mariusweiss/TaxBERT-Tax-Risk-Sentiment

---

license: mit language:

• en

TaxBERT-Risk-Sentiment-Analysis

This repository accompanies the paper: Hechtner, F., Schmidt, L., Seebeck, A., & Weiß, M. (2026). How to design and employ specialized large language models for accounting and tax research: The example of TaxBERT.

TaxBERT is a domain-adapted RoBERTa model, specifically designed to analyze qualitative corporate tax disclosures.

SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5146523 The paper provides an ‘A-to-Z’ description of how to design and employ specialized Bidirectional Encoder Representation of Transformers (BERT) models that are environmentally sustainable and practically feasible for accounting and tax researchers.

GitHub: https://github.com/TaxBERT/TaxBERT

Intended Use: This model is intended for sentence-level risk sentiment analysis in the context of qualitative corporate tax disclosures. It can be used to classify individual disclosure sentences according to whether they refer to taxes and, if so, whether they indicate tax risk, tax risk mitigation, or a tax-neutral statement. Performance outside this domain may be limited and should be validated separately before use.

Labels:

• 0: No tax reference
• 1: Tax Risk
• 2: Tax Risk Mitigation
• 3: Tax Neutral

Definition of Tax Risk: Tax Risk is defined as the risk of cash outflow due to taxes.

If the following Guide/Repository is used for academic or scientific purposes, please cite the paper:

Hechtner, F., Schmidt, L., Seebeck, A., & Weiß, M. (2026). How to design and employ specialized large language models for accounting and tax research: The example of TaxBERT.