Knowledge graphs have been widely used to transform fragmented information into structured and inferable relations, thereby supporting research on bird migration, conservation, and biodiversity. However, due to the high specialization and complexity of avian knowledge, as well as the diversity of entity relations, existing bird-domain knowledge graphs have remained limited in coverage and exhibited shallow relation extraction. To address these limitations, a bird-domain knowledge graph completion framework was proposed by integrating large language models with fine-tuning and a dynamic feedback mechanism. The framework was designed to enhance the understanding and reasoning of fine-grained avian knowledge for more efficient graph completion. Specifically, the Qwen model was fine-tuned using a domain-oriented instruction set to improve its specialized inference capability, and a dynamic feedback mechanism was incorporated to further strengthen reasoning and information extraction. Under relaxed evaluation settings, the proposed method achieved a Hits@1 of 63.0%, representing a 50.76% improvement over ChatGPT. With feedback enhancement, the performance increased to 95.3%, surpassing DeepSeek by 3.1%. The correct model outputs were iteratively integrated back into the graph, forming a closed-loop process of fine-tuning, completion, feedback, and expansion, which continuously strengthened self-learning and vertical-domain completion ability.
近年来,国内学者在鸟类知识图谱构建方面已取得初步成果。尹俊[5]通过本体建模与三层序列标注结构(Bidirectional Encoder Representations from Transformers-Bidirectional Long Short-Term Memory-Conditional Random Field,BERT-BiLSTM-CRF)模型联合抽取实体和关系构建了中国鸟类知识图谱联合抽取实体和关系,构建了面向领域知识的鸟类知识图谱(Bird Knowledge Graph for Domain Knowledge,BKG4DK),解决了数据整合难题;姚津[6]进一步拓展至亚洲鸟类,引入基于基于路径的平移嵌入模型(PTransE)的循环神经网络模型(PTransE-based Recurrent Neural Network,PT-RNN)推理方法,增强图谱的深度与关联性;李忠伟等[7]聚焦迁徙要素构建相关图谱,并通过RoBERTa提升识别精度,量化要素关联。但这些研究均未涉及区域特异性数据,且缺乏对鸟类生理特征的深度刻画。国际研究方面,现有成果多集中于通用生态知识图谱,如生物多样性网络[8];或是基于计算机视觉的鸟类识别[9],如单阶段目标检测模型(You Only Look Once, YOLO)系列模型在物种检测中的应用[10];世界鸟类数据库Avibase[11]虽提供结构化分类数据,但缺乏语义关联与推理能力;Wu等[12]通用知识图谱虽涵盖部分鸟类实体,却未针对生态特征建模。当前针对鸟类生态知识的结构化建模仍处于起步阶段,特别是在形态特征的细粒度抽取与部分关键属性整合方面,仍存在显著空缺。本研究构建的鸟类知识图谱旨在填补该领域知识组织与语义推理的空白,为生物多样性保护提供精细化数据支撑,推动知识图谱由“通用架构”向“深度垂直”方向转型。
1.2 知识图谱补全
