透過您的圖書館登入
IP:18.227.52.11
  • 學位論文

基於混合神經網路之知識複雜推理系統設計與實作-以豬隻疾病知識建立為例

An Intelligent Knowledge Complex Inference System Based on Combined Neural Networks: Design and Implementation - A Case of Pig Disease Knowledge Building

指導教授 : 龔旭陽
若您是本文的作者,可授權文章由華藝線上圖書館中協助推廣。

摘要


本論文對畜牧豬隻飼養知識進行整理,目標為將分散於網路的飼養資料執行自然預言處理後建置豬隻飼養知識圖譜,然而大量知識轉化到知識圖譜的過程,有部分知識是沒有直接關聯的,這造成了飼養知識圖譜的不完整性問題,據此,針對畜牧豬隻飼養知識圖譜的補全,本論文提出知識複雜推理機制(Knowledge Complex Reasoning Mechanism, KCRM)解決以下議題:(1)為了有效抽取知識中之實體,本論文採用雙向長短期記憶結合條件隨機域實現知識文本的命名實體識別並將其分類,再透過注意力機制之雙向門控循環單元針對句子中的實體間的關係進行抽取。(2)由於知識來源廣泛,抽取的實體、值與關係往往會有一定程度錯誤,包括有被錯誤識別的實體或被錯誤分類的關係,因此為了提高知識圖譜的置信度,透過知識融合對其進行處理,並將半結構化與非結構化的知識進行整合。(3)最後針對知識圖譜的不完整性,透過推理來補全知識圖譜,做法為利用卷積神經網路(Convolutional Neural Networks)對三元組進行卷積後透過權重計算得到三元組推理的分數藉以判斷三元組的有效性,三元組中的實體或關係可以通過知識推理得出,實現知識圖譜中已知的知識來推理未知且隱藏的知識,進而達到知識圖譜的擴展。於知識圖譜實作部分,本論文之知識來源包括動物疾病知識庫、豬隻生產醫學平台和畜產試驗所收集畜牧豬隻知識與用語,建立台灣一套豬隻疾病知識圖譜辭典範例,並採用Bi-Lstm-CRF與Attention-Bi-GRU進行知識抽取,及採用ConvKB實現知識推理。最後論文以準確率、召回率以及 F-measure驗證常用分詞及知識抽取方法之效率,並針對常用推理模型之平均排名MR值與Hit@k進行實驗,證明ConvKB於畜牧疾病推論的整體精確性優於ConvE與TransE。

並列摘要


This thesis organizes the knowledge of animal husbandry and pig breeding. The goal is to implement the natural language process of the feeding data scattered on the network to build a knowledge graph of pig breeding. However, when knowledge is converted into a knowledge graph, some of the knowledge is not direct related, i.e., some knowledge is implicit. The implicit knowledge has caused the incompleteness of the feeding knowledge graph. To complete the knowledge graph of feeding pigs, this thesis proposed the Knowledge Complex Reasoning Mechanism (KCRM) to solve the following issues: (1) To effectively extract the entities in the knowledge, this thesis adopted the Bidirectional Long-short Term Memory combined with Conditional Random Fields to identify and classify the named entities of the knowledge text, and then uses the bi-directional gated loop unit of the attention mechanism to target the entities in the sentence. (2) Due to the wide range of knowledge sources, the extracted entities, values and relationships often have a certain degree of error, including misrecognized entities or misclassified relationships. Therefore, to improve the confidence of the knowledge graph, knowledge fusion Process and integrate semi-structured and unstructured knowledge. (3) Finally, for the incompleteness of the knowledge graph, the knowledge graph is completed by reasoning. The Convolutional Neural Networks scheme is adopted to convolve the triplet and then the weighted calculation is used to obtain the triplet inference score to judge the validity of the triplet. The unknown and implicit knowledge can be inferred by the knowledge of the known knowledge graph, and the entities or relations in the triple can be obtained by knowledge reasoning, and then the knowledge graph can be expanded. This thesis implemented the knowledge graph of pig breeding using the combined neural network schemes. The knowledge sources include the animal disease knowledge base, the pig production medical platform and the animal husbandry laboratory to collect the knowledge and terms of animal husbandry pigs. This thesis established the first set of examples of the pig disease knowledge dictionary and graph in Taiwan. This thesis adopted Bi-LSTM-CRF and Attention-Bi-GRU for knowledge extraction, and ConvKB for knowledge reasoning. Finally, the paper verifies the efficiency of word segmentation and knowledge extraction methods with accuracy, recall, and F-measure. This thesis conducts experiments on the average ranking MR value of common reasoning models and Hit@k, which proves that ConvKB has excellent overall accuracy in animal husbandry disease inference than ConvE and TransE.

參考文獻


斎藤康毅. (2019). Deep Learning 2|用Python進行自然語言處理的基礎理論實作: 歐萊禮.
呂國豪. (2018). 結合NLP與Bi-LSTM之智慧飼養知識計算系統設計與實作: guo li ping dong ke ji da xue zi xun guan li xi.
张仲伟, 指挥控制工程学院, 中. 陆., 张仲伟, 曹雷, 陈希亮, 寇大磊, . . . 73671部队, 中. (2019). 基于神经网络的知识推理研究综述. [Survey of Knowledge Reasoning Based on Neural Network]. 计算机工程与应用(2019年 12), 8-19,36.
官赛萍, 网络数据科学与技术重点实验室, 官赛萍, 靳小龙, 贾岩涛, 王元卓, . . . 中国科学院大学计算机与控制学院, 北. (2018). 面向知识图谱的知识推理研究进展. [Knowledge Reasoning Over Knowledge Graph: A Survey]. 软件学报(2018年 10), 2966-2994.
林宜潔 (Producer). (2019, 07 01). 口蹄疫拔針成功!但台灣豬想出口還得等1年. 遠見雜誌. Retrieved from https://www.gvm.com.tw/article/67040