Taxon: AI Framework for Hierarchical Tax Code Prediction
A team of researchers has introduced Taxon, a new framework designed to forecast hierarchical tax codes within the e-commerce sector. This innovative system employs a feature-gating mixture-of-experts architecture to direct multi-modal features through various taxonomy levels. Additionally, it incorporates a semantic consistency model, derived from extensive language models, to ensure that product titles align with official tax definitions. To tackle noisy supervision, a multi-source training pipeline integrates curated tax databases, invoice validation logs, and merchant registrations. The findings of this research have been made available on arXiv.
Key facts
- Taxon is a framework for hierarchical tax code prediction.
- It uses a feature-gating mixture-of-experts architecture.
- A semantic consistency model is distilled from LLMs.
- The training pipeline combines multiple data sources.
- The paper is on arXiv with ID 2601.08418.
- Tax code prediction is crucial for e-commerce compliance.
- Errors in tax codes lead to financial and regulatory risks.
- The framework addresses noisy supervision in business records.
Entities
Institutions
- arXiv