Source-linked AI summary
NLP-based Message Prioritisation for Opportunistic Routing: A Case Study of Nepal Earthquake Emergency Communication
Yihang Cao, Milena Radenkovic
TL;DR
Constrained transmission and buffer capacity can make urgent-message treatment difficult in opportunistic emergency communication, while the end-to-end effects of NLP-derived priorities require evaluation. The paper integrates humanitarian-message classifiers with priority-aware routing in a Nepal earthquake simulation. BERT improves priority classification measures over Logistic Regression, but network gains are mixed, with fewer observed P1 buffer losses under BERT and no clear latency reduction.
Problem
Limited transmission and buffer capacity complicate urgent-message treatment, and the end-to-end interaction between humanitarian classification and priority-aware routing requires investigation.
Source: §I · p. 1Method
The study maps classifier-predicted humanitarian categories to routing priorities and integrates them into a modified Binary Spray and Wait evaluation in the ONE simulator.
Source: §I · p. 1, D. Semantic Priority Mapping and… · p. 5Results
BERT improves priority accuracy and P1 recall and lowers under-priority and severe priority errors versus Logistic Regression, while P1 buffer losses fall from five to one.
Source: A. Conclusion · p. 10, E. Kathmandu Priority Routing… · p. 10Takeaways & Limitations
The findings support evaluating classification through network-level effects, with the clearest urgent-message protection difference appearing under buffer pressure.
Source: A. Conclusion · p. 10, E. Kathmandu Priority Routing… · p. 10The models use only 2015 Nepal earthquake social-media messages, and the category-to-priority mapping may affect routing outcomes.
Source: B. Research Limitations · p. 10
Abstract
from arXiv · showhide
This dissertation investigates whether Natural Language Processing (NLP) based semantic message prioritisation can improve the treatment of urgent disaster messages in opportunistic emergency communication, using the 2015 Nepal earthquake as a case study. TF IDF with Logistic Regression and BERT are used to classify social media messages into nine humanitarian categories, which are mapped to four routing priorities from P1 to P4. These priorities are integrated into the ONE simulator within a controlled Kathmandu earthquake scenario using a modified Binary Spray and Wait router. BERT achieves higher priority accuracy and P1 recall than Logistic Regression, while reducing under prioritisation and severe priority errors. However, improved NLP performance does not produce uniform network gains. P1 delivery probability remains high and similar across routing conditions, and no consistent latency reduction is observed. The clearest difference appears under buffer pressure, where the BERT condition shows fewer observed ground truth P1 buffer losses than Logistic Regression. The results indicate that semantic prioritisation is most valuable when limited transmission and storage resources are contested, although it cannot overcome mobility and contact constraints in opportunistic networks.
I. INTRODUCTION
Opportunistic networks can preserve communication when infrastructure fails, but constrained transmission and buffer capacity make urgent-message treatment difficult. The paper connects humanitarian message classification with routing priorities, addressing a gap in evidence about their end-to-end interaction.
I. INTRODUCTION
- Limited transmission opportunities and buffer capacity make it difficult to ensure urgent disaster messages receive appropriate treatment.Opportunistic networking supports communication without continuous end-to-end connectivity through store-carry-forward.Source: §I · p. 1
- Existing opportunistic routing approaches emphasize network conditions and resource availability, leaving humanitarian urgency in message content potentially unaddressed.NLP can identify humanitarian information in disaster-related text and translate message semantics into routing priorities.Source: §I · p. 1
- Classification errors may leave genuinely urgent messages with insufficient transmission-scheduling and buffer-management priority.The interaction between message classification and priority-aware routing requires further investigation.Source: §I · p. 1
A. Opportunistic Routing and Resource Management
- Epidemic routing offers delivery opportunities through high replication but uses substantial bandwidth and buffer capacity, while PRoPHET forwards using encounter history.Spray and Wait limits replication to reduce communication overhead; fixed replication may not fit varying congestion and resource conditions.Source: A. Opportunistic Routing and Resource… · pp. 1–2
B. NLP for Humanitarian Message Classification
- Statistical text features such as TF-IDF are computationally lightweight, but lexical features may miss contextual information in disaster messages.BERT provides contextual representations, while short, casual messages and imbalanced categories complicate humanitarian classification and make accuracy alone potentially misleading.Source: B. NLP for Humanitarian Message… · p. 2
C. Semantic Prioritisation and Research Gap
The paper examines how NLP-predicted humanitarian priorities affect opportunistic routing, where classification uncertainty interacts with resource allocation and network constraints. It evaluates a Nepal earthquake case study using controlled routing comparisons and an integrated NLP-to-simulation architecture.
