Automatic Validation of Textual Attribute Values in E-commerce Catalog by Learning with Limited Labeled Data

Yaqing Wang, Yifan Ethan Xu, Xian Li, Xin Luna Dong, Jing Gao

Abstract

Maintaining high-quality product catalogs is a significant challenge for large-scale e-commerce platforms due to the prevalence of self-reported, noisy attribute data. Traditional Natural Language Inference (NLI) models require extensive labeled datasets for each product category, which is impractical for the diverse and ever-changing "long tail" of e-commerce. We present MetaBridge, a novel framework that combines meta-learning with latent variable modeling to validate product attributes using only a handful of labeled examples. By leveraging vast amounts of unlabeled data through a unique "Bridging Regularizer," MetaBridge captures category-specific nuances and achieves superior performance over state-of-the-art baselines in low-data regimes.

Introduction

E-commerce catalogs are often populated by millions of retailers, leading to frequent errors in product attributes (e.g., a "flavor" attribute listed as "Non-GMO" instead of "Vanilla"). Accurate attribute validation is essential for effective search and recommendation systems. However, the sheer scale and diversity of product categories make manual validation impossible. While deep learning models like BERT can treat this as an NLI task, they struggle with the data scarcity inherent in niche or new categories. MetaBridge addresses this by treating attribute validation as a few-shot learning task, effectively bridging the gap between limited labels and abundant unlabeled descriptions.

The MetaBridge Framework

MetaBridge is built upon a stochastic neural network architecture designed to handle the uncertainty of diverse product categories. The framework introduces several key innovations to optimize performance with minimal supervision.

Architecture and Methodology

  1. Meta-Learning Foundation: The model utilizes a Model-Agnostic Meta-Learning (MAML) approach to identify a global parameter set that can be rapidly adapted to specific categories with minimal labeled data.
  2. Latent Variable Modeling: To account for category-specific uncertainty, the model incorporates a latent variable ($z$). This allows the network to learn the underlying distribution of a category rather than memorizing specific training instances.
  3. The Bridging Regularizer: This core component enforces consistency between the distribution of labeled "query" samples and the larger "support" set of unlabeled samples within the same category using a KL-divergence term.
  4. Entropy Minimization: During the adaptation phase, the model minimizes the entropy of its predictions on unlabeled data, encouraging confident and feature-rich learning without explicit labels.
  5. Transformer Encoders: Twin Transformers are employed to encode unstructured product profiles and attribute values into a shared semantic feature space.
Architecture of MetaBridge showing the Validation and Adaptation phases, incorporating latent variable modeling and the Bridging Regularizer.

Experimental Evaluation

The framework was evaluated on two large-scale internal Amazon datasets: Flavor (321 categories) and Ingredient (658 categories). The experiments simulated a few-shot environment, providing the model with 100 unlabeled products and only 3 to 10 labeled examples per category for adaptation.

Key Results

MetaBridge consistently outperformed state-of-the-art baselines, including BERT, MAML, and Meta-SGD. On the Flavor dataset, MetaBridge achieved a Precision-Recall (PR) AUC of 0.7852, surpassing BERT (0.7599) and Meta-SGD (0.7575).

One of the most significant improvements was observed in high-precision scenarios. At a precision threshold of 0.8 (R@P=0.8), MetaBridge demonstrated a 10% improvement over the strongest BERT baseline. Furthermore, the model proved highly robust to extreme data scarcity; when labeled query samples were reduced to just 3 per category, MetaBridge maintained a 5.5% performance lead over standard MAML.

Performance comparison between MAML and MetaBridge across different numbers of query samples per task for PR AUC and Recall at Precision 0.8.

Discussion and Conclusion

The success of MetaBridge highlights the value of using unlabeled data as a prior for category-specific features. By enforcing distribution consistency through the Bridging Regularizer and utilizing stochasticity for regularization, the model avoids the overfitting common in few-shot settings.

While highly effective for textual attributes, challenges remain for specialized domains with rare vocabularies, such as complex chemical ingredients. Future research directions include multi-modal validation—integrating product images to verify textual claims—and expanding the framework to handle complex hierarchical relationships within e-commerce catalogs.

Summary updated

Abstract

Product catalogs are valuable resources for eCommerce website. In the catalog, a product is associated with multiple attributes whose values are short texts, such as product name, brand, functionality and flavor. Usually individual retailers self-report these key values, and thus the catalog information unavoidably contains noisy facts. Although existing deep neural network models have shown success in conducting cross-checking between two pieces of texts, their success has to be dependent upon a large set of quality labeled data, which are hard to obtain in this validation task: products span a variety of categories. To address the aforementioned challenges, we propose a novel meta-learning latent variable approach, called MetaBridge, which can learn transferable knowledge from a subset of categories with limited labeled data and capture the uncertainty of never-seen categories with unlabeled data. More specifically, we make the following contributions. (1) We formalize the problem of validating the textual attribute values of products from a variety of categories as a natural language inference task in the few-shot learning setting, and propose a meta-learning latent variable model to jointly process the signals obtained from product profiles and textual attribute values. (2) We propose to integrate meta learning and latent variable in a unified model to effectively capture the uncertainty of various categories. (3) We propose a novel objective function based on latent variable model in the few-shot learning setting, which ensures distribution consistency between unlabeled and labeled data and prevents overfitting by sampling from the learned distribution. Extensive experiments on real eCommerce datasets from hundreds of categories demonstrate the effectiveness of MetaBridge on textual attribute validation and its outstanding performance compared with state-of-the-art approaches.
BibTeX
@inproceedings{wang2020automatic,
  author = {Yaqing Wang and Yifan Ethan Xu and Xian Li and Xin Luna Dong and Jing Gao},
  title = {{Automatic Validation of Textual Attribute Values in E-commerce Catalog by Learning with Limited Labeled Data}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2020},
  doi = {10.1145/3394486.3403303},
}
Figure