Client
United Nations Office for the Coordination of Humanitarian Affairs (OCHA) , United Nations Women (UNW), and the Accountability to Affected Populations Task Force (AAP Task Force)
Context
Amidst the ongoing war in Gaza, reliable reporting on humanitarian conditions became increasingly difficult due to limited media access, restricted movement, and intermittent communication blackouts. In this environment, publicly available posts from local residents, journalists, and institutions emerged as critical sources of information on urgent needs. UN agencies required a structured system to extract, categorize, and track this data consistently.
Our Approach
Social Studio Analytics developed a full pipeline to monitor and classify humanitarian needs content in near real time, with a focus on Gaza and supplementary coverage of the West Bank. The system included:
- Source selection: Curated hundreds of public Arabic-language pages and personal accounts with relevance to the topic and local presence.
- Data scraping: Deployed automated scraping tools to collect public posts from Facebook and Telegram on a daily basis.
- Cleaning & filtering: Removed irrelevant content, spam, and reposts; retained only location-based, first-hand, or verified second-hand humanitarian content.
- Large Language Model (LLM) -based classification: Applied a language model to:
- Tag posts by humanitarian theme (e.g. Aid, Needs, Protection, Gender Based Violence)
- Identify both general and exact needs (e.g. food insecurity → infant formula shortage)
- Flag mentions of protected groups (children, older persons, people with disabilities, female-headed households)
Technical Approach and Workflow
- Data Ingestion & Cleaning
Collected structured and semi-structured data from Facebook and Telegram, removed duplicates, normalized timestamps, and filtered out irrelevant entries based on content type and priority level. - Schema Unification & Feature Engineering
Standardized heterogeneous data sources into a unified schema, generated synthetic identifiers (e.g.URLs), mapped metadata (classification, affiliation, gender), and addressed missing values using appropriate imputation techniques. - LLM-Powered Annotation
Applied NLP-based batch processing to extract structured labels (location, sentiment, themes, needs, etc.) using Arabic-language content, with chunk-based execution for stability. - Merging & Export
Consolidated all processed outputs using indexed joins, finalized a clean, and exported to CSV/XLSX for downstream use.
Outputs & Delivery
Results were delivered through a combination of weekly reporting and live dashboards:
- Interactive dashboard with filters by theme, geography, date, and population group
- Comparative analytics showing weekly trend shifts in needs expression
- Sample post library with original Arabic text and English metadata
- Archived datasets organized by week, theme, and location
- Alert system -daily report
- AdHoc report inferred from the trends
- Stakeholders Mapping
This system provided field teams, analysts, and planners with structured, high-volume insight directly from affected populations, overcoming the gaps left by traditional reporting mechanisms.
Demo
[Explore the Dashboard Sample] (link placeholder)
