Source-linked AI summary
Processing Social Media Messages in Mass Emergency: A Survey
Muhammad Imran, Carlos Castillo, Fernando Diaz, Sarah Vieweg
TL;DR
Emergency social media contains potentially timely, actionable information, but extracting it requires handling massive, heterogeneous, informal, and uneven-quality message streams. The paper surveys computational methods for filtering, classifying, extracting, aggregating, and summarizing such content, finding a mature research area while noting that operational value and adoption remain unresolved.
Problem
Emergency social media is abundant, heterogeneous, informal, and variable in quality, making it difficult to identify credible and useful information for response.
Method
The paper surveys computational methods for processing crisis-related social media, covering approaches from event detection through filtering, classification, extraction, aggregation, and summarization.
Results
The research community has achieved a fairly high degree of maturity in filtering, classifying, processing, and aggregating social media data during crises.
Takeaways & Limitations
The survey provides methods for building tools that identify, filter, and organize overwhelming crisis data for emergency-response stakeholders.
Takeaways & Limitations
Social media use remains experimental for many organizations and is not yet part of their normal day-to-day operations.
Abstract
from arXiv · showhide
Social media platforms provide active communication channels during mass convergence and emergency events such as disasters caused by natural hazards. As a result, first responders, decision makers, and the public can use this information to gain insight into the situation as it unfolds. In particular, many social media messages communicated during emergencies convey timely, actionable information. Processing social media messages to obtain such information, however, involves solving multiple challenges including: handling information overload, filtering credible information, and prioritizing different classes of messages. These challenges can be mapped to classical information processing operations such as filtering, classifying, ranking, aggregating, extracting, and summarizing. We survey the state of the art regarding computational methods to process social media messages, focusing on their application in emergency response scenarios. We examine the particularities of this setting, and then methodically examine a series of key sub-problems ranging from the detection of events to the creation of actionable and useful summaries.
1. INTRODUCTION
The survey examines computational methods for processing social media content during disasters, where timely information can support situational awareness but overwhelming data volumes and organizational barriers complicate use.
- Sudden-onset crises create uncertainty and demand rapid decisions with minimal information, while social media can enhance situational awareness.
- The survey aims to help researchers and developers build tools that identify, filter, and organize crisis-related social media data for end users.The intended users include formal response agencies, humanitarian organizations, and other stakeholders.
- Social media use in disaster response developed from earlier Internet-based coordination and user-generated crisis information toward broader, more immediate communication channels.
- Formal agencies are interested in incorporating social media, but monitoring, verifying, and responding to messages can impose substantial personnel and administrative costs.
- The paper systematically analyzes computational methods for processing social media messages, including operations spanning detection, filtering, classification, and summarization.
2. USERS AND INFORMATION NEEDS
The section presents disaster information needs as role-dependent and emphasizes that the public, responders, and other stakeholders actively produce, exchange, and use crisis information.
- People in disasters make decisions using available information, and members of the public often act as early responders for themselves and others.
- Public, agency, and other stakeholders use social media to communicate, organize relief, and perform tasks such as reporting hospital availability and coordinating medical responses.
- Social media is recognized as a source of timely information across the disaster lifecycle, including preparation, impact, response, and recovery.
- Information needs vary by stakeholder responsibilities; firefighters may need reports about smoke, precautions, and traffic, whereas individuals may seek personal safety or neighborhood updates.
- Because useful information depends on each stakeholder’s goals, technological methods are needed to locate relevant messages quickly.
3. SYSTEMS FOR CRISIS-RELATED SOCIAL MEDIA MONITORING
Existing crisis-monitoring systems combine social media collection, language processing, extraction, detection, grouping, classification, translation, and visualization to present crisis information across temporal, spatial, and thematic dimensions.
- Academic systems vary in maturity, with some deployed in real situations and others still under development.
- Most systems use dashboards that summarize crisis social media through message lists, time series, maps, and proportional visualizations.
- Core capabilities include collecting messages, applying NLP and named-entity processing, extracting information and geotags, and monitoring message volume for events or alerts.
- Systems also cluster similar messages, classify messages or groups, and automatically translate messages.
- Many systems focus on processing social media rather than addressing specific stakeholder needs, which may hinder practitioner adoption.
- Some systems enhance situational awareness by combining event detection, text classification, online clustering, geotagging, and temporal and spatial presentation.
