Multi-Value-Product Retrieval-Augmented Generation for Industrial Product Attribute Value Identification

Huike Zou, Haiyang Yang, Yindu Su, Chen Li Yu, Qinye Xie, Chengbao Lian, Qingheng Zhang, Shuguang Han, Fei Huang, Jufeng ChenEMNLP 2025

TL;DR

This paper introduces MVP-RAG, a novel approach for identifying attribute values (like brand, size, color) for products in e-commerce. By combining retrieval and generation techniques, MVP-RAG significantly improves accuracy and handles "out-of-distribution" values better than previous methods, demonstrating strong performance on real-world industrial data.

What Problem Does This Paper Address?

In e-commerce, accurately identifying product attributes and their values is crucial for effective search, recommendations, and business analytics. This process, known as Product Attribute Value Identification (PAVI), is challenging, especially on platforms with many individual sellers who may not provide complete or standardized information. Existing PAVI methods struggle with:

  • Cascading errors: Errors in early stages propagate and worsen later results.
  • Out-of-Distribution (OOD) values: Difficulty identifying attribute values not seen during training.
  • Generalization: Inability to adapt to new or evolving attribute taxonomies.
  • Implicit values: Attributes that are implied but not explicitly stated.

These limitations hinder the efficiency and user experience of e-commerce platforms.

Methods

The paper proposes Multi-Value-Product Retrieval-Augmented Generation (MVP-RAG), which frames PAVI as a retrieval-generation task. The core idea is to leverage both similar products and attribute information to guide a generative model.

MVP-RAG has two main components:

  1. Multi-level Attribute Value-Product Retrieval:
    • Attribute Value Retrieval: Uses a retrieval model (TACLR) to find candidate attribute values relevant to the product's category.
  2. Product Retrieval: Employs a general-purpose representation model (BGE) to find similar products within the same category. This provides contextual "few-shot" examples.
  3. Attribute Value Generation:
    • A large language model (LLM), specifically Qwen2.5, is fine-tuned.
  4. The LLM is prompted with a structured input including task definition, similar products, product information, and candidate attribute values.
  5. The model then generates the standardized attribute values.
  6. To improve OOD handling, the training incorporates a separate batch of OOD attribute values.

Datasets

The authors evaluate their model on two datasets:

  • Xianyu-PAVI dataset: A proprietary dataset sourced from Xianyu, a C2C e-commerce platform. It's described as large-scale, containing 8,803 product categories, 26,645 category-attribute pairs, and 6.3 million category-attribute-attribute value triples.
  • WDC-PAVI dataset: Another dataset comprising products across five categories, with a significant portion of null values in product-attribute pairs.

Key Insights and Ideas

  • Retrieval as a Stabilizer: By retrieving both similar products and candidate attribute values, MVP-RAG reduces the "hallucination" common in pure generation models and provides a more stable foundation for attribute value identification.
  • Multi-level Context is Key: Combining product-level context (similar items) with attribute-level context (candidate values) is more effective than using either alone. This synergy helps in both accuracy and generalization.
  • OOD Handling through Targeted Training: Explicitly training with out-of-distribution attribute values is a direct strategy to improve the model's ability to identify unseen values.
  • Real-world Deployment Matters: The successful deployment of MVP-RAG on Xianyu, processing millions of listings and adapting to dynamic taxonomies, validates its practical utility beyond academic benchmarks.

Key Results

  • MVP-RAG achieves state-of-the-art F1 scores on both the Xianyu-PAVI and WDC-PAVI datasets.
  • On the Xianyu-PAVI dataset, MVP-RAG achieved an F1 score of 89.5%, outperforming previous retrieval-based methods like TACLR by 3.3% and generation-based methods (Product-RAG) by 26.3%.
  • Ablation studies show that removing either the attribute value retrieval or product retrieval module leads to a significant drop in F1 scores, highlighting the importance of each component.
  • The model demonstrates robustness, with attribute value accuracy remaining high even when up to 75% of recalled product attributes are incorrect for simple attributes like brand.

Limitations and Future Work

The authors acknowledge several limitations:

  • Text-only Information: The current model only uses text data. Incorporating image or video information could provide crucial context for attributes like color, material, or shape, especially for difficult-to-infer cases.
  • LLM Inference Latency: The reasoning process of the LLM is time-consuming. Future work should focus on accelerating and optimizing the model's inference speed for industrial-scale applications.

Summary updated

BibTeX
@inproceedings{zou2025multi,
  author = {Huike Zou and Haiyang Yang and Yindu Su and Chen Li Yu and Qinye Xie and Chengbao Lian and Qingheng Zhang and Shuguang Han and Fei Huang and Jufeng Chen},
  title = {{Multi-Value-Product Retrieval-Augmented Generation for Industrial Product Attribute Value Identification}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure