Amandhi17/Asymmetric-Evidence-Alignment-Network

Asymmetric Evidence Alignment for Cross-Channel, Cross-Lingual Incident Coreference in Low-Resource Disaster Response # C4 — Multi-Channel, Cross-Lingual Incident Coreference (AEAN) Working repository for Component 4. Build order follows the implementation roadmap. ## Setup ```bash python -m venv .venv .venv/Scripts/python -m pip install -r requirements.txt # Windows ``` ## Regenerating the excluded artifacts Three artifacts are too large for GitHub (1.2-2.4 GB each) and are gitignored. Everything needed to rebuild them IS committed: | Artifact | Rebuild with | Cost | |---|---|---| | `.venv/` | `pip install -r requirements.txt` | minutes | | `aean-labse-ft-best/` | `aean/finetune.py` via `AEAN_Step4_Colab.ipynb` | ~35 min, free Colab T4 | | `results/tok_ft/` | `python -m aean.encode` | ~40 min CPU | | `results/emb_*.npy` | `python -m aean.baselines ` | ~80 min CPU (first run only) | Seeds are fixed throughout, so a rebuild reproduces the reported numbers. The result tables in `results/*.csv` ARE committed -- those are the thesis output. ## Status Build order follows `AEAN_Full_Implementation_Roadmap` (10 steps, two-level contribution: domain-adapted encoder + AEAN architecture). | Step | State | |------|-------| | 1 — environment + tokenising loader | **done** — `aean/data.py` | | 2 — dataset audit | **done** — `audit.py`, 13/13 PASS on v7 | | 3 — baselines + RQ1 diagnostic | **done (CPU part)** — `aean/baselines.py`; trained baselines need a GPU | | 4 — fine-tune encoder (Stage A) — *contribution #1* | **done** — `aean/finetune.py`, 0.573 -> 0.905 | | 5 — AEAN on frozen encoder (Stage B) — *contribution #2* | next | | 6 — cost-sensitive decision + abstention | not started | | 7 — joint fine-tune (Stage C, optional) | not started | | 8–10 — evaluation, interpretability, integration | not started | **Staging rule:** never train encoder and AEAN together first. Stage A alone is a self-contained result ("domain-adapted beats zero-shot") and is the safety net if AEAN runs late. The 2×2 — {zero-shot, fine-tuned} × {cosine, AEAN} — is the headline table. ## Data **`SL-D …