MAVE: A Product Dataset for Multi-source Attribute Value Extraction

Li Yang, Qifan Wang, Zac Yu, Anand Kulkarni, Sumit Sanghai, Bin Shu, Jon Elsas, Bhargav Kanagal

Abstract

Attribute value extraction refers to the task of identifying values of an attribute of interest from product information. Product attribute values are essential in many e-commerce scenarios, such as customer service robots, product ranking, retrieval and recommendations. While in the real world, the attribute values of a product are usually incomplete and vary over time, which greatly hinders the practical applications. In this paper, we introduce MAVE, a new dataset to better facilitate research on product attribute value extraction. MAVE is composed of a curated set of 2.2 million products from Amazon pages, with 3 million attribute-value annotations across 1257 unique categories. MAVE has four main and unique advantages: First, MAVE is the largest product attribute value extraction dataset by the number of attribute-value examples. Second, MAVE includes multi-source representations from the product, which captures the full product information with high attribute coverage. Third, MAVE represents a more diverse set of attributes and values relative to what previous datasets cover. Lastly, MAVE provides a very challenging zero-shot test set, as we empirically illustrate in the experiments. We further propose a novel approach that effectively extracts the attribute value from the multi-source product information. We conduct extensive experiments with several baselines and show that MAVE is an effective dataset for attribute value extraction task. It is also a very challenging task on zero-shot attribute extraction.

Introduction

In e-commerce, structured data is critical for search filters, recommendation engines, and comparison tools. However, essential information—such as a laptop's "7 Hours" of "Battery Life"—is often buried in unstructured text across titles, bulleted lists, and long descriptions. Existing research in Attribute Value Extraction (AVE) faces three major hurdles: small or private datasets, a narrow focus on single sources (like titles), and the "long sequence" problem where standard Transformer models truncate detailed product pages.

The MAVE Dataset

The MAVE dataset addresses these limitations. Curated from millions of Amazon product pages with a 97.8% precision bar, it features 2.2 million products across 1,257 categories and 2,535 unique category-attribute pairs. Unlike previous datasets, MAVE includes multi-source data (titles, descriptions, features), forcing models to aggregate information from across the entire product page.

Figure 1: Distribution of categories and attributes in MAVE. The inner circle shows categories, while the outer circle shows specific attributes.

Methodology: MAVEQA

To leverage this multi-source data, we propose MAVEQA. We frame extraction as a Question Answering (QA) problem: given an attribute (the question) and product text (the context), the model identifies the answer span.

MAVEQA utilizes the Extended Transformer Construction (ETC) backbone to handle long sequences. It employs a "global-local" attention mechanism:

  • Global Tokens: Summarize high-level information from different sources (e.g., title vs. description).
  • Long Tokens: Represent the actual text words.
  • Efficiency: Words attend to a local neighborhood and global tokens, allowing the model to process sequences of 1024 tokens or more without the quadratic cost of standard self-attention.
Figure 2: The MAVEQA architecture uses global tokens to summarize different information sources and local attention to handle long sequences efficiently.

Results and Discussion

MAVEQA consistently outperforms standard baselines, particularly as sequence length increases. While standard models (AVEQA) see performance drops beyond 512 tokens due to truncation, MAVEQA maintains high precision and recall by processing the full context.

The dataset also reveals interesting insights into attribute rarity. "Tail" attributes with consistent linguistic patterns (e.g., "Black Tea Variety") are easily learned with few examples. However, "Zero-shot" extraction—identifying attributes never seen during training—remains a significant challenge, especially for attributes requiring domain knowledge like "Compatibility."

Conclusion

MAVE provides a large-scale, diverse, and multi-source benchmark for the AVE task. Our proposed MAVEQA model demonstrates the importance of handling long-form, multi-source text in e-commerce. Future work will explore using MAVE for pre-training general-purpose product representations.

Summary updated

Abstract

Attribute value extraction refers to the task of identifying values of an attribute of interest from product information. Product attribute values are essential in many e-commerce scenarios, such as customer service robots, product ranking, retrieval and recommendations. While in the real world, the attribute values of a product are usually incomplete and vary over time, which greatly hinders the practical applications. In this paper, we introduce MAVE, a new dataset to better facilitate research on product attribute value extraction. MAVE is composed of a curated set of 2.2 million products from Amazon pages, with 3 million attribute-value annotations across 1257 unique categories. MAVE has four main and unique advantages: First, MAVE is the largest product attribute value extraction dataset by the number of attribute-value examples. Second, MAVE includes multi-source representations from the product, which captures the full product information with high attribute coverage. Third, MAVE represents a more diverse set of attributes and values relative to what previous datasets cover. Lastly, MAVE provides a very challenging zero-shot test set, as we empirically illustrate in the experiments. We further propose a novel approach that effectively extracts the attribute value from the multi-source product information. We conduct extensive experiments with several baselines and show that MAVE is an effective dataset for attribute value extraction task. It is also a very challenging task on zero-shot attribute extraction. Data is available at {\it \url{https://github.com/google-research-datasets/MAVE}}.
BibTeX
@inproceedings{yang2022mave,
  author = {Li Yang and Qifan Wang and Zac Yu and Anand Kulkarni and Sumit Sanghai and Bin Shu and Jon Elsas and Bhargav Kanagal},
  title = {{MAVE: A Product Dataset for Multi-source Attribute Value Extraction}},
  booktitle = {WSDM},
  year = {2022},
  doi = {10.1145/3488560.3498377},
}
Figure