4. DATA CHARACTERIZATION, ACQUISITION, AND PREPARATION
This section characterizes crisis social-media data, explains how message content varies across events and locations, and surveys acquisition, preparation, and processing challenges. It emphasizes that collection choices and imperfect data shape subsequent analysis.
- Data characteristics: Social media activity during emergencies increases and contains varied information about social, built, and physical environments.Messages can describe people and reactions, property and infrastructure, hazards, weather, and other environmental factors.
- Data characteristics: Across six crisis events, tweets reporting negative consequences comprised 20%-60%, donation offers or requests 15%-70%, and warnings or advice 5%-20%.
- Data characteristics: Tweets from affected areas contain more requests for help and warnings, whereas distant-area tweets more often express concern and condolences.
- Acquisition and limitations: Social-media datasets are often incomplete because research relies heavily on Twitter and limited access to other platforms or providers.
- Preparation: Preprocessing commonly includes natural-language processing and filtering to prepare messages for analysis.
- Processing trade-offs: Retrospective analysis favors accuracy through fuller context, while live analysis favors lower latency despite partial information.The preferred approach depends on the use case: responders may prioritize latency, whereas forensic analysts may prioritize accuracy.
- Challenges: Crisis-scale processing must handle millions of messages, unavoidable redundancy, noisy language, heterogeneous sources, multilingual content, and limited contextual understanding.
5. EVENT DETECTION AND TRACKING
This section surveys event-detection and tracking methods adapted to rapidly arriving, high-volume, short, noisy social-media streams. It distinguishes retrospective and online new-event detection and reviews representative feature, clustering, graph, and burst-based approaches.
- Background: Crisis event-detection methods historically adapt techniques from news detection, topic detection, and document-stream research.
- Background: Social-media event detection requires approaches suited to faster, larger-volume streams of short, noisy, unstructured content.
- New Event Detection: New Event Detection identifies whether a continuously monitored message describes an event not previously reported, usually using similarity-based novelty.
- Retrospective detection: Retrospective detection clusters past messages using textual, social, temporal, or spatial similarity, with graph analysis and community detection among surveyed strategies.
- Online detection: Online detection operates with low latency and avoids previously seen messages or prior event knowledge, making it suitable for real-time monitoring.
- Representative approaches: Representative systems detect events through bursty keywords or query frequencies, wavelet-based signal clustering, message segments, named entities, dates, and latent-variable classification.
6. CLUSTERING, CLASSIFICATION, EXTRACTION AND SUMMARIZATION
This section introduces a broad processing stage for crisis messages after event- or topic-related content has been identified. It separates methods that classify whole data items from methods that extract useful information across content.
- Processing tasks: Post-detection processing divides techniques into whole-item classification and extraction of useful information from one or more data items.
6.1. Classifying Social Media Items
This section surveys classification approaches for organizing crisis-related social media, emphasizing that categories vary with available data and response-agency needs. It contrasts supervised and unsupervised methods, their uses, and their evaluation challenges.
- Classification dimensions: There is no single standard categorization scheme for crisis-related social media messages.Different studies use different approaches, and the relevant categories can change across crises.
- Classification dimensions: Crisis-related messages are classified across dimensions including content, information type, source, credibility, time, and location.The selected categories depend on the social media data available and the information needs of response agencies.
- Supervised classification: Supervised classification learns a predictive model from labeled training examples and evaluates it on a separate labeled testing set.The model uses features from example items to label new, unseen messages.
- Supervised classification: Training sets typically range from a few hundred to a few thousand examples, while accuracy generally decreases when models transfer across crises or time periods.More examples usually improve results, with diminishing returns after a certain point.
- Evaluation: Reported classification accuracies range from 0.60 to 0.90, but comparisons remain difficult because studies use different datasets, baselines, and performance measures.The survey notes that newer research collections may make systems easier to compare.
- Unsupervised classification: Clustering groups similar unlabeled items and can reduce the number of messages humans must examine, whereas unsupervised outputs may align less well with responders’ expected categories.Supervised methods require labeled examples, while unsupervised methods are useful when labels are costly, delayed, or unavailable.
6.2. Sub-Document Analysis
This section covers sub-document analysis methods that extract structured information or generate summaries from crisis-related social media. These methods support filtering and aggregation but face linguistic, temporal, and scalability challenges.
- Information extraction: Named entity extraction detects text referring to people, organizations, or locations and supports semantic enrichment.It is described as the most common information-extraction task.
