CB-SentiLex: Bangladesh Bank Monetary Policy Stance Detection Benchmark

CB-SentiLex: Bangladesh Bank Monetary Policy Stance Detection Benchmark

Version: 1.0  
DOI: [10.5281/zenodo.21337423](doi.org)  
License: MIT  
Paper: "CB-SentiLex: An Auditable Weak-Supervision Framework for Central Bank Stance Detection with a Bangladesh Bank Benchmark"

Overview

This release contains the complete reproducibility package for CB-SentiLex, including the classified corpus, curated lexicon, trained transformer models, and three-annotator calibration set for Bangladesh Bank Monetary Policy Statements (2006–2026).

 Contents

```
CB-SentiLex-Benchmark-v1.0/
├── data/
│   ├── bb_sentences_stance.csv          # 7,432 classified sentences
│   ├── bb_document_stance.csv           # 39 document-level stance scores
│   └── calibration_300.csv              # 300-sentence, 3-annotator calibration set
├── lexicon/
│   └── dcs_lexicon.json                 # 75-term CB-SentiLex lexicon (34 hawkish, 41 dovish)
├── models/
│   └── distilbert/                      # Fine-tuned DistilBERT checkpoint
│       ├── config.json
│       ├── model.safetensors
│       ├── tokenizer.json
│       └── tokenizer_config.json
├── scripts/                             # Reproducibility scripts
│   ├── dcs_labeling.py                  # Core labeling function
│   ├── train_classifier.py              # Transformer fine-tuning
│   ├── error_analysis.py                # Error analysis
│   ├── policy_decision_validation.py    # Policy-decision correlation
│   ├── external_validation_bootstrap_cis.py  # Bootstrap confidence intervals
│   ├── falsification_tests.py           # Falsification checks
│   ├── monthly_fx_validation.py         # Monthly FX correlation
│   ├── forward_macro_validation.py      # Annual macro correlation
│   ├── temporal_stability.py            # Temporal split evaluation
│   ├── adversarial_testing.py           # Adversarial robustness
│   └── publication_figures.py           # Figure generation
├── LICENSE
├── README.md
└── citation.cff
```

 File Descriptions

Corpus

- **bb_sentences_stance.csv**: 7,432 sentence-level classifications with weak labels, neural model predictions, and document metadata. Columns: `sentence_id`, `text`, `source_document`, `period`, `language`, `doc_type`, `page`, `stance_score`, `stance_label`, `pred_hawkish`, `pred_dovish`, `pred_neutral`, `predicted_label`.

- **bb_document_stance.csv**: 39 document-level stance scores aggregated from sentence labels. Columns: `bank`, `doc_id`, `date`, `year`, `total_sentences`, `hawkish_count`, `dovish_count`, `neutral_count`, `stance_score`.

 Calibration Set

- **calibration_300.csv**: 300 sentences annotated by three independent annotators (Persons 1–3) with labels, confidence levels, notes, and consensus status. Includes weak labels for comparison. Blind two-annotator Fleiss' κ = 0.204; full three-annotator κ = 0.231.

Lexicon

- **dcs_lexicon.json**: The CB-SentiLex-Bengali lexicon with 75 terms (34 hawkish, 41 dovish) organized into four semantic families: rate keywords, inflation keywords, stance keywords, and policy keywords. Format: JSON with `hawkish` and `dovish` arrays.

 Model

- **distilbert/**: Fine-tuned `distilbert-base-uncased` checkpoint for three-class stance classification (HAWKISH, DOVISH, NEUTRAL). Trained on the weakly labeled BB corpus. Benchmark performance: ~0.87 macro-F1 on the fixed test split.

 Quick Start

```python
import pandas as pd
import json
from transformers import AutoModelForSequenceClassification, AutoTokenizer

# Load sentence-level data
df = pd.read_csv("data/bb_sentences_stance.csv")

# Load lexicon
with open("lexicon/dcs_lexicon.json") as f:
    lexicon = json.load(f)

# Load calibration set
cal = pd.read_csv("data/calibration_300.csv")

# Load trained model
model = AutoModelForSequenceClassification.from_pretrained("models/distilbert")
tokenizer = AutoTokenizer.from_pretrained("models/distilbert")

# Classify a sentence
inputs = tokenizer("BB has decided to reduce the policy rate by 50 basis points", 
                   return_tensors="pt", truncation=True, max_length=128)
outputs = model(**inputs)
predicted = ["HAWKISH", "DOVISH", "NEUTRAL"][outputs.logits.argmax(-1).item()]
```

 Citation

If you use this dataset, please cite:

```bibtex
@misc{nabil2026cbsentilex,
  title={CB-SentiLex: An Auditable Weak-Supervision Framework for Central Bank Stance Detection with a Bangladesh Bank Benchmark},
  author={Ann Naser Nabil and Umme Hafsa},
  year={2026},
  publisher={Zenodo},
  doi={10.5281/zenodo.21337423},
  url={zenodo.org
}
```

 License

MIT License. See [LICENSE](LICENSE) for details.

Contact

- Ann Naser Nabil: ann.n.nabil@gmail.com
- Umme Hafsa: hafsa.ju.2026@gmail.com