Research roadmap

Phase 1 — Reproducible foundations

• Reproduce the original LSTM result.
• Establish classical baselines on a fixed split.
• Add data cards and event-aware splits.

Phase 2 — Modern encoders

• Benchmark DistilBERT, RoBERTa, DeBERTa and XLM-R.
• Compare full fine-tuning with LoRA/QLoRA where appropriate.
• Publish ablation and calibration studies.

Phase 3 — Multilingual crisis intelligence

• Evaluate English, Turkish, Persian, Arabic and Spanish.
• Measure cross-lingual transfer and code-switching robustness.
• Separate translated, synthetic and human-authored examples.

Phase 4 — Multimodal and evidence-aware systems

• Add CrisisMMD-compatible text-image pipelines.
• Detect duplicate and recycled imagery.
• Evaluate evidence retrieval only where retrieval is task-relevant.

Phase 5 — Human-in-the-loop operations

• Build a review queue and annotation interface.
• Measure reviewer workload, coverage and error cost.
• Add drift alerts, incident playbooks and external validation.