Taxon: Hierarchical Tax Code Prediction with Semantically Aligned LLM Expert Guidance

TL;DR

Taxon tackles the problem of hierarchical tax code prediction for large-scale e-commerce by integrating a hierarchical feature-gating mixture-of-experts (MoE) with an LLM-guided semantic consistency module. The framework is trained via a multi-source pipeline that combines structural supervision across a ten-level taxonomy with semantic alignment signals distilled from domain-expert LLM judgments, and it is reinforced by a path reconstruction technique (RePath) to enforce hierarchical validity. Extensive experiments on proprietary TaxCode data and the public WOSE dataset show state-of-the-art performance, with ablations highlighting the contributions of the MoE architecture, category-name fusion, hierarchical supervision, and semantic consistency. Taxon is deployed in Alibaba’s tax service system, handling hundreds of thousands of daily queries with high accuracy and robustness, demonstrating practical impact for automated invoicing and compliance at scale.

Abstract

Tax code prediction is a crucial yet underexplored task in automating invoicing and compliance management for large-scale e-commerce platforms. Each product must be accurately mapped to a node within a multi-level taxonomic hierarchy defined by national standards, where errors lead to financial inconsistencies and regulatory risks. This paper presents Taxon, a semantically aligned and expert-guided framework for hierarchical tax code prediction. Taxon integrates (i) a feature-gating mixture-of-experts architecture that adaptively routes multi-modal features across taxonomy levels, and (ii) a semantic consistency model distilled from large language models acting as domain experts to verify alignment between product titles and official tax definitions. To address noisy supervision in real business records, we design a multi-source training pipeline that combines curated tax databases, invoice validation logs, and merchant registration data to provide both structural and semantic supervision. Extensive experiments on the proprietary TaxCode dataset and public benchmarks demonstrate that Taxon achieves state-of-the-art performance, outperforming strong baselines. Further, an additional full hierarchical paths reconstruction procedure significantly improves structural consistency, yielding the highest overall F1 scores. Taxon has been deployed in production within Alibaba's tax service system, handling an average of over 500,000 tax code queries per day and reaching peak volumes above five million requests during business event with improved accuracy, interpretability, and robustness.

Figures (11)

• Figure 1: An illustration of tax code prediction in e-commerce.
• Figure 2: Overview of the proposed framework. The system processes both textual and structured features through a hierarchical MoE to predict tax codes at multiple levels.
• Figure 3: Overview of the training workflow, integrating hierarchical feature-gating MoE with LLM-assisted semantic consistency. The multi-stage pipeline ensures robust learning across the tax code hierarchy.
• Figure 4: Four-stage data processing and training pipeline.
• Figure 5: Log-scale Cumulative distribution of prediction confidence for correctly predicted samples. Only 2.88% of predictions fall below 0.9 confidence, motivating balanced sampling.