💻 computer science

CB-SentiLex: An Auditable Weak-Supervision Framework for Central Bank Stance Detection with a Bangladesh Bank Benchmark

This paper introduces CB-SentiLex, an auditable weak-supervision framework and the first reproducible NLP benchmark for a South Asian central bank, which successfully generates stance labels for a Bangladesh Bank corpus and demonstrates strong model performance alongside significant correlations with monetary policy directions, despite challenges from temporal concept drift.

Original authors: Ann Naser Nabil, Umme Hafsa

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ann Naser Nabil, Umme Hafsa

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of economics as a giant, bustling control room where the "Central Bank" acts like the captain of a massive ship. Their job is to steer the economy through stormy waters of inflation and calm seas of growth. To do this, they don't just shout orders; they send out carefully worded messages called "Monetary Policy Statements." These messages are the ship's radar, telling everyone on board (investors, businesses, and regular people) whether the captain plans to speed up, slow down, or just hold steady. For decades, scientists have tried to use computers to read these messages and figure out the captain's mood. This field, called "Natural Language Processing" (NLP), is like teaching a robot to understand human nuance. But until now, most of these robots have only been trained on the captains of the biggest, richest ships (like the US Federal Reserve), leaving the captains of smaller, emerging economies in the dark. This paper asks a simple but tricky question: Can we build a robot that understands the specific, cautious language of a smaller bank, like the Bangladesh Bank, without needing a team of expensive human experts to read every single sentence?

The authors of this study, Ann Naser Nabil and Umme Hafsa, decided to build a new kind of "mood detector" for the Bangladesh Bank. They realized that hiring experts to manually label thousands of sentences as "hawkish" (wanting to raise rates), "dovish" (wanting to lower rates), or "neutral" would be too slow and expensive. Instead, they created a clever shortcut called CB-SentiLex. Think of this as a "keyword flashlight." They hand-picked a list of 75 specific words and phrases—like "tightening," "inflationary pressure," or "accommodative"—that usually signal a specific mood. The computer scans the bank's statements, counts these keywords, and makes a quick guess at the stance. It's like a teacher grading a test by just checking if the student used the right vocabulary words, rather than reading the whole essay for deep meaning.

To make sure their "keyword flashlight" wasn't just guessing wildly, they ran a reality check. They took 300 sentences and had three human experts read them blindly to see if they agreed with the computer's guesses. The humans didn't always agree with each other (which is normal, as the bank's language is often intentionally vague), but they did agree enough to show the computer's method was a solid starting point. The computer then used these "weak" labels to train a smart AI model (a type of digital brain called a Transformer). The results were promising: the AI learned to mimic the keyword system with high accuracy, reaching a score of about 0.87 on a scale where higher is better.

However, the paper is very honest about the limits of this tool. When they tested the AI on new, future data, its performance dropped, suggesting that the bank's language changes over time (a phenomenon called "concept drift"). Furthermore, when they compared the AI's "mood readings" to the actual interest rate decisions the bank made, they found a clear link: when the AI said the bank was "hawkish," the bank was indeed likely to raise rates in that same quarter. But the AI couldn't perfectly predict the future; it was better at describing what the bank was thinking now than guessing what they would do next. The study also found that the AI struggled most with "dovish" (easing) language, often missing subtle hints that the bank was trying to be gentle, instead labeling those sentences as "neutral."

In the end, this paper doesn't claim to have built a crystal ball that predicts the economy. Instead, it offers a transparent, reproducible, and low-cost toolkit for understanding the Bangladesh Bank's communication. It proves that you don't need a massive budget or a team of PhDs to start analyzing central bank language in South Asia. By releasing their data, their keyword list, and their trained models to the public, the authors have handed the keys to a new kind of economic radar, allowing anyone to listen more closely to the whispers of the central bank, even if the robot still needs a little help understanding the most subtle hints.

Technical Summary: CB-SentiLex for Central Bank Stance Detection

Problem Statement
Natural Language Processing (NLP) research on central bank communication has historically concentrated on advanced-economy institutions, specifically the Federal Reserve, the European Central Bank, and the Bank of England. This geographic bias leaves a significant gap in understanding the monetary policy stances of South Asian central banks, which operate under different institutional constraints and communication norms. Specifically, there are no existing NLP resources for the Bangladesh Bank (BB), despite its role in managing the monetary policy of a large emerging economy. Furthermore, the creation of high-quality, manually annotated corpora for central bank text is often prohibitively expensive and requires scarce domain expertise, limiting the scalability of stance detection in resource-constrained settings.