知识图谱补全技术历经多阶段范式演进,其核心目标在于解决知识图谱固有的实体与关系缺失问题。早期研究依赖手工规则与统计推理方法[13],通过路径排序算法等人工定义规则进行关系推断,虽具备可解释性优势,但面临规则制定成本高、领域迁移性差的瓶颈。这一局限性推动了表示学习方法的兴起,以平移嵌入(Translation Embedding, TransE)等平移距离模型[14]通过向量空间几何约束建模关系语义,DistMult等语义匹配模型[15]则采用张量分解捕捉潜在关联。但这类嵌入模型在建模复杂语义交互与多跳推理任务时表现不足。随着深度学习技术的发展,图神经网络(Graph Neural Network, GNN)通过多层信息传播机制实现了结构化特征的有效聚合[16],而预训练语言模型(Pre-trained Language Model, PLM)的引入开创了文本与结构联合建模的新范式[17],基于双向编码器表征转换器的知识图谱补全模型(BERT-based Model for Knowledge Graph Completion, KG-BERT)通过将三元组编码为自然语言序列的创新实践[18]取得了一定进展,但受限于模型容量,在处理复杂推理任务时仍显不足。大语言模型LLM的快速发展,推动了知识图谱补全方法的范式转变[4]。凭借其强大的语言理解与生成能力,LLM可以更好地结合知识图谱中的文本与结构信息,从而进行更精准的补全。知识图谱大语言模型 (Knowledge Graph Large Language Model, KG-LLM)通过指令微调实现三元组建模[19],但现有方法普遍依赖单向推理流程,缺乏对预测结果的动态验证机制,导致错误传播风险累积。尽管Zhu等[20]验证了LLM在零样本补全中的潜力,但直接应用仍面临幻觉问题与领域适配性挑战。
2 方法
2.1 鸟类知识图谱构建
目前,网络上关于鸟类的数据虽体量庞大,但往往零散且冗余,难以支撑精细化的研究与应用需求。与此同时,鸟类生态知识具有高维度语义依赖与实体多样性特征,使得知识图谱构建面临知识碎片化、语义歧义与推理不确定性等多重挑战。因此,本研究从鸟类观测记录、学术论文、图书馆馆藏书籍、鸟类保护机构及生态调查报告等多种资料中,收集并整理了三江源、青海湖与祁连山等典型生态区域的鸟类相关数据,包括栖息地、羽毛特征、眼部特征等信息。数据通过光学字符识别(Optical Character Recognition, OCR)技术提取出有效信息,并结合数据清洗、标准化与去重等步骤,形成一个结构化的鸟类文本数据集,作为构建知识图谱的原始数据[21]。通用的知识图谱本体构建方法,有骨架法、本体方法体系(Toronto Virtual Enterprise, TOVE)法、本体工程方法论(Methontology)和基于叙词表的领域本体构建法等[22]。鉴于骨架法自顶向下的清晰层次、易于扩展的知识结构,和七步法规范的系统流程、成熟的建模方案,本文结合两者优势进行本体构建,以此高效支撑鸟类垂直领域的高质量图谱构建需求。根据鸟类学和生态学的特点,本文将本体中的实体分为核心类和细节类,核心类代表鸟类物种的基本分类和生物学特征,细节类描述鸟类的外观特征与具体的生态行为。Protégé支持本体的图形化编辑,便于构建本体结构并建模实体间复杂关系[23]。通过Protégé,可定义类、属性、关系和约束,并以Web本体语言(Web Ontology Language, OWL)格式输出本体文件。
动态反馈机制DFM通过递进式反馈架构增强生成模型的外部知识支持能力。其中每一层反馈表示特定粒度和置信度的知识输入层,用于动态增强模型的生成过程。在每一层中,系统基于主语-谓语对从知识图谱中抽取相关对象,构成候选对象集合,并引入一定比例的噪声对象,其采样自预设的噪声分布,以模拟真实语境中的信息不全与认知干扰,检验模型的鲁棒性与容错能力。模型当前输出与目标答案之间的语义相似度由度量函数计算,其中表示模型当前输出,表示目标答案。基于该相似度,采用启发式控制策略(Heuristic Control Strategy),动态地选择反馈层级,以调整知识密度。当相似度超过设定阈值时,触发早停策略(Early Stopping),节省推理资源。DFM包含三个递进层级,各层级对应不同的知识传递策略:
YinJ. Construction and Application of Knowledge Graph in Chinese Avian Domain[D]. Nanchang: Jiangxi University of Finance and Economics, 2022. DOI: 10.27175/d.cnki.gjxcu.2022.001130 .
YaoJ. Research on Ontology Construction and Reasoning for Asian Bird Knowledge Atlas[D]. Nanchang: Jiangxi University of Finance and Economics, 2023. DOI: 10.27175/d.cnki.gjxcu.2023.000640 .
LiZ W, LiM X, LiY, et al. Construction of Knowledge Graph Based on Bird Migration Associated Factor Data[J]. Comput Eng Des, 2024, 45(12): 3704-3711. DOI: 10.16208/j.issn1000-7024.2024.12.024 .
[12]
PageR D M. Ozymandias: A Biodiversity Knowledge Graph[J]. PeerJ, 2019, 7: e6739. DOI: 10.7717/peerj.6739 .
[13]
TanM, LeQ V. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks[C]//Proceedings of the 36th International Conference on Machine Learning (ICML 2019). Long Beach: PMLR, 97, 2019: 6105‑6114.
[14]
YangH, FengQ, ZouyiL Y, et al. Artificial Intelligence in Aquatic Biology: Identifying and Conserving Aquatic Species[J]. Water Ecol, 2025, 1(2): 100008. DOI: 10.1016/j.wateco.2025.100008 .