- Information extraction: Information extraction converts unstructured crisis messages into normalized, machine-readable records that can be filtered, sorted, or aggregated.For example, casualty reports can be represented with affected-person counts, report types, and locations.
- Information extraction: Crisis information extraction is difficult because social media contains informal writing and many ungrammatical sentences.Probabilistic sequential models are presented as generally more effective than regular-expression heuristics.
- Information extraction: Supervised extraction systems can identify problem expressions or apply category-dependent extraction to messages about infrastructure damage, donations, and caution or advice.One approach matches extracted problems with messages describing solutions, while another first classifies tweets and then extracts category-specific information.
- Text summarization: Text summarization produces concise representations of core topics, commonly by extracting key sentences from one or more documents.Crisis summarization must operate incrementally and temporally to provide updates beyond what users have already read.
- Text summarization: Temporal summarization research defines time-sensitive precision and recall metrics for updates to unfolding crisis events.The TREC Temporal Summarization track focuses on generating updates immediately after events occur.
- Challenges: Crisis summarization remains preliminary because feature importance is poorly understood and aggressive inter-sentence similarity computation scales poorly.Existing work has primarily reused features from batch text summarization.
6.3. Challenges
The section identifies practical challenges in applying learning methods to crisis social media, especially the cost and reliability of labels and the difficulty of transferring models across crises.
- Labeling: Human labeling can be costly, unreliable, or unavailable during disasters, particularly for multilingual populations and domain-specific tasks.Hybrid human–automatic labeling and active learning are proposed to address these constraints.
- Domain adaptation: Reusing a classifier trained on a previous crisis can cause a substantial loss in accuracy, even when the crises share several elements.This motivates domain adaptation methods that adjust classifiers to datasets with different characteristics.
- Domain adaptation: Domain adaptation methods had not yet been applied to crisis-related social media data according to the survey.The methods are intended to preserve classifier accuracy when moving between datasets.
7. SEMANTIC TECHNOLOGIES IN DISASTER RESPONSE
This section surveys semantic technologies for enriching crisis social media, enabling concept-based search, and supporting interoperability among systems used by different disaster-response organizations.
- Semantic enrichment: Semantic technologies link data elements to machine-readable concepts, representing a crisis as an interconnected set of elements.They help address varied expressions and relationships among concepts in social media.
- Semantic enrichment: Named entity linking detects entities and maps each one to a contextually matching concept, often represented by a URL.For example, “JFK” can be linked to the Wikipedia concept for John F. Kennedy International Airport.
- Semantic enrichment: Semantically enriched messages support faceted search, allowing users to search and navigate by concepts as well as keywords.Adaptive faceted search frameworks enrich tweets with entities and external links to create navigable facets.
- Ontologies: Interoperability is especially important during crises because software from different agencies, developers, and vendors must communicate across coordinated response efforts.Semantic interoperability requires shared machine-understandable vocabularies and relationships.
- Ontologies: Ontology mapping connects concepts from separate ontologies when systems use different or overlapping representations of a domain.The survey identifies ontology heterogeneity as a problem for communication among ontology-based systems.
- Ontologies: Crisis ontologies cover areas such as organizations, disasters, geography, damage, processes, and resources, using formats including OWL and RDF.The surveyed ontologies are crisis-specific but can be combined with ontologies describing social-media concepts.
8. SUMMARY AND FUTURE RESEARCH DIRECTIONS
The surveyed research has matured in processing social-media data during crises, but its practical value for emergency organizations remains unestablished. Future work should deepen processing capabilities while building usable, implementable systems that support decisions and coordinated action.
- Within roughly 4 to 6 years, crisis-social-media research reached substantial maturity in filtering, classifying, processing, and aggregating data.
- Whether this research benefits the public or formal response agencies, including saving lives or property, remains unresolved.
- Social-media use remains experimental for many organizations and is not yet part of routine, day-to-day operations.
- Future research should both deepen data-processing capabilities and engage human-centered approaches to create viable systems emergency responders can implement.
- Deepening Data Processing Capabilities: Next-generation systems should progress from situational awareness toward decision support, potentially including forecasting from social-media signals.
- Deepening Data Processing Capabilities: Processing should integrate social media with SMS, traditional news, blogs, images, and credibility-verification methods.
- Beyond Data Processing: User-centered design should guide how emergency information is presented and how users interact with dashboards and visualizations.
- Beyond Data Processing: Computational methods can support communication strategies and match reported needs with available resources or intended donations to coordinate action.