C. Semantic Prioritisation and Research Gap
- Better classification does not necessarily improve delivery or latency because mobility, contact availability, and buffer size also affect network outcomes.The study examines predicted semantic priority against conventional routing and ground-truth priority, focusing on urgent-message delivery, latency, and buffer protection.Source: C. Semantic Prioritisation and… · p. 2
A. Nepal Earthquake Study Design
- The study maps nine humanitarian message categories to four routing priorities and evaluates standard, oracle, Logistic Regression, and BERT routing conditions.The controlled comparison uses the same message schedule, Kathmandu road map, node configuration, and mobility seeds, retaining ground-truth priorities separately to assess classification errors.Source: A. Nepal Earthquake Study Design · pp. 2–3
B. Nepal Earthquake NLP–Routing Architecture Design
- The architecture links data and NLP processing, NLP-to-simulation integration, opportunistic routing simulation, and evaluation and analysis.Predicted and ground-truth priorities are mapped separately for use in the modified routing evaluation.Source: B. Nepal Earthquake NLP–Routing… · p. 3
C. Nepal Earthquake Message Preprocessing and Classification Design
The study prepares a fixed Nepal earthquake social-media dataset for a controlled comparison of Logistic Regression and BERT humanitarian-message classifiers. It documents category imbalance, preprocessing, and classifier configurations before routing-priority evaluation.
C. Nepal Earthquake Message Preprocessing and Classification Design
- The cleaned dataset contains 3,003 English-language messages, split into 2,402 training and 601 test messages using a fixed stratified 80:20 partition.Both classifiers use the same partition for consistent comparison.Source: C. Nepal Earthquake Message… · p. 3, Table I · p. 3
- Donation and volunteering is the largest category at 887 messages, while caution and advice is least represented at 35.The imbalance may bias classifiers toward frequent categories; Logistic Regression uses balanced class weights and BERT weighted cross-entropy.Source: C. Nepal Earthquake Message… · p. 3, Figure 2 · p. 3
- Preprocessing replaces URLs and user mentions, normalises hashtags and whitespace, and removes duplicates before classification.Original and cleaned messages are retained with identifiers and ground-truth labels for traceability and error analysis.Source: C. Nepal Earthquake Message… · p. 3, Figure 3 · p. 4
- TF–IDF with Logistic Regression provides the conventional baseline, while BERT-base-uncased is fine-tuned using the same training and test partition.Both classifiers produce humanitarian categories and confidence scores that are converted into routing priorities.Source: C. Nepal Earthquake Message… · p. 4
D. Semantic Priority Mapping and NLP–Simulator Integration Design
Nine humanitarian message categories are mapped to four urgency levels, and both ground-truth and classifier-predicted priorities are carried into the simulator to distinguish classification errors from routing effects.
Semantic priority mapping
- The fixed mapping converts nine humanitarian categories into four routing priorities, from P1 (highest urgency) to P4 (lowest urgency).Injured or dead people and missing, trapped, or found people map to P1; displacement, infrastructure damage, and caution map to P2; donation and useful information map to P3; sympathy and irrelevant messages map to P4.Source: D. Semantic Priority Mapping and… · p. 4
- The same mapping function π converts ground-truth and classifier-predicted categories into separate routing priorities, with lower indices indicating greater urgency.Keeping the two priority values separate supports distinguishing classification errors from their downstream routing effects.Source: D. Semantic Priority Mapping and… · pp. 4–5
NLP–simulator integration
- A CSV message schedule stores ground-truth and predicted priorities, which the PriorityMessageEventGenerator transfers to generated messages as truePriority and predictedPriority.This separates classification from routing while enabling both priority sources to be evaluated under the same simulation conditions.Source: D. Semantic Priority Mapping and… · p. 5
- Four routing conditions use no semantic priorities, ground-truth priorities, or priorities predicted by Logistic Regression or BERT.The standard baseline isolates priority-aware routing from classification-error effects; the oracle condition provides a ground-truth reference.Source: D. Semantic Priority Mapping and… · p. 5
E. Kathmandu Earthquake Network Scenario Design
The controlled Kathmandu scenario models post-earthquake communication with intermittent peer-to-peer contacts, combining a road network, mobile response roles, and disaster-related messages without claiming to reproduce the 2015 earthquake exactly.
