptation(LoRA) / divergent chain of thought(DCoT)

参考文献

[1] 钱亮宏, 王福德, 孙晓海. 基于预训练Transformer语言模型的源代码剽窃检测研究[J]. 吉林大学学报(信息科学版), 2024, 42(4): 747-753.

[2] QIAN L H, WANG F D, SUN X H. A Study on Source Code Plagiarism Detection Based on Pre-Trained Transformer Language Model[J]. Journal of Jilin University (Information Science Edition), 2024, 42(4): 747-753.

[3] 袁杰, 王军博, 陈歆, 等. 人工智能在超高性能混凝土中的应用研究进展[J]. 吉林大学学报(工学版), 2025, 55: 771-789.

[4] YUAN J, WANG J B, CHEN X, et al. Research Progress on the Application of Artificial Intelligence in Ultra-High Performance Concrete[J]. Journal of Jilin University (Engineering and Technology Edition), 2025, 55: 771-789.

[5] 张汝波, 常世淇, 张天一. 基于深度学习的图像信息隐藏方法综述[J/OL]. 吉林大学学报(工学版), 1-19 [ 2026-05-04]. https://link.cnki.net/doi/10.13229/j.cnki.jdxbgxb.20240381.

[6] ZHANG R B, CHANG S Q, ZHANG T Y. A Review of Image Information Hiding Methods Based on Deep Learning[J/OL]. Journal of Jilin University (Engineering and Technology Edition), 1-19 [2026-05-04]. https://link.cnki.net/doi/10.13229/j.cnki.jdxbgxb.20240381.

[7] KOJIMA T, GU S S, REID M, et al. Large Language Models Are Zero-Shot Reasoners[J]. Advances in Neural Information Processing Systems, 2022, 35: 22199-22213.

[8] WEI J, TAY Y, BOMMASANI R, et al. Emergent Abilities of Large Language Models[J/OL]. 2022 [2026-05-04]. https://arxiv.org/abs/2206.07682.

[9] WANG L, XU W, LAN Y, et al. Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models[J/OL]. 2023 [2026-05-04]. https://arxiv.org/abs/2305.04091.

[10] YUE X, NI Y, ZHANG K, et al. MMMU: A Massive Multi-Discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 9556-9567.

[11] 陈正非, 席皛, 李志勇, 等. 基于多源数据融合的电建知识图谱构建研究[J]. 吉林大学学报(信息科学版), 2024, 42(5): 921-929.

[12] CHEN Z F, XI S, LI Z Y, et al. Research on the Construction of Electric Construction Knowledge Graph Based on Multi-Source Data Fusion[J]. Journal of Jilin University (Information Science Edition), 2024, 42(5): 921-929.

[13] 彭建祥. 深度学习模式下大数据特征集成分类算法[J]. 吉林大学学报(信息科学版), 2025, 43(2): 231-237.

[14] PENG J X. A Feature-Integrated Classification Algorithm for Big Data in Deep Learning Mode[J]. Journal of Jilin University (Information Science Edition), 2025, 43(2): 231-237.

[15] 王祥, 谭国真, 彭衍飞, 等. 基于语言推理和认知记忆的自动驾驶决策模型[J]. 吉林大学学报(工学版), 2025, 55(12): 3918-3927.

[16] WANG X, TAN G Z, PENG Y F, et al. Autonomous Driving Decision-Making Model Based on Linguistic Reasoning and Cognitive Memory[J]. Journal of Jilin University (Engineering and Technology Edition), 2025, 55(12): 3918-3927.

[17] TOUVRON H, MARTIN L, STONE K, et al. Llama 2: Open Foundation and Fine-Tuned Chat Models[J/OL]. 2023 [2026-05-04]. https://arxiv.org/abs/2307.09288.

[18] 袁满, 张维罡, 李明轩. 基于认知图谱的智能问答系统推理模型研究[J]. 吉林大学学报(信息科学版), 2021, 39(5): 589-595.

[19] YUAN M, ZHANG W G, LI M X. Research on Reasoning Model of Intelligent Q&A System Based on Cognitive Mapping[J]. Journal of Jilin University (Information Science Edition), 2021, 39(5): 589-595.

[20] HU E J, SHEN Y, WALLIS P, et al. LoRA: Low-Rank Adaptation of Large Language Models[J]. International Conference on Learning Representations, 2022, 1(2): 3-10.

[21] PUERTO H, CHUBAKOV T, ZHU X, et al. Fine-Tuning with Divergent Chains of Thought Boosts Reasoning through Self-Correction in Language Models[J/OL]. 2024 [2026-05-04]. https://arxiv.org/abs/2407.03181.

[22] 赵东, 肖承文, 郭玉庆, 等. 测井软件程序库资源化与运行时部署集成框架[J]. 吉林大学学报(信息科学版), 2024, 42(5): 972-978.

[23] ZHAO D, XIAO C W, GUO Y Q, et al. An Integrated Framework for Logging Software Library Resourcing and Runtime Deployment[J]. Journal of Jilin University (Information Science Edition), 2024, 42(5): 972-978.

[24] GROENEVELD D, BELTAGY I, WALSH P, et al. OLMo: Accelerating the Science of Language Models[J/OL]. 2024 [2026-05-04]. https://arxiv.org/abs/2402.00838.

[25] BIGSCIENCE WORKSHOP, SCAO T L, FAN A, et al. BLOOM: A 176B-Parameter Open-Access Multilingual Language Model[J/OL]. 2022 [2026-05-04]. https://arxiv.org/abs/2211.05100.

[26] ZHANG H, ZENG P, GAO L, et al. MPT: Multi-Grained Prompt Tuning for Text-Video Retrieval[C]// Proceedings of the 32nd ACM International Conference on Multimedia. New York: ACM, 2024: 1206-1214.

[27] GAO X, XIE W, XIANG Y, et al. Falcon: Faster and Parallel Inference of Large Language Models through Enhanced Semi-Autoregressive Drafting and Custom-Designed Decoding Tree[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Washington DC: AAAI Press, 2025, 39: 23933-23941.

[28] TEAM G, MESNARD T, HARDIN C, et al. Gemma: Open Models Based on Gemini Research and Technology[J/OL]. 2024 [2026-05-04]. https://arxiv.org/abs/2403.08295.

[29] JIANG A Q, SABLAYROLLES A, ROUX A, et al. Mixtral of Experts[J/OL]. 2024 [2026-05-04]. https://arxiv.org/abs/2401.04088.

基金资助

国家自然科学基金资助项目(42172161)

国家自然科学基金资助项目(52474035)

黑龙江省自然科学基金联合基金重点资助项目(ZL2024D003)

中国石油创新基金资助项目(2024DQ02-0114)