NLP course project — University at Buffalo, 2022.
Team Tesla: Jayesh Suryavanshi · Sumeet Aher · Krati Sharma
Course taught by Dr. Rohini K. Srihari (TAs: Sougata Saha, Souvik Das). Full write-up: docs/report.pdf.
🔭 Revived in 2026 → social-unrest-forecasting — a from-scratch rebuild as a rigorous out-of-time forecasting study on current ACLED/GDELT data. It replaces this project's leaky "prediction" step with a proper leakage-free backtest, extends the test across 8 countries and deep GKG-theme text, and includes an adversarial-review pass. Honest headline finding: news text adds no forecasting skill beyond the event record.
The project builds an early-warning pipeline around ACLED (Armed Conflict Location & Event Data): extract structure from conflict-event descriptions, generate ACLED-style event summaries, and predict upcoming unrest events from news streams.
Built in spring 2022 — before ChatGPT — with the then-standard NLP toolkit: TF-IDF + classical scikit-learn models, GloVe embeddings + a Keras LSTM, and a fine-tuned T5 transformer. The notebooks are preserved as a faithful record of that work; see Reproducibility and the provenance note at the top of each notebook for what was (and was not) touched in the 2026 cleanup.
| What it does | Approach | Where | |
|---|---|---|---|
| Task 1 — Information extraction | Classify an ACLED event description (NOTES) into one of 6 event types (Battles, Explosions/Remote violence, Protests, Riots, Strategic developments, Violence against civilians) |
TF-IDF (1–2 grams) → LinearSVC / Logistic Regression / Multinomial NB / Random Forest | notebooks/task1_event_classification/ |
| Task 2 — Event summarization | Generate an ACLED-style narrative summary from structured event fields (date, actors, location, fatalities, …) | T5-base fine-tuned on ~90k ACLED rows (structured-to-text) | notebooks/task2_event_summarization/ |
| Task 3 — Event prediction | Predict whether an unrest event will occur, from news text | M2: GloVe + LSTM over daily event sequences · M3: Bing News ↔ ACLED matching + SGD/LogReg/MNB classifiers | notebooks/task3_event_prediction/ |
The course ran in three milestones. Not everything described in the report has code in this repository — this matrix is the honest inventory:
| Milestone 1 | Milestone 2 | Milestone 3 | |
|---|---|---|---|
| Task 1 | report only | ✅ Task1Classifier.ipynb |
same notebook (M2/M3 submissions were byte-identical; the report's BERT tagging upgrade was never committed) |
| Task 2 | report only | — (never implemented) | ✅ Tesla_SUP_Subtask2_2.ipynb (T5) |
| Task 3 | report only | ✅ Task3_milestone2.ipynb (LSTM) |
✅ 3-notebook news pipeline |
├── docs/report.pdf # final project report
├── data/README.md # how to obtain the datasets (not versioned)
└── notebooks/
├── task1_event_classification/
│ └── Task1Classifier.ipynb # TF-IDF + 4 classical models, confusion-matrix analysis
├── task2_event_summarization/
│ └── Tesla_SUP_Subtask2_2.ipynb # T5-base fine-tuning + ROUGE evaluation
└── task3_event_prediction/
├── milestone2_lstm/
│ └── Task3_milestone2.ipynb # GloVe + spaCy day-vectors → LSTM
└── milestone3_news_pipeline/
├── DownloadBingNews.ipynb # 1) pull news via Bing News Search v7 (API retired 2025)
├── Labelling_Event_with_ACLED.ipynb # 2) label news by cosine-matching ACLED events
└── Training_and_Prediction.ipynb # 3) train/evaluate event-prediction classifiers
Each notebooks/task*/ directory also carries its own README with run instructions and per-task caveats.
All numbers below are from the notebooks' saved outputs and match the results presented in the 2022 project report (see each notebook for the full classification reports, and its provenance cell for caveats about where its outputs came from).