Scenario scope
- The scenario is a reproducible urban opportunistic-network case study, not an exact reconstruction of communication conditions during the 2015 Nepal earthquake.It combines the Kathmandu road network, disaster-related messages, and representative emergency-response roles.Source: E. Kathmandu Earthquake Network… · p. 5
Network and mobility design
- The ONE simulation uses a 3268×2482-metre Kathmandu road map and 95 road-bound nodes grouped as residents, rescue workers, and rescue vehicles.Residents source messages, rescue workers serve as destinations and mobile relays, and faster rescue vehicles provide additional forwarding opportunities.Source: E. Kathmandu Earthquake Network… · p. 5
- Different group speeds and buffer capacities create heterogeneous resources, while message exchange remains intermittent because nodes must enter wireless range.Residents have the smallest buffers; rescue vehicles have larger buffers and higher mobility.Source: E. Kathmandu Earthquake Network… · p. 5
Traffic and resource conditions
- Each 12-hour run uses a 50-metre transmission range, 50 KB/s transmission speed, 100–500 KB messages, and 30–60-second creation intervals.The high-load configuration generates 959 offered messages per run, from residents to rescue workers.Source: E. Kathmandu Earthquake Network… · pp. 5–6
- The same network configuration and message schedule are retained across routing conditions to compare priority strategies under competition for contacts and buffer space.This controlled setting examines transmission scheduling, buffer management, and urgent-message delivery.Source: E. Kathmandu Earthquake Network… · p. 6
F. Priority-Aware Routing and Buffer Management Design
The modified Binary Spray and Wait router uses semantic priority in transmission selection and buffer management while retaining its original copy-management mechanism.
Routing design
- Priority-aware routing changes transmission selection and buffer management while retaining Binary Spray and Wait copy management.These are the two routing operations through which semantic priorities influence forwarding and buffer eviction.Source: F. Priority-Aware Routing and Buffer… · p. 6
Transmission selection
- Direct delivery is attempted first; otherwise, eligible spray-phase messages are queue-ordered and then stably sorted by priority, with P1 ahead of P2, P3, and P4.Messages with equal priority retain the original queue ordering, and the router attempts transmission over available connections.Source: F. Priority-Aware Routing and Buffer… · p. 6
Buffer management
- When buffer space is insufficient, eligible non-transmitting messages at the incoming message’s priority or lower are considered for eviction before admission.If those removals cannot free enough space without violating priority constraints, the incoming message is rejected.Source: F. Priority-Aware Routing and Buffer… · p. 6
IV. EVALUATION
The evaluation compares classification and network routing under matched message and mobility conditions. It assesses humanitarian categories, mapped priorities, and network outcomes, with particular attention to ground-truth P1 messages.
A. Evaluation Setup
- Logistic Regression and BERT are tested without retraining on the same fixed test set of 601 Nepal earthquake messages at both category and routing-priority levels.The network evaluation compares Standard Binary Spray and Wait, Oracle, Logistic Regression, and BERT under the same traffic, scenario, and mobility configuration.Source: A. Evaluation Setup · p. 7, Table VII · p. 7
- Five mobility seeds are reused across all four routing conditions, yielding 20 runs and identical mobility trajectories for comparison.Network outcomes are grouped by ground-truth priority, so a true P1 message remains in the P1 group even when a classifier assigns it a lower priority.Source: A. Evaluation Setup · p. 7
B. Evaluation Metrics
- Classification is assessed with accuracy, macro-F1, and weighted-F1, while priority prediction uses priority accuracy, priority macro-F1, P1 recall, under-priority rate, and severe priority error rate.Under-priority means predicting a lower priority than ground truth; severe error means a difference of at least two levels.Source: B. Evaluation Metrics · p. 7
- Network metrics include ground-truth P1 delivery probability, latency among successfully delivered P1 messages, total buffer-capacity drops, and ground-truth P1 buffer losses.These metrics assess delivery performance and protection of urgent messages.Source: B. Evaluation Metrics · p. 7
C. Nepal Earthquake NLP Performance
BERT outperforms Logistic Regression on the nine-category humanitarian classification task and on mapped priority prediction, including P1 recall. The models’ macro F1 scores remain below their weighted F1 scores, indicating uneven classification performance.
