CBRS: Cognitive Blood Request System with Bilingual Dataset and Dual-Layer Filtering for Multi-Platform Social Streams

1. Problem Statement

The paper addresses the critical challenge of identifying and acting upon urgent blood donation requests posted on social media (e.g., Facebook, Telegram, Discord).

• Information Overload: Emergency requests are often buried within the massive volume of daily social media communications, leading to delayed responses.
• Unstructured Data: Requests are posted in unstructured, free-form text, often containing slang, dialects, and mixed languages (Bengali, English, and Transliterated Bengali), making automated extraction difficult.
• Low-Resource Language Gap: Existing Natural Language Processing (NLP) models and datasets are heavily skewed toward high-resource languages. There is a lack of curated datasets for Bengali and its transliterated variants, which are crucial for the target demographic.
• Asymmetric Error Costs: In emergency contexts, False Negatives (missing a real request) are far more detrimental than False Positives (flagging a non-request). Traditional classifiers often fail to prioritize recall sufficiently without incurring high computational costs.
• Scalability vs. Cost: Using Large Language Models (LLMs) for every message is too expensive and slow for high-volume streams, while lightweight Machine Learning (ML) models often lack the precision to parse complex, free-form text into structured data.

2. Methodology

The authors propose CBRS, a multi-platform framework featuring a Dual-Layered Filtering (DLF) architecture and a specialized parsing pipeline.

A. Dataset Curation

• Novel Dataset: The team curated a dataset of 11,000+ parsed blood donation requests.
• Linguistic Diversity: It includes Bengali, English, and Transliterated Bengali (Romanized Bengali).
• Adversarial Negatives: To improve robustness, the dataset includes "hard negatives" (e.g., donation acknowledgments, expressions of willingness to donate) and curated adversarial examples generated using Deepseek-V3. These adversarial samples contain keywords like "blood," "urgent," and "emergency" but are not actual requests, forcing the model to learn semantic context rather than just keyword matching.

B. Dual-Layered Filtering (DLF) Architecture

To balance speed, cost, and accuracy, the system employs a two-stage filtering process:

1. Layer 1: Lightweight Classifier (TF-IDF + Asymmetric LogReg):
   • Function: Acts as a high-speed gatekeeper to filter the vast majority of non-relevant messages.
   • Technique: Uses TF-IDF vectorization with a Logistic Regression classifier.
   • Asymmetric Weighting: The loss function is weighted (α = 12) to heavily penalize False Negatives. This ensures high Recall (catching almost all real requests) even if it increases False Positives.
   • Efficiency: Extremely low inference time (~1.10 seconds for the whole batch), processing thousands of messages cheaply.
2. Layer 2: LLM-Based Verification & Parsing:
   • Function: Processes only the messages flagged as positive by Layer 1.
   • Technique: Uses GPT-4o-mini (or similar LLMs) for secondary classification and detailed parsing.
   • Cost Optimization: Since Layer 1 filters out ~95% of non-relevant traffic, the expensive LLM API calls are reduced significantly, making the system cost-effective.

C. Structured Parsing

• Goal: Convert unstructured text into a predefined JSON schema containing fields like blood_group, bags_needed, patient_info, location, hospital_name, time, and contacts.
• Model Strategy:
  • Fine-tuning: The authors fine-tuned Llama-3.2-3B using LoRA (Low-Rank Adaptation) with 4-bit quantization.
  • Performance: The fine-tuned model achieves 92% zero-shot accuracy, outperforming the base model by 41.54% and surpassing the few-shot performance of larger models like GPT-4o-mini and Gemini-2.0-flash.
  • Efficiency: The fine-tuned model requires only ~78 input tokens (no few-shot examples needed), whereas other LLMs require ~3,000 tokens (including prompts and examples), resulting in a 35× reduction in token usage.

D. System Integration

• Multi-Platform Bots: Deployed on Telegram and Discord.
• Donor Matching: Uses geo-location (Haversine distance) to match requests with registered donors nearby.
• Notification Strategy: Implements a stage-wise notification system to avoid spamming donors, stopping alerts once a donor responds.

3. Key Contributions

1. First Bilingual Blood Donation Dataset: Introduction of a 11K-sample dataset covering Bengali, English, and Transliterated Bengali with curated adversarial negatives.
2. Cost-Efficient Dual-Layer Architecture: A novel framework that combines a lightweight, asymmetrically weighted ML classifier with a powerful LLM, optimizing for both recall and operational cost.
3. High-Performance Lightweight Parser: Demonstration that a LoRA-fine-tuned 3B parameter model can outperform massive LLMs in zero-shot parsing tasks while drastically reducing token consumption and inference costs.
4. Real-World Validation: A user study involving 114 participants across 30 active groups (Telegram/Discord) validating the system's effectiveness in reducing response times and improving user satisfaction.

4. Results

• Filtering Accuracy: The DLF system achieved 99% Accuracy, Precision, and Recall, outperforming standard baselines (SVM, Random Forest, DistilBERT) while maintaining the fastest inference time.
• Parsing Performance:
  • The LoRA-finetuned Llama-3.2-3B achieved 92% accuracy in zero-shot parsing.
  • It surpassed GPT-4o-mini and Gemini-2.0-flash in few-shot settings and significantly outperformed the base Llama-3.2-3B.
• Cost Efficiency:
  • The dual-layer approach reduced operational costs by 33% to 47% compared to using an LLM for all messages.
  • Input token usage was reduced by 35× compared to standard few-shot prompting with other LLMs.
• User Study Findings:
  • Response Time: Parsing and retrieval times were reduced to an average of 4 seconds and 5 seconds respectively.
  • Satisfaction: 61% of users rated the functionality as "Above Average," and 44% found the slash-command interface "Very easy" to use.
  • Correlation: Strong positive correlations were found between Command Usability/Intuitiveness and User Satisfaction.

5. Significance

• Scalable Humanitarian Aid: CBRS provides a scalable, low-cost solution for extracting critical health information from social media, directly applicable to emergency response in low-resource settings.
• Advancing Low-Resource NLP: The work demonstrates that with careful dataset curation and parameter-efficient fine-tuning (LoRA), small models can outperform massive proprietary models in specific, high-stakes domains, democratizing access to advanced AI tools.
• Asymmetric Learning: The paper highlights the importance of asymmetric loss functions in emergency detection, prioritizing the saving of lives (recall) over computational efficiency in the initial filtering stage.
• Open Science: The authors have made the dataset, code, and trained models publicly available, fostering further research in crisis informatics and multilingual NLP.

In conclusion, CBRS successfully bridges the gap between the chaotic nature of social media emergency posts and the structured data requirements of blood donation logistics, offering a robust, cost-effective, and linguistically inclusive solution.