眾包災情回報篩選器

Abstract

災情回報的效率和正確性會直接影響救災的品質，大部份的災情回報是由政府處理，相關人員從各種來源收集資訊、確認它的正確性，再發佈正確的訊息。這樣的流程確保了資訊的品質，卻時常造成時間上的延遲。近年由於網路及行動裝置的發展，許多群眾平台如Sahana、 Ushahidi 平台、莫拉克颱風災情地圖開始應用在災情回報上。然而，雖然這些平台在防災上扮演著日益重要的角色， 它有三項明顯的不利因素：正確性、重複回報、格式不一致且沒有重要資訊的欄位。本研究發展一套結合電腦計算和群眾智慧的方法 (ACI filter) 來降低這些不利因素。ACI filter包含人工智慧演算法和群眾智慧。我們使用人工智慧演算法找出很有可能為正確回報的資料，並且篩除很有可能為錯誤回報的資料。剩下的回報則交由群眾智慧處理。群眾有三項工作：合併重複的災情回報資訊、篩選錯誤的回報、將資訊的格式一致化。我們使用2012 06010豪雨事件的災情回報資料進行可行性評估，當中包含876筆資訊。研究成果顯示ACI filter篩除26.25%的回報，偽陰性率為0.00%，偽陽性率為3.91%，同時ACI filter將11.3%的災情回報格式化。總結來說，我們開發了ACI filter，提高群眾平台在在災情回報上的可用性。

The quality of disaster mitigation is directly connected to the efficiency and quality of disaster responses. In current state of practice, most disaster responses require manual process, mainly handled by government officers, to eliminate the incorrectness. They usually collect the information from multiple sources, verify the correctness and announce the verified information. This process ensures the quality of the information but often result in time-delay. Recently, due to the rapid development of Internet and mobile devices, many crowd-based platforms, such as Sahana, Ushahidi crisis map and Typhoon Morakot Crisis Map, have been developed and employed for disaster responses. Although these platforms sometimes play important role in disaster mitigation, they have three obvious drawbacks: correctness, duplication, and inconsistent format. In this research, we aim to develop a computational method, ACI filter, to eliminate the drawbacks. ACI filter integrates both artificial intelligence (AI) filter and the human intelligence using crowd sources. We used AI filter to retrieve the responses that are highly possible to be correct and eliminate the responses that are highly possible to be incorrect. Remaining responses are filtered by crowd sources. We used the crowd for three purposes: to combine the duplicated responses, to eliminate incorrect responses, and to synchronize the formats of the responses. To verify the ACI filter, we used 876 disaster responses collected from a real disaster in Taiwan caused by a serious torrential rain in June 10 2012. We recruited 284 volunteers from the Internet to participate the test. Each participant is asked to answer twenty short questions. Average testing time for each participant is 212 seconds, meaning 10.6 seconds per question. The research results show that ACI filter eliminates 26.25% inaccurate responses. False negative rate (i.e. mistakenly validation) is 0.00%. False positive rate (i.e. mistakenly elimination) is 3.91%. In conclusion, ACI filter, combining the machine and human power, can successfully improve the accuracy of the responses with the crowd. The method can be extended and applied to cope with large-scale disaster responses.