請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73668
標題: | 基於分段卷積神經網絡進行生物醫學關係萃取 Using Piecewise Convolutional Neural Networks for Biomedical Relation Extraction |
作者: | Yu-Tang Kuo 郭毓棠 |
指導教授: | 魏志平(Chih-Ping Wei) |
關鍵字: | 關係萃取,生醫關係萃取,關係分類,深度學習,分段卷積神經網絡, Relation extraction,Biomedical relation extraction,Relation classification,Deep learning,Piecewise convolutional neural network, |
出版年 : | 2019 |
學位: | 碩士 |
摘要: | 關係萃取是自然語言處理一項重要的任務,它可以被應用在很多的工作,像是知識圖譜的建立、問句回答、文章總結等。但是,自然語言處理其實是一個複雜的過程,很容易造成錯誤的傳遞以及累積,使得最後的表現不理想。為了避免這個問題,最近這幾年,許多研究捨棄自然語言處理的方法,而採用類神經網路的方式進行關係萃取。PCNN 是一種類神經網路的技術,他根據一段上下文以及兩個名詞的位置,然後產出不同關係的機率分佈。
雖然採用 PCNN 來進行關係萃取已經能獲得不錯的表現,可是它其實還有兩個可以改善的方向。首先,關係是具有方向性的,然而在兩個反向的關係中事實上存在一些共同資訊是可以一起被萃取的,如果我們能讓 PCNN 學習反向關係中的共同資訊關係,那麼它將有機會獲得更好的表現。第二,PCNN 是個關係萃取的通用架構,並沒有針對特定的領域做出調整。然而這些特定領域其實都有一些領域知識可以幫助 PCNN 適應該領域。如此 PCNN 針對該領域的關係萃取也會變得比較好。在我們的研究中,生醫領域其實存在一些領域知識可以讓 PCNN 學習,因此我們修改了 PCNN 的架構,讓 PCNN 可以適應生醫領域以達到更好的表現。 綜合上述兩者,我們提出了一套改良 PCNN 的架構,用以從生醫文獻摘要中萃取生物醫療關係,而根據我們的實驗結果,我們發現生醫領域的知識的確會讓 PCNN 的表現變得更好,而在反向關係之間的共同資訊這方面,則表現的沒有比 PCNN 好。 Relation extraction is an important task for natural language processing (NLP) and has many applications, such as the construction of knowledge graph, question answering, and document summarization. However, NLP is a complicated process. Errors in NLP will propagate and accumulated to its application task and accordingly cause negative influences to the effectiveness of the focal application task. To avoid this problem, recent studies have devoted themselves to the use of neural networks for relation extraction without involving NLP. Zeng et al. (2015) proposed PCNN, which is one of the neural networks for relation extraction. PCNN requires a context (textual passage) and locations of two entities in the passage. Then it outputs the probability distribution of the relation classes. However, PCNN incurs two limitations.First, although a relation is directional, there still exists some common information between the same relations but with different directions (referred to as reverse relations).If we can exploit common information inherent to these reverse relations, the effectiveness of relation extraction may be improved.Second, PCNN is a generic architecture that can applied to various application domains. Given a specific domain, we can exploit domain knowledge of specific domain and incorporate it into the generic architecture to improve its effectiveness. In this research, we focus on extracting biomedical relations from the abstracts of biomedical articles. To address the aforementioned limitations of PCNN, we will propose a PCNN-based method that will incorporate biomedical domain knowledge for biomedical relation extraction. According to our evaluation results, we observe that the incorporation of biomedical domain knowledge into the PCNN model indeed improves the effectiveness of relation extraction. However, the common information between reverse relations cannot help improve relation extraction. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73668 |
DOI: | 10.6342/NTU201901946 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊管理學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-108-1.pdf 目前未授權公開取用 | 1.48 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。