融合语义结构感知与动态序列建模的蛋鸡疫病知识联合抽取模型

Joint Knowledge Extraction Model for Laying-hens Diseases Integrating Semantic-structural Awareness and Dynamic Sequence Modeling

摘要:

针对蛋鸡疫病诊断知识图谱构建中存在实体重叠和长程依赖等问题,提出了一种融合语义结构感知与动态序列建模的蛋鸡疫病知识联合抽取模型 SynGraph-HenJE(Syntactic and graph for hen disease joint extraction)。 首先,模型采用 BERT(Bidirectional encoder representation from transformers)预训练模型对蛋鸡疫病文本进行语义编码, 并通过自主构建的 Semantic-guided edge-aware GATv2(SGEA-GATv2)图注意力机制增强语义编码,提升模型对重叠实体的区分能力与关系判别能力;其次,通过关系感知门控机制,融合时间卷积网络(Temporal convolutional network, TCN)提取的局部特征和双向长短时记忆网络(Bidirectional long short-term memory,BiLSTM)提取的全局特征,有效解决模型对蛋鸡疫病文本的长程依赖问题;最后,改进模型在自建的蛋鸡疫病数据集和两个公共数据集进行了验证试验,结果表明,精确率分别达到 95. 0% 、84. 2% 和 82. 5% ,召回率分别达到 92. 5% 、81. 5% 和 78. 5% ,F1 值分别达到 93. 7% 、82. 8% 和 80. 4% ,与 Spert、CasRel、TPLinker 和 OneRel 模型对比,精确率平均提升了 7. 71 个百分点,召回率平均提升了 4. 58 个百分点,F1 值平均提升了 6. 16 个百分点,模型表现出较好的泛化能力。 SynGraph-HenJE 在 6 类实体上的精确率均超过 90% ;在 5 类关系上的精确率也显著优于对比模型,精确率均超过 88. 0% 。 本研究有效解决了实体重叠和长程依赖问题,提高了蛋鸡疫病文本的三元组提取质量,为智能诊断提供技术支持。

Abstract:

Aiming to address the challenges of overlapping entities and long-range dependencies in constructing a knowledge graph for diagnosing laying hens' diseases, syntactic and graph for hen disease joint extraction (SynGraph-HenJE ), a joint extraction model that integrated semantic-structural awareness with dynamic sequence modeling was proposed. The model employed the pre-trained bidirectional encoder representations from transformers (BERT )architecture to perform semantic encoding on text data related to poultry diseases. To enhance this semantic representation, a custom semantic-guided edge-aware GATv2 (SGEA-GATv2)module was introduced, which strengthened the model’ s ability to distinguish overlapping entities and improve relational discrimination. Through a relationship-aware gating mechanism, it integrated local modeling extracted by a temporal convolutional network (TCN)with global modeling extracted by a bidirectional long short-term memory (BiLSTM)network, effectively enhancing the model’ s capacity to capture long-range dependencies in laying hens'??disease texts. The proposed model was evaluated on both the self-constructed laying hens' disease dataset and two public datasets. The results showed that the model achieved accuracies of 95. 0% , 84. 2% , and 82. 5% , recall rates of 92. 5% , 81. 5% , and 78. 5% , and F1-scores of 93. 7% , 82. 8% , and 80. 4% , respectively. Compared with the reference model, the proposed model demonstrated an average accuracy improvement of 7. 71 percentage points, an average recall improvement of 4. 58 percentage points, and an average F1 score improvement of 6. 16 percentage points, demonstrating the strong generalization capability of the proposed model. SynGraph-HenJE achieved accuracies above 90% across all six entity categories. For the five relation categories, the model also outperformed the baselines, with all accuracy values exceeding 88. 0% . The research effectively addressed the challenges of entity overlap and long- range dependencies, improving the quality of triplet extraction from laying hens?? disease texts and providing technical support for intelligent diagnosis.