運用社群媒體建置時空災害知識圖譜於災害管理

Abstract

台灣深受自然災害影響，每當災害發生時，即時且適當地應變係為重要任務，可以減少財物損失和人員傷亡。而社群媒體具有即時性的特色，同時擁有高度可及性，此外，所蒐集的社群媒體亦可成為歷史資訊。因此，如能有效運用社群媒體上的資訊，可以迅速反映來自民眾傳遞之災害發生時間、空間、事件和嚴重程度等訊息，同時在未來的備災時提供防災建議與參考，協助災害管理。然而，過去研究中較缺乏發展自動化獲得社群媒體中資訊並以標準化知識架構表示，以及表達來自群眾由下而上 (bottom-up)所建構之時空知識 。為改善前述研究缺口，本研究選擇使用社群媒體，自動化建構以標準框架表示知識的時空災害知識圖譜(spatio-temporal diaster knowledge graph)來表達災害知識語意，此外，將知識圖譜應用於災害知識查詢和推論，以支援災害管理。 本研究主要利用人工智慧模型 (artificial intelligence, AI)進行自然語言處理任務，將從社群媒體資源中萃取災害知識。本研究以臺灣最受歡迎之社群媒體平台之一的批踢踢實業坊 (PTT)為研究材料，同時，並以BERT (Bidirectional Encoder Representations from Transformers) 模型進行自然語言處理。在完成文本前處理後，首先透過指代消解任務 (ellipsis and coreference resolution)，將社群媒體文字中缺漏之主詞和使用之代名詞，還原至原本的字詞以利後續知識圖譜的建構；接著根據建構的災害知識本體 (disaster ontology)，訂定實體辨識和關係萃取應涵蓋之災害知識並整合，再依照訂定的規則，找出三元組並建構時空災害知識圖譜，並應用SPARQL語法和地理資訊系統操作進行災害知識語意查詢。成果評估方面，使用機器學習指標評估模型的性能，並和ChatGPT 建構成果比較衡量本研究之知識圖譜建構成果，並選定災害個案分析和官方報告比對來探討時空災害知識圖譜中可獲得之資訊特性。 研究結果顯示，結合指代消解、實體辨識和關係萃取之方法，可有效萃取社群媒體文字中的三元組，並建構成時空災害知識圖譜，同時也能以此圖譜內容進行知識查詢，並將其實際運用於地震及颱風的個案，瞭解不同災害之群眾討論內容差異，同時應用於災害管理。本研究所提出的方法可提升社群媒體資訊存取之效率，並補充以往缺乏的由下而上及時空相關知識。在實務應用層面，可運用災害知識和空間查詢應用於災害管理，作為災害救援人員或決策者提供在災害採取行動之建議。

Natural disasters have a profound impact on Taiwan, necessitating immediate and effective responses to minimize damage and casualties. Social media, with its real-time and highly accessible features, serves as a promising resource for disaster management. Moreover, data collected from social media can also serve as a valuable source of historical information. Therefore, effectively leveraging social media information can help reflect the time, location, events, and severity of disasters as reported by the public. Additionally, it can provide disaster preparedness suggestions and references for future disaster management. However, existing studies have not focused on developing automatic methods to extract information from social media into a knowledge representation format, as well as to express spatiotemporal knowledge constructed from bottom-up contributions by the public. This study aims to construct a spatio-temporal disaster knowledge graph (ST-DKG) for disaster semantic interpretation and disaster management from social media. In addition, the ST-DKG is applied to disaster knowledge queries and reasoning in geographic information systems, achieving a better solution and understanding in disaster management. This study primarily utilizes artificial intelligence (AI) models for natural language processing. The study uses PTT, one of Taiwan's most popular social media platforms, and employs the BERT model for natural language processing. After completing text preprocessing, the study first restores omitted subjects and pronouns in social media texts through an ellipsis and coreference resolution task, facilitating subsequent knowledge graph construction. Then, based on the constructed disaster ontology, the study identifies and integrates disaster knowledge for entity recognition and relation extraction. It then follows defined rules to identify triples and construct the ST-DKG, applying SPARQL and geographic information system operations for disaster knowledge semantic querying. For evaluation, machine learning metrics are used to assess the model's performance, and the ST-DKG results are compared with those generated by ChatGPT. The study also performs disaster case analysis and compares the findings with official reports to explore the characteristics of information obtainable from the ST-DKG. The results show that by utilizing the techniques of coreference resolution, entity recognition, and relation extraction, it is feasible to obtain triples from social media texts, thereby enabling the construction of an ST-DKG. This knowledge graph enables knowledge querying. It can also be applied to the cases of earthquakes and typhoons to understand the differences in the content of public discussions on different disasters, and also use it in disaster management This study proposes an automated method for processing social media information and constructing an ST-DKG, improving the efficiency of accessing social media information and supplementing the previously lacking bottom-up and spatiotemporal knowledge. In practical applications, disaster knowledge and spatial queries can be applied to disaster management, contributing to providing disaster workers or authorities with the most appropriate actions in disasters.