利用指向圖資訊分類多標籤引用目的之研究

Abstract

當撰寫科技論文時，若在內文提及先前相關研究，作者往往會使用規定格式來引用相關論文。引用除了可以給予被引用論文之作者應有的尊重外，並提供讀者獲得更詳細訊息的查找方向。由於引用往往是作者在撰寫時有目的的選取相關之研究，並在內文中提及提及，因此引用的目的也應該是可以被大眾所利用的資訊。以往已經有一些先前研究是以引用句當作唯一的線索，去分析預測作者撰寫時引用對應論文之目的。然而要判斷引用目的所需要的資訊，有時卻不只內涵引用符號的引用句而已，相關的上下文內容在最近的研究也也被證實對判斷引用目地有幫助。但現存的方法中，並沒有一個模型可以同時預測引用目的，同時說明為何模型會認為這項引用包含此引用目的。我們相信引用的目的和對應的上下文，對於研究論文之間關係有所幫助，因此在此篇碩士論文，我們提出一個結合Transformer encoder，並利用GGCN把指代連結圖的訊息一併納入的方法。實驗結果顯示我們的方法除了可以有效預測引用句的引用目的，也可輸出對判斷各個目的有幫助的上下文。

In a scientific paper, authors usually use the prescribed format to cite the related papers when mentioning the previous works in the text. Citations not only give credit to the authors of the cited papers, but also provide readers the detailed information with a clear search direction. Because citations are often the authors' purposeful selection of the related research, the intention of the references should also be useful information to the public. Although there have been several previous works using the citing sentence as the only clue to analyze and predict the author's purpose of citing the corresponding paper, the recent work shows that it is necessary to judge the citation intent with the information from both the citing sentence and the related contexts. However, none of the existing methods predict the citation intent and offer the evidence simultaneously when the model deems that the citation contains the citing intent. We believe that the identifying the purpose of citation and the corresponding context is useful for investigating the relationship between papers. In this work, we propose a deep learning method that combines a Transformer encoder and uses Gated Graph Convolution Network (GGCN) to incorporate information implied in the coreference graph. Experimental results show that our method not only can effectively predict the citation intent of the citing sentence but also recognize the contexts that help judge each intent.