[15]
LepageD. Avibase: The World Bird Database[DB/OL]. (2003-06-24)[2025-3-20].
[16]
WuW T, LiH S, WangH X, et al. Probase: A Probabilistic Taxonomy for Text Understanding[C]//Proceedings of the 2012 ACM SIGMOD International Conference on Management of Data. Scottsdale: ACM, 2012: 481-492.. DOI: 10.1145/2213836.2213891 .
[17]
LaoN, CohenW W. Relational Retrieval Using a Combination Of Path-constrained Random Walks[J]. Mach Learn, 2010, 81(1): 53-67. DOI: 10.1007/s10994-010-5205-8 .
[18]
BordesA, UsunierN, Garcia-duranA, et al. Translating Embeddings for Modeling Multi-relational Data[C]//Advances in Neural Information Processing Systems 26 (NeurIPS 2013).Red Hook: Curran Associates Inc., 2013: 2787-2795.
[19]
ZhongJ, ZhaoH C, ZhaoQ C, et al. A Knowledge Graph-based Method for Drug-drug Interaction Prediction with Contrastive Learning[J]. IEEE/ACM Trans Comput Biol Bioinform, 2024, 21(6): 2485-2495. DOI: 10.1109/TCBB.2024.3477410 .
[20]
AroraS. A Survey on Graph Neural Networks for Knowledge Graph Completion[PP/OL]. (2020-07-24)[2025-3-21].
[21]
HuA, KuangL N, YangD H. LPBERT: A Protein-Protein Interaction Prediction Method Based on a Pre-trained Language Model[J]. applied Sci, 2025, 15(6): 3283. DOI: 10.3390/app15063283 .
HuangB, WuS A, WangW G, et al. KG-LLM-MCom: a Survey on Integration of Knowledge Graph and Large Language Model[J]. J Wuhan Univ Nat Sci Ed, 2024, 70(4): 397-412. DOI: 10.14188/j.1671-8836.2024.0040 .
[24]
ZhangL, WangW, JiangZ X, et al. Exploring Knowledge from Knowledge Graphs and Large Language Models for Link Prediction[C]//2024 IEEE International Conference on Knowledge Graph (ICKG). Washington: IEEE, 2024: 485-491. DOI: 10.1109/ICKG63256.2024.00068 .
[25]
ZhuY Q, WangX H, ChenJ, et al. LLMS for Knowledge Graph Construction and Reasoning: Recent Capabilities and Future Opportunities[J]. World Wide Web, 2024, 27(5): 58. DOI: 10.1007/s11280-024-01297-w .
YangS C, LiB M, TianR X. Exploration on Optimization of ETC Publishing Service Process Based on OCR and NLP Technology[J]. China ITS J, 2025(2): 16-23. DOI: 10.13439/j.cnki.itsc.2025.02.001 .
ZhaoW Y. Research on the Construction Method of Multi-modal Tangka Knowledge Graph[D]. Xining: Qinghai Normal University, 2024. DOI: 10.27778/d.cnki.gqhzy.2024.000724 .
GuoY Y. Research on the Construction Method of Blue and White Porcelain Based on Protégé[J]. Ceram Sci Art, 2023, 57(11): 39-41. DOI: 10.13212/j.cnki.csa.2023.11.130 .
[32]
LewisP, PerezE, PiktusA, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]//Advances in Neural Information Processing Systems 33 (NeurIPS 2020). Red Hook: Curran Associates Inc., 2020: 28613-28624.
[33]
YangA, YangB, HuiB, et al. Qwen2.5 Technical Report[PP/OL]. (2024-12-19)[2025-3-25].
[34]
HuJ E, ShenY, WallisP, et al. LoRA: Low-Rank Adaptation of Large Language Models[C]//International Conference on Learning Representations (ICLR 2022). Appleton: ICLR 2022.
[35]
ZengA, XuB, WangB, et al. ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools[PP/OL]. (2024-06-18)[2025-4-2].
[36]
GuoD, YangD, ZhangH, et al. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning[PP/OL]. (2025-01-22)[2025-4-15].
[37]
DemirhanH, ZadroznyW. Survey of Multimodal Medical Question Answering[J]. BioMedInformatics, 2024, 4(1): 50-74. DOI: 10.3390/biomedinformatics4010004 .