Task 1 — event-type classification (5-fold CV accuracy; run on a ~100-row sample of the training set, hence the modest numbers — the report presents this same run, with Multinomial NB scoring highest):
| Model | CV accuracy |
|---|---|
| Logistic Regression | 0.59 |
| Multinomial Naive Bayes | 0.59 |
| LinearSVC | 0.57 |
| Random Forest | 0.50 |
Task 2 — T5 summarization (1-epoch fine-tune; first 100 validation examples):
| Metric | F1 | Precision | Recall |
|---|---|---|---|
| ROUGE-1 | 0.375 | 0.648 | 0.272 |
| ROUGE-2 | 0.184 | 0.351 | 0.129 |
| ROUGE-L | 0.359 | 0.619 | 0.261 |
Task 3 — event prediction, Milestone 3 pipeline (1,065 news articles labeled by ACLED matching; 10% held-out test set, n=107):
⚠️ Read this before the numbers below. These labels were generated by cosine-matching each news article to an ACLED event, and the classifier is then trained on the same news text to predict those derived labels. That is circular — the 0.85 is a label-leakage artifact, not out-of-sample forecasting skill. The 2026 rebuild (social-unrest-forecasting) redoes this as a proper leakage-free, out-of-time backtest and finds that news text adds no forecasting skill beyond the event record. Treat the table below as a historical record of the coursework pipeline, not as an evaluation of predictive performance.
| Model | Accuracy | F1 (weighted) |
|---|---|---|
| SGDClassifier (hinge) | 0.85 | 0.85 |
| Logistic Regression | 0.85 | 0.85 |
| Multinomial Naive Bayes | 0.83 | 0.83 |
Among matched news–event pairs, the news article preceded its ACLED event by ≈5 days on average — but per the caveat above, this reflects the news-to-event matching window, not a validated predictive lead time.
(Task 3 Milestone 2, the LSTM approach: training/evaluation outputs were not captured in the saved run, so no metrics survive; the report discusses it qualitatively.)
None of the datasets are versioned here (ACLED's terms don't permit redistribution, and the files total ~130 MB). data/README.md lists every file the notebooks expect and how to obtain it:
- the seven course-provided ACLED-derived TSVs (train/valid/test per task),
- an ACLED export (Jan–Feb 2019) for the news-labeling step,
- GloVe 300-d embeddings,
- the Bing News CSVs — these can no longer be regenerated: Microsoft retired the Bing News Search v7 API in August 2025.
These notebooks are a historical record, not a maintained pipeline. They were written for Google Colab (GPU for Tasks 2/3) and still contain /content/... and Google Drive paths. requirements.txt documents the stack; re-execution end-to-end is not guaranteed and was deliberately not attempted during cleanup — no saved output was produced after 2022, and nothing was re-executed. (For two of the Milestone 3 pipeline notebooks the saved outputs demonstrably come from an even older, pre-2022 environment — see their provenance cells and Known gaps below.)
- Milestone 1 produced reports only; no code exists for it.
- Task 2 had no Milestone 2 implementation, and the report's BERT-based tagging upgrade for Task 1 Milestone 3 was never committed.
- The Task 2 fine-tuned T5 model was never saved, so its ROUGE numbers can't be reproduced without retraining.
- Task 1's saved metrics come from a ~100-row sample; Task 2's ROUGE from 100 validation examples.
- The Task 3 M2 LSTM run had a GloVe loading error (traceback preserved in the notebook) and its metrics were never captured.
- The Task 3 M3 pipeline's saved outputs come from a pre-2022 environment (macOS Anaconda, Python 3.5, scikit-learn ≤ 0.19) matching no team member, which suggests parts of that pipeline were adapted from earlier third-party code with its outputs retained; the numbers do match the 2022 report's screenshots. Details in those notebooks' provenance cells.
- Several small 2022 bugs were found and fixed during the 2026 cleanup (a validation-features copy-paste bug, a page-fetch loop-indentation bug, a removed-in-sklearn-0.20 import, a missing seaborn import). Each notebook's provenance cell states exactly what changed; all saved outputs predate the fixes.
2026 cleanup. The repo was restructured from the original Team Tesla/codebase/Milestone N/Task M/ layout into the task-centric tree above; a byte-identical duplicate notebook was deduplicated, 1.7 MB of junk cell output was stripped, the bugs above were fixed, and this README replaced a generic placeholder. The full original layout is preserved in git history.
Security disclosure. The 2022 code hardcoded an Azure Bing News Search v7 subscription key, and ~135 MB of course-provided ACLED-derived TSVs were committed early on. In the 2026 cleanup the git history was rewritten to purge both (plus a zip that duplicated the key): the key string is redacted everywhere, and the datasets are gone from history in line with ACLED's redistribution terms. The key was inert regardless — Microsoft retired the Bing Search APIs in August 2025. If you hold an old clone or a pre-2026 commit link, it will not match the rewritten history.
MIT. ACLED data is subject to ACLED's terms of use and is not distributed with this repository.