C. Nepal Earthquake NLP Performance
- For both models, macro F1 is lower than weighted F1, indicating uneven performance across categories, with some less common or harder classes remaining difficult to classify.The paper notes that humanitarian classification is not yet perfect.Source: C. Nepal Earthquake NLP Performance · pp. 7–8
- BERT performs better than Logistic Regression across all nine humanitarian classifications, increasing accuracy from 0.6772 to 0.7188, weighted F1 from 0.6748 to 0.7096, and macro F1 from 0.5549 to 0.5838.Source: C. Nepal Earthquake NLP Performance · p. 7
- After mapping categories to four priorities, BERT leads Logistic Regression in priority accuracy (0.8070 vs. 0.7804), priority macro F1 (0.7560 vs. 0.7256), and P1 recall (0.8868 vs. 0.8302).Both classifiers achieve higher accuracy on the mapped priority task than on the original nine-category task.Source: Table IX · p. 8, C. Nepal Earthquake NLP Performance · p. 8
- BERT reduces under-prioritisation from 0.1048 to 0.0899 and severe priority errors from 0.0632 to 0.0349 versus Logistic Regression.The paper characterizes these reductions as fewer errors that may compromise urgent message handling.Source: C. Nepal Earthquake NLP Performance · p. 8
D. Nepal Earthquake Priority Error Analysis
BERT improves urgent-priority classification, but Kathmandu routing results show mixed network gains: P1 delivery remains high and latency has no consistent reduction, while P1 buffer losses are lower than with Logistic Regression.
D. Nepal Earthquake Priority Error Analysis
- For ground-truth P1 messages, BERT correctly classifies 94 of 106 versus Logistic Regression’s 88, reducing P3/P4 errors from 15 to 7 while increasing P2 errors from 3 to 5.The improvement therefore does not extend to every type of priority error.Source: Figure 5 · p. 8, D. Nepal Earthquake Priority Error… · p. 8
E. Kathmandu Priority Routing Performance
- P1 delivery probability is 0.9586 for both Standard and Oracle (810 messages), versus 0.9527 (805) for Logistic Regression and 0.9574 (809) for BERT.The comparison aggregates results across five mobility seeds.Source: Table X · p. 8, E. Kathmandu Priority Routing… · p. 8
- Oracle has the lowest average P1 latency at 1826.23 s and BERT the highest at 1842.80 s; Logistic Regression has the lowest median at 1377.9 s, while Standard is highest at 1392.6 s.No single routing condition leads on both latency metrics.Source: Figure 7 · p. 9, E. Kathmandu Priority Routing… · p. 9
- BERT has the fewest total buffer losses (65, versus Oracle’s 73 and Logistic Regression’s 74), but Oracle loses no P1 messages, compared with five for Logistic Regression and one for BERT.The buffer-loss comparison covers priority-aware routing conditions.Source: Table XI · p. 9, E. Kathmandu Priority Routing… · p. 9
- BERT’s network gains are mixed: P1 delivery probability rises from 0.9527 to 0.9574 versus Logistic Regression, but latency shows no clear reduction and P1 losses fall from five to one.The end-to-end comparison reports classification and network performance together.Source: Table XII · p. 9, E. Kathmandu Priority Routing… · p. 10
A. Conclusion
BERT improves priority classification and reduces observed P1 buffer losses relative to Logistic Regression, but these gains do not yield corresponding network-wide improvements. The findings are limited to the controlled Nepal earthquake setting, and future work proposes broader data, priority mappings, and network conditions.
A. Conclusion
- BERT improves priority accuracy and P1 recall while reducing under-priority and severe priority errors; observed P1 buffer losses fall from five to one versus Logistic Regression.The network-level gains are qualified: only a small improvement in P1 delivery probability is observed, with the same delivery latency for both models.Source: A. Conclusion · p. 10
- The study highlights the potential of combining NLP-based priority prediction with priority-aware routing and buffer handling, and recommends evaluating classifiers by network-level effects..Source: A. Conclusion · p. 10
B. Research Limitations
- Results are limited to 2015 Nepal earthquake social media, category-to-priority mappings, and a controlled Kathmandu simulation; the routing also treats different prediction uncertainties alike.Different category-to-priority mappings can lead to different routing, and performance can vary in other environments.Source: B. Research Limitations · p. 10
C. Future Work
- Future work proposes testing other disaster data, social media, and languages, evaluating priority mappings with emergency professionals, and varying network conditions and mobility traces.The authors also propose using prediction confidence in transmission scheduling and buffer control.Source: C. Future Work · p. 10