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

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

TL;DR

This paper introduces NARMADA, a semi-automated system that uses social media (specifically tweets) to identify and match resource needs with available resources during disaster relief operations. It tackles the critical challenge of efficiently connecting those who need help with those who can provide it, a crucial step often hampered by information overload.

What Problem Does This Paper Address?

In the chaotic aftermath of a disaster, effectively coordinating relief efforts is paramount. While social media platforms like Twitter are invaluable for gathering real-time information, they also present a massive challenge: sifting through an overwhelming volume of posts to identify critical needs (e.g., "need food") and available resources (e.g., "have water to share"). Existing systems often focus on either identifying these needs/availabilities or matching them, but rarely integrate both. This gap leaves a critical bottleneck in disaster response, delaying the delivery of essential aid. NARMADA aims to bridge this gap by providing a unified platform for automated identification and matching of these vital pieces of information.

Methods

NARMADA employs a multi-stage approach leveraging Natural Language Processing (NLP) and Information Retrieval (IR) techniques:

  • Pre-processing: Tweets undergo standard cleaning, removing URLs, mentions, special characters, and segmenting compound words. Crucially, case-folding and stemming are avoided to preserve proper nouns for later analysis.
  • Identifying Needs and Availabilities: The system uses a supervised, three-class classification approach (need, availability, or other) to categorize tweets. They experiment with Convolutional Neural Networks (CNNs) as a baseline and propose a BERT-based model. The BERT model, pre-trained and fine-tuned, significantly outperforms CNNs, especially in identifying less frequent "need" and "availability" tweets.
  • Extracting Relevant Fields: To make the identified needs and availabilities actionable, the system extracts five key fields: resource, quantity, location, source, and contact. This is achieved using a combination of unsupervised techniques, including dependency parsing and semantic similarity checks against a pre-compiled list of disaster-related resources. A notable adaptation is the use of the BERT model for multi-label classification to handle "covert tweets" where the resource isn't explicitly stated but implied.
  • Matching Needs and Availabilities: A fast, real-time algorithm matches needs to availabilities based on the proportion of common resources identified in the extracted fields. Tweets are ranked by the fraction of matching resources.

Datasets

The research utilizes a dataset of tweets collected from two major disaster events:

  • Nepal-quake (April 2015): Tweets related to the Nepal earthquake.
  • Italy-quake (August 2016): Tweets related to the Central Italy earthquake.

These datasets were collected using the Twitter Search API. The paper notes that the Nepal-quake dataset is richer in mentions of varied resources, likely due to the severity and developing context of the disaster compared to the Italy-quake.

Key Insights and Ideas

  • BERT's Superiority in Disaster Tweet Classification: The paper demonstrates that pre-trained BERT models, when fine-tuned, significantly outperform traditional CNN models for classifying disaster-related tweets into "need," "availability," and "other" categories. This highlights BERT's ability to capture nuanced contextual information crucial for this task.
  • The Challenge of "Covert Tweets": A significant insight is the existence of "covert tweets" where the need or availability is implied rather than explicitly stated. The adaptation of the BERT model for multi-label classification to handle these tweets is a key innovation.
  • Importance of Recall for Disaster Response: The authors emphasize that optimizing for recall (identifying all needs and availabilities) is crucial, especially in sparser datasets like the Italy-quake. This is because missing even a single critical need can have severe consequences.
  • Unsupervised Extraction for Scalability: The reliance on unsupervised methods for extracting specific fields (resource, location, etc.) is a pragmatic choice, acknowledging the difficulty of obtaining large amounts of labeled data in disaster scenarios and enabling generalizability.

Key Results

  • In-domain Classification: The proposed BERT model achieved the best F1-scores for in-domain classification of tweets into needs, availabilities, and others, outperforming CNN-based baselines. For Nepal-quake, the BERT model achieved an F1-score of 0.828, and for Italy-quake, 0.826.
  • Cross-domain Classification: The BERT model also showed strong performance in cross-domain classification, demonstrating its robustness. For instance, when trained on Italy-quake and tested on Nepal-quake, it achieved an F1-score of 0.110.
  • Information Extraction: The system achieved high F1-scores for extracting key fields, with scores ranging from 0.58 (contact) to 1.00 (resource) for need-tweets and 0.65 (contact) to 1.00 (resource) for availability-tweets.
  • Matching Performance: The proposed matching methodology achieved competitive F1-scores, with 0.84 for Nepal-quake and 0.87 for Italy-quake, outperforming several baseline methods.

Limitations and Future Work

  • Limited to English Tweets: The current system is primarily designed for English tweets, limiting its applicability in regions where local languages are prevalent.
  • Semi-Automated Nature: While automated for core tasks, the system is described as semi-automated, implying human oversight or intervention might still be necessary.
  • Future Work:
    • Login System and Access Levels: Implementing a login system to differentiate user roles (e.g., visitor, volunteer, administrator) is planned.
  • Inter-Volunteer Communication: Enabling communication between volunteers within the platform to discuss resources.
  • Vernacular Language Support: Expanding support to include vernacular languages, which is a significant undertaking requiring appropriate tools.
  • Geographical Visualization: The discussion mentions marking needs and availabilities on a map, suggesting future integration of geospatial visualization.
  • Handling Misinformation: Acknowledging the prevalence of misinformation on Twitter, future work could involve mechanisms to verify the authenticity of posted needs and availabilities.

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 Kripa 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},
}
Figure