Abstract: Social media can be a highly valuable source of information during disasters. A crisis' development over time is of particular interest here, as social media messages can convey unfolding events in near-real time. Previous approaches for the automatic detection of information in such messages have focused on a static analysis, not taking temporal changes and already-known information into account. In this paper, we present a novel method for detecting new topics in incoming Twitter messages (tweets) conditional upon previously found related tweets. We do this by first extracting latent representations of each tweet using pre-trained sentence embedding models. Then, Infinite Mixture modeling is used to dynamically cluster these embeddings anew with each incoming tweet. Once a cluster reaches a minimum number of members, it is considered to be a new topic. We validate our approach on the TREC Incident Streams 2019A data set.

Abstract: The massive crowdsourced data generated on social networking platforms (e.g. Twitter and Flickr) provide free, real-time data for damage assessment (DA) even during catastrophes. Recent studies leveraging crowdsourced data for DA mainly focused on analyzing textual formats. Crowdsourced images can provide rich and objective information about damage conditions, however, are rarely researched for DA purposes. The highly-varied content and loosely-defined damage forms make it difficult to process and analyze the crowdsourced images. To address this problem, we propose a data-driven DA method based on multi-, crowd-sourced images, which includes five machine learning classifiers organized in a hierarchical structure. The method is validated with a case study investigating the damage condition of the City of Miami caused by Hurricane Irma. The outcome is then compared with a metric derived from NFIP insurance claims data. The proposed method offers a resource for rapid DA that supplements conventional DA methods.

Abstract: Integration of social media in emergency response environments presents specific organizational challenges, such as lack of resources or information credibility. Additionally, there exists individual resistance to change in these environments that could potentially discourage adoption. To identify and understand these challenges, we conducted semi-structured group interviews with emergency call takers and dispatchers. We find that these PSAP operators desire participation and explanation of changes throughout the organizational change process. Participants also articulated they desired training regarding change even when not directly affected. Though change management procedures often call for these strategies, they are commonly overlooked, leaving individuals to imagine worse case scenarios that manifest as additional stress in an already stressful work environment. It is suggested that a formalized change management process which directly addresses the identified challenges within the organizational technology adoption process (OTAP) is needed in order to mitigate undue stress.

Abstract: Social media are rich information sources during crisis events such as earthquakes and terrorist attacks. Despite myriad challenges, with the right tools, significant insight can be gained to assist emergency responders and related applications. However, most extant approaches are incomparable, using bespoke definitions, models, datasets and even evaluation metrics. Furthermore, it's rare that code, trained models, or exhaustive parametrisation details are openly available. Thus, even confirming self-reported performance is problematic; authoritatively determining state of the art (SOTA) is essentially impossible. Consequently, to begin addressing such endemic ambiguity, this paper makes 3 contributions: 1) replication and results confirmation of a leading technique; 2) testing straightforward modifications likely to improve performance; and 3) extension to a novel complimentary type of crisis-relevant information to demonstrate it's generalisability.

Abstract: We report on the design, development, and evaluation of a user labeling framework for social media monitoring by emergency responders. By labeling Twitter user accounts based on behavior and content, this novel approach identifies tweets from accounts belonging to Individuals generating Personalized content and captures information that might otherwise be missed. We evaluate the framework using training data from the 2018 Camp, Woolsey, and Hill fires. Approximately 30% of the Individual-Personalized tweets contain first-hand information, providing a rich stream of content for social media monitoring. Because it can quickly eliminate most redundant tweets, this framework could be a critical first step in an end-to-end information extraction pipeline. It may also generalize more easily for new disaster events since it relies on general user account attributes rather than tweet content. We conclude with next steps for refining and evaluating our framework in near real-time during a disaster response.

Abstract: Social media is perceived as a rich resource for disaster management and relief efforts, but the high class imbalance between disaster-related and non-disaster-related messages challenges a reliable detection. We analyze and compare the effectiveness of three state-of-the-art machine learning models for detecting disaster-related tweets. In this regard we introduce the Disaster Tweet Corpus~2020, an extended compilation of existing resources, which comprises a total of 123,166 tweets from 46~disasters covering 9~disaster types. Our findings from a large experiments series include: detection models work equally well over a broad range of disaster types when being trained for the respective type, a domain transfer across disaster types leads to unacceptable performance drops, or, similarly, type-agnostic classification models behave more robust at a lower effectiveness level. Altogether, the average misclassification rate of~3,8\% on performance-optimized detection models indicates effective classification knowledge but comes at the price of insufficient generalizability.

Abstract: Abstractive summarization is intended to capture key information from the full text of documents. In the application domain of disaster and crisis event reporting, key information includes disaster effects, cause, and severity. While some researches regarding information extraction in the disaster domain have focused on keyphrase extraction from short disaster-related texts like tweets, there is hardly any work that attempts abstractive summarization of long disaster-related documents. Following the recent success of Reinforcement Learning (RL) in other domains, we leverage an RL-based state-of-the-art approach in abstractive summarization to summarize disaster-related documents. RL enables an agent to find an optimal policy by maximizing some reward. We design a novel hybrid reward metric for the disaster domain by combining \underline{Vec}tor Similarity and \underline{Lex}icon Matching (\textit{VecLex}) to maximize the relevance of the abstract to the source document while focusing on disaster-related keywords. We evaluate the model on a disaster-related subset of a CNN/Daily Mail dataset consisting of 104,913 documents. The results show that our approach produces more informative summaries and achieves higher \textit{VecLex} scores compared to the baseline.

Abstract: Crisis responders need to locate events reported in social media messages that typically lack geographic metadata such as geotags. Toponyms, places names referenced in messages, provide another source of geographic information, however, the availability and granularity of toponyms in crisis social media remain poorly understood. This study examines toponym usage and granularity across six categories of crisis-related information posted on Twitter during a severe storm. Findings show users often include geographic information in messages describing local and remote storm events but do so rarely when discussing other topics, more often use toponyms than geotags when describing local events, and tend to include fine-grained toponyms in reports of infrastructure damage and service disruption and course-grained toponyms in other kinds of storm-related messages. These findings present requirements for hyperlocal geoparsing techniques and suggest that social media monitoring presents more immediate affordances for course-grained damage assessment than fine-grained situational awareness during a crisis.

Abstract: The usage of non-authoritative data for disaster management provides timely information that might not be available through other means. Wikipedia, a collaboratively-produced encyclopedia, includes in-depth information about many natural disasters, and its editors are particularly good at adding information in real-time as a crisis unfolds. In this study, we focus on the most comprehensive version of Wikipedia, the English one. Wikipedia offers good coverage of disasters, particularly those having a large number of fatalities. However, by performing automatic content analysis at a global scale, we also show how the coverage of floods in Wikipedia is skewed towards rich, English-speaking countries, in particular the US and Canada. We also note how coverage of floods in countries with the lowest income is substantially lower than the coverage of floods in middle-income countries. These results have implications for analysts and systems using Wikipedia as an information source about disasters.