財務社交媒體的論點關係識別與時間推理

Abstract

生活中眾多討論與對話由各個論點與論點間的關係所構成，例如金融市場相關討論，社會議題辯論，以及論文寫作內容。因此在自然語言處理中，有相當多作品處理論點關係識別這個重要主題。雖然論點之間的關係相當多元，可能有各式各樣的分類，但是使用支持、攻擊、無關或其他，是一個基本且常見的分類組合，很多相關分類也是基於此稍作轉變或細分。除了論點之間的關係，時間在許多討論與人們的生活場景中也扮演重要角色，因此長期以來也一直是重要且熱門的研究主題。從傳統處理時態，到現今針對不同任務，使用不同方法標記與處理時間的資料集與模型之相關研究中，經常提及因為時間資訊在不同語境下意涵多變且經常隱含等相關性質，使時間推理在不同領域的應用仍具挑戰。 因此，這篇論文的主要貢獻為邀請具金融工作與知識背景的專業人士，標記中文金融社交媒體討論中的論點關係和時間資訊，建立了一個資料集。並且蒐集其他相似論點關係識別的研究，討論此類任務跨語言和跨領域的學習能力，並且提出一個易於實作，且高效率的複習方法，處理學習中常見的的遺忘問題。近年來，小型資料標註和學習也日漸重要，因此我們討論語言、領域和任務順序外，也討論資料集大小對效能的影響，以提供未來研究於任務順序安排的參考方向。此外，因為時間資訊的多變性，我們的資料集提供了針對中文金融社交討論平台的時間標記，並結合相關中文金融研究，比較與分析文字時間資訊推理的影響。

Many discussions are composed of arguments and their interactions, such as discussions on financial market, debates of social topics and essay writing. Therefore, argument relation identification is an important topic in language processing. Although there are numerous relations, support, attack and none or other, is an essential and common set with many variations. In addition to the relation of arguments, temporal knowledge is also crucial not only in many discussions, but also for many aspects of people's daily life, and thus has long been a popular language research topic as well. There are multiple temporal datasets focusing on various targets, e.g. event relation and duration. We can see from previous works that temporal knowledge and its influence on many other tasks still remain complex and puzzling because they're often implicit and versatile in different scenarios, which make it hard to have universal resources and standard for all domains and purposes. Therefore, our work enrich argumentative and temporal resources with a Chinese financial dataset, TREE (Time Reveals valuE Expression), which has argumentative and temporal labels annotated by experts to further understand argument relation and temporal knowledge. We discuss the challenge and quality of our dataset and also collaborate with other relative and similar works to examine our methods. Inspired by post-training works, we develop a simple and resource efficient method to help models overcoming forgetting problem when learning form transferring relative tasks. We find not only the order of training tasks matters but also language families, domains and sizes of datasets. For the financial temporal inference task, we collaborate and compare with other Chinese financial works to analyze the influence of temporal inference task based on text.