NARMADA: Need and Available Resource Managing Assistant for Disasters and Adversities

Kaustubh Hiware, Ritam Dutt, Sayan Sinha, Sohan Patro, Kripabandhu Ghosh, Saptarshi Ghosh

TL;DR

NARMADA is a semi-automated platform that leverages NLP and Information Retrieval to identify, extract, and match post-disaster resource needs and availabilities directly from social media data. By integrating fine-tuned BERT models with unsupervised extraction techniques, it provides an end-to-end pipeline to streamline relief coordination during crises.

What Problem Does This Paper Address?

During natural disasters, platforms like Twitter become vital lifelines for gathering situational information. Rescue workers look for "actionable" information: what resources (food, water, medical supplies, shelter) are needed and where they are available.

However, finding these needles in the haystack is incredibly difficult. Actionable tweets make up a tiny fraction of total disaster-related posts (often less than 4%). Furthermore, existing tools typically handle classification or resource matching in isolation, failing to provide an integrated, end-to-end system. They also struggle with "covert tweets"—posts that imply a need without explicitly naming a resource (e.g., "people are shivering in the cold" implies a need for shelter). NARMADA bridges these gaps by creating a unified, human-in-the-loop management platform.

Methods

NARMADA utilizes a multi-step pipeline to process microblogs:

  1. Identification: The system uses a pre-trained BERT model (bert-base-uncased) with a fully connected layer to perform a three-class classification task, filtering tweets into Need, Availability, or Other.
  2. Field Extraction: To extract specific details—Resource, Quantity, Location, Source, and Contact—NARMADA adapts an unsupervised dependency parsing approach. To handle covert tweets, they introduce a multi-label BERT classifier that categorizes the text into four broad resource buckets: Food, Health, Shelter, and Logistics.
  3. Matching: A real-time matching algorithm ranks availability posts for any given need post based on the proportion of overlapping resources.

The system architecture connects a web-based user interface (built with Angular) to a Python Flask and NodeJS backend, utilizing spaCy for NLP tasks and MongoDB for storage.

Figure 1: NARMADA’s architecture overview

Datasets

The system was evaluated using Twitter datasets from two major earthquake events:

  • 2015 Nepal Earthquake (Nepal-quake): A corpus of 50,068 tweets containing 499 needs and 1,333 availabilities.
  • 2016 Central Italy Earthquake (Italy-quake): A corpus of 70,487 tweets containing 177 needs and 233 availabilities.

The datasets consist exclusively of English tweets to reflect the primary medium used for rapid communication between international relief agencies and local populations.

Key Insights and Ideas

  • Context Matters for Covert Needs: Traditional keyword or pattern-matching methods completely miss implicit requests. Utilizing BERT’s bidirectional self-attention allows the system to classify a tweet like "villagers fear starvation" as a Food need without requiring the explicit word "food" to be present.
  • Asymmetry in Cross-Domain Generalization: A model trained on a severe disaster in a developing region (Nepal) generalizes much better to a milder disaster in a developed region (Italy) than vice versa. This happens because severe disasters generate a much wider, more varied vocabulary of resource demands.
  • The Necessity of Human-in-the-Loop: Fully automated matching can be risky in high-stakes disaster scenarios. NARMADA is intentionally designed as a semi-automated platform, providing a dashboard for human administrators to verify, modify, and finalize resource allocations.
Figure 2: Dashboard of NARMADA – (a) Navigation Buttons. (b) Needs and Availabilities List: tweets are displayed in reverse chronological order; gray tweet: already matched; black tweet: unmatched; each tweet con- tains a notch at the bottom-right corner, clicking on which reveals more details. (c) Search Box: when a query is entered, the needs and availabilities containing the query-phrase are displayed. (d) Matching List: displays the matched needs and availabilities; clicking a matching displays its resources, and gives the user an option to mark it as completed.

Key Results

  • Superior Classification: In-domain classification using BERT achieved high F1-scores (0.823 for Nepal, 0.826 for Italy), significantly outperforming standard CNN baselines.
  • Cross-Domain Robustness: In the challenging cross-domain setting (training on one earthquake and testing on the other), BERT achieved an F1@100 score of 0.234 for needs and 0.168 for availabilities when trained on Nepal and tested on Italy, doubling or tripling the performance of previous state-of-the-art CNN models.
Table 3: Performance of the neural architectures when trained on Italy-quake and tested on Nepal-quake. Best F1-scores in boldface.
  • Effective Matching: The resource matching algorithm achieved competitive F1-scores of 0.84 for the Nepal dataset and 0.87 for the Italy dataset.

Limitations and Future Work

  • Language Constraints: The current pipeline only processes English tweets, which limits its utility in non-English speaking regions unless translation layers or multilingual models are introduced.
  • Lack of Verification: The system assumes social media posts are genuine, leaving it vulnerable to misinformation or exaggerated resource counts during a crisis.
  • User Management: Currently, the platform lacks a role-based access control system (e.g., separating casual visitors, volunteers, and system admins) and does not support intra-platform communication between volunteers.

Summary updated

Abstract

Although a lot of research has been done on utilising Online Social Media during disasters, there exists no system for a specific task that is critical in a post-disaster scenario -- identifying resource-needs and resource-availabilities in the disaster-affected region, coupled with their subsequent matching. To this end, we present NARMADA, a semi-automated platform which leverages the crowd-sourced information from social media posts for assisting post-disaster relief coordination efforts. The system employs Natural Language Processing and Information Retrieval techniques for identifying resource-needs and resource-availabilities from microblogs, extracting resources from the posts, and also matching the needs to suitable availabilities. The system is thus capable of facilitating the judicious management of resources during post-disaster relief operations.
BibTeX
@inproceedings{hiware2020narmada,
  author = {Kaustubh Hiware and Ritam Dutt and Sayan Sinha and Sohan Patro and Kripabandhu Ghosh and Saptarshi Ghosh},
  title = {{NARMADA: Need and Available Resource Managing Assistant for Disasters and Adversities}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}

Loading related content...

Figure