JPAVE: A Generation and Classification-based Model for Joint Product Attribute Prediction and Value Extraction

Zhongfen Deng, Hao Peng, Tao Zhang, Shuaiqi Liu, Wenting Zhao, Yibo Wang, Philip S. Yu

Abstract

Product attribute value extraction is an important task in e-Commerce which can help several downstream applications such as product search and recommendation. Most previous models handle this task using sequence labeling or question answering method which rely on the sequential position information of values in the product text and are vulnerable to data discrepancy between training and testing. This limits their generalization ability to real-world scenario in which each product can have multiple descriptions across various shopping platforms with different composition of text and style. They also have limited zero-shot ability to new values. In this paper, we propose a multi-task learning model with value generation/classification and attribute prediction called JPAVE to predict values without the necessity of position information of values in the text. Furthermore, the copy mechanism in value generator and the value attention module in value classifier help our model address the data discrepancy issue by only focusing on the relevant part of input text and ignoring other information which causes the discrepancy issue such as sentence structure in the text. Besides, two variants of our model are designed for open-world and closed-world scenarios. In addition, copy mechanism introduced in the first variant based on value generation can improve its zero-shot ability for identifying unseen values. Experimental results on a public dataset demonstrate the superiority of our model compared with strong baselines and its generalization ability of predicting new values.

Introduction

In the competitive landscape of e-commerce, accurately extracting attribute-value pairs (e.g., Attribute: "Color", Value: "Midnight Blue") from product titles and descriptions is essential for enhancing search relevance and recommendation accuracy. Traditional methodologies typically frame this as a sequence labeling (e.g., NER) or Question Answering (QA) task. However, these approaches suffer from significant limitations: they are heavily dependent on exact token positions, vulnerable to "data shifting" across different platforms, and struggle with zero-shot extraction of previously unseen values.

Methodology

To address these challenges, we introduce JPAVE (Joint Product Attribute Prediction and Value Extraction). JPAVE utilizes a multi-task learning framework that decouples value extraction from rigid sequential positions, focusing instead on semantic content.

The JPAVE architecture featuring a shared text encoder and two extraction variants: JPAVE-GEN and JPAVE-CLS.

The architecture consists of a shared text encoder (Bi-GRU) and an attribute predictor, integrated with one of two specialized extraction modules:

  1. JPAVE-GEN (Open-World): Employs a GRU-based decoder with a copy mechanism (Pointer-Generator). This allows the model to either generate values from a fixed vocabulary or copy them directly from the input text, enabling the identification of novel values.
  2. JPAVE-CLS (Closed-World): Utilizes a value attention module to match input text against a pre-defined set of known values, optimized for scenarios where the attribute space is constrained.

By avoiding traditional BIO-tagging, JPAVE remains position-agnostic, effectively treating sentence structure as noise and focusing on the core semantic tokens.

Experimental Evaluation

We evaluated JPAVE on the MEPAVE dataset, a large-scale Chinese e-commerce repository containing 26 attributes and over 2,100 unique values across 71,194 training instances.

Robustness and Efficiency

To test the model's position-agnostic nature, we conducted experiments on permuted datasets where word order was randomized. Unlike sequence labeling models that fail under such conditions, JPAVE maintained stable performance. Furthermore, JPAVE is highly efficient; with only 12.3M parameters, it outperforms significantly larger models like T5-base and GPT-2, making it ideal for high-throughput production environments.

Results and Discussion

JPAVE demonstrated superior performance across all metrics, outperforming both text-only and multimodal baselines that utilize product images.

Performance comparison of JPAVE variants against existing baselines on the MEPAVE dataset.

As shown in the results, JPAVE-GEN achieved a state-of-the-art F1 score of 98.40% for attribute prediction and 97.51% for value extraction. Notably, in open-world scenarios, it achieved an 83.15% F1 score on unseen values with 100% precision, highlighting its robust zero-shot capabilities.

Conclusion

JPAVE represents a significant shift from position-dependent extraction to semantic-based joint prediction. By leveraging copy mechanisms and attention modules, it overcomes the hurdles of data discrepancy and unseen value identification. Future work will explore cross-lingual applications and the integration of visual features to further enhance extraction in text-sparse environments.

Summary updated

Abstract

Product attribute value extraction is an important task in e-Commerce which can help several downstream applications such as product search and recommendation. Most previous models handle this task using sequence labeling or question answering method which rely on the sequential position information of values in the product text and are vulnerable to data discrepancy between training and testing. This limits their generalization ability to real-world scenario in which each product can have multiple descriptions across various shopping platforms with different composition of text and style. They also have limited zero-shot ability to new values. In this paper, we propose a multi-task learning model with value generation/classification and attribute prediction called JPAVE to predict values without the necessity of position information of values in the text. Furthermore, the copy mechanism in value generator and the value attention module in value classifier help our model address the data discrepancy issue by only focusing on the relevant part of input text and ignoring other information which causes the discrepancy issue such as sentence structure in the text. Besides, two variants of our model are designed for open-world and closed-world scenarios. In addition, copy mechanism introduced in the first variant based on value generation can improve its zero-shot ability for identifying unseen values. Experimental results on a public dataset demonstrate the superiority of our model compared with strong baselines and its generalization ability of predicting new values.
BibTeX
@inproceedings{deng2023jpave,
  author = {Zhongfen Deng and Hao Peng and Tao Zhang and Shuaiqi Liu and Wenting Zhao and Yibo Wang and Philip S. Yu},
  title = {{JPAVE: A Generation and Classification-based Model for Joint Product Attribute Prediction and Value Extraction}},
  booktitle = {IEEE Big Data},
  year = {2023},
  doi = {10.1109/bigdata59044.2023.10386204},
}
Figure