[Paper Review] Civique: Using Social Media to Detect Urban Emergencies

Diptesh Kanojia, Vishwajeet Kumar+1 more|arXiv (Cornell University)|Oct 14, 2016

Public Relations and Crisis Communication7 references3 citations

TL;DR

Civique is a real-time system that detects urban emergencies—such as fires, accidents, and earthquakes—by analyzing geotagged Twitter posts using natural language processing and machine learning. It classifies tweets into emergency categories with over 90% F-score and visualizes incidents on a map while notifying users and authorities via web and Android interfaces.

ABSTRACT

We present the Civique system for emergency detection in urban areas by monitoring micro blogs like Tweets. The system detects emergency related events, and classifies them into appropriate categories like "fire", "accident", "earthquake", etc. We demonstrate our ideas by classifying Twitter posts in real time, visualizing the ongoing event on a map interface and alerting users with options to contact relevant authorities, both online and offline. We evaluate our classifiers for both the steps, i.e., emergency detection and categorization, and obtain F-scores exceeding 70% and 90%, respectively. We demonstrate Civique using a web interface and on an Android application, in realtime, and show its use for both tweet detection and visualization.

Motivation & Objective

• To develop a real-time system that detects urban emergencies from social media, particularly Twitter, to support faster emergency response.
• To classify geotagged tweets into specific emergency categories such as fire, accident, or earthquake using NLP and machine learning.
• To visualize detected emergencies on interactive maps and notify users and authorities through web and mobile interfaces.
• To address challenges such as informal language, false alarms, and credibility of users in emergency detection from social media.
• To provide both online and offline contact options for civic authorities to ensure response continuity during network disruptions.

Proposed method

• The system ingests real-time Twitter streams using the Twitter Streaming API, filtering tweets based on a predefined list of keywords related to emergencies.
• A text-cleaning pipeline removes noise (e.g., hashtags, URLs, excessive punctuation) and normalizes misspellings and slang using rule-based and NLP techniques.
• The system employs a two-stage classification model: first detecting if a tweet is emergency-related, then categorizing it into specific types like fire, accident, or earthquake.
• For classification, the system uses feature engineering including unigrams, bigrams, trigrams, POS tags, and named entity recognition, with Naive Bayes and SVM models evaluated.
• Spatio-temporal data (location and time) are extracted from cleaned tweets and visualized on a Google Maps interface for real-time situational awareness.
• Users receive notifications via web and Android apps, with embedded links to online contacts and offline numbers for emergency authorities, based on the detected event type.

Experimental results

Research questions

• RQ1 Can real-time analysis of geotagged Twitter posts effectively detect urban emergencies such as fires, accidents, and earthquakes?
• RQ2 How accurately can a two-stage machine learning pipeline classify emergency-related tweets and assign them to specific categories?
• RQ3 To what extent can NLP techniques handle informal language, misspellings, and slang in social media posts for reliable emergency detection?
• RQ4 How effective is the system in reducing false positives and ensuring credible alerts through user and content verification?
• RQ5 Can a combined web and mobile interface enable timely user notification and facilitate contact with relevant emergency authorities?

Key findings

• The emergency detection classifier achieved an F-score exceeding 70%, demonstrating strong performance in identifying emergency-related tweets.
• The categorization classifier achieved an F-score above 90%, indicating high accuracy in assigning detected emergencies to specific types such as fire or accident.
• The system successfully visualized real-time emergency events on a map interface, enabling spatial awareness of ongoing incidents.
• Notifications were generated and delivered via both web and Android applications, including online and offline contact details for relevant authorities.
• The word cloud analysis confirmed that the model was trained on semantically relevant emergency keywords such as 'fire', 'earthquake', 'accident', and 'respond'.
• The system demonstrated robustness in handling informal language through a combination of text-cleaning and machine learning, reducing noise without sacrificing detection accuracy.