Methodology
The paper introduces CB-SentiLex, an auditable weak-supervision framework designed to generate reproducible stance labels without requiring corpus-wide expert annotation. The methodology proceeds through four main stages:

Corpus Construction: The authors compiled a dataset of 39 Monetary Policy Statements (MPS) from the Bangladesh Bank, spanning 2006 to 2026. This corpus covers 8,647 preprocessed sentences and captures two distinct policy regimes: monetary targeting (pre-2017) and inflation targeting (post-2017).
Weak Labeling (CB-SentiLex): A domain-specific lexicon containing 75 terms (34 hawkish, 41 dovish) was curated from central bank glossaries and manual inspection of BB, Fed, and RBI statements. Sentences are labeled via a transparent majority-vote rule based on keyword counts: if hawkish matches exceed dovish matches, the label is HAWKISH; if dovish exceeds hawkish, it is DOVISH; otherwise, NEUTRAL. This approach prioritizes reproducibility and transparency over the nuance of manual annotation.
Human Calibration and Audit: To assess the quality of the weak labels, a 300-sentence calibration set was created, sampled across four time periods. Three annotators labeled these sentences independently. Two annotators worked blindly (without seeing the weak labels), while a third acted as a guided validator. Inter-annotator agreement was measured using Fleiss' kappa.
Model Training and Validation: Transformer models (DistilBERT, RoBERTa, and FinBERT) were fine-tuned on the weakly labeled corpus. The primary evaluation metric was macro-F1, chosen to address the severe class imbalance (approx. 89% NEUTRAL). External validation was performed by correlating the derived stance index with BB repo-rate changes, macroeconomic indicators (inflation, exchange rates), and through adversarial testing.

Key Contributions
The paper makes four specific contributions to the field:

First BB Stance Corpus: The release of the first large-scale, reproducible NLP benchmark for a South Asian central bank, comprising 39 documents and 8,647 sentences.
Auditable Weak-Supervision Framework: The introduction of CB-SentiLex, which couples a transparent 75-term lexicon with a human-calibration audit protocol. This allows for stance measurement without the high cost of full manual annotation.
Benchmarking and Reproducibility: The release of all artifacts, including the corpus, the curated lexicon, trained models, and the 300-sentence calibration set. The study benchmarks DistilBERT and RoBERTa against the weak labels, demonstrating that the labeling scheme is learnable despite class imbalance.
Empirical Validation: The paper provides a multi-dimensional validation of the stance index, including external correlation with policy rate directions and an analysis of concept drift across temporal splits.

Results

Model Performance: On the fixed benchmark split, DistilBERT and RoBERTa achieved a macro-F1 score of approximately 0.870, indicating strong agreement with the weak labels. FinBERT achieved 0.815 ± 0.020 under 5-fold cross-validation.
Label Quality: The blind inter-annotator agreement (Fleiss' $\kappa$ ) between the two independent human annotators was 0.204, confirming genuine task ambiguity in central bank text. However, the consensus of the human panel vs. the weak labels yielded a macro-F1 of 0.664, suggesting the lexical labels are informative but conservative in borderline cases.
External Validation: The stance index showed a statistically significant correlation with same-quarter repo-rate direction ( $\rho = 0.406, p = 0.011$ ). However, correlations with next-quarter rates, annual macroeconomic outcomes, and monthly exchange rates were either weak or statistically insignificant.
Robustness and Drift: A temporal split evaluation (training on 2006–2015, testing on 2016–2026) revealed significant concept drift, with macro-F1 dropping from 0.870 to 0.603.
Error Analysis: The dominant error pattern was the misclassification of DOVISH sentences as NEUTRAL (21.11% error rate), particularly when easing language was indirect or hedged. Conversely, NEUTRAL sentences containing inflation-adjacent language were frequently misclassified as HAWKISH.

Significance and Claims
The paper positions CB-SentiLex not as a perfect policy-decision predictor, but as a communication-measurement tool. The authors claim that their work demonstrates the feasibility of constructing reproducible stance indices for emerging-market central banks using low-cost, weak-supervision methods.

The study highlights that while transformer models can learn the internal coherence of the lexical labeling scheme, the approach is limited by the inherent ambiguity of central bank language and the difficulty of capturing subtle, hedged policy signals. The authors emphasize that the framework provides a scalable path for researchers to analyze monetary policy in underrepresented regions, offering a starting point for financial NLP in South Asia that can be refined through iterative human annotation. The work explicitly avoids claiming that textual stance can directly predict market outcomes or future rate moves with high accuracy, instead framing the index as a useful metric for understanding expectation management and communication effectiveness.

Technical Summary: CB-SentiLex for Central Bank Stance Detection

More like this