[Paper Review] MAVE: A Product Dataset for Multi-source Attribute Value Extraction

Li Yang, Qifan Wang+6 more|arXiv (Cornell University)|Dec 16, 2021

Sentiment Analysis and Opinion Mining4 citations

TL;DR

MAVE is a large-scale, multi-source product dataset with 2.2 million products and 3 million attribute-value annotations across 1257 categories, designed to advance attribute value extraction in e-commerce. It enables state-of-the-art performance in zero-shot and few-shot learning, with a novel model (MAVEQA) outperforming baselines on long, complex product profiles and demonstrating strong generalization on unseen attributes.

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}}.

Motivation & Objective

• To address the lack of large-scale, diverse, and high-quality datasets for product attribute value extraction in e-commerce.
• To improve model generalization by enabling zero-shot and few-shot learning on unseen attributes.
• To provide a context-rich dataset with multi-source product information (title, description, specs, etc.) for better attribute coverage.
• To support research in robust, scalable attribute extraction under real-world conditions of missing or incomplete data.
• To establish a benchmark for evaluating attribute extraction models on long, complex product profiles.

Proposed method

• The dataset is constructed by extracting raw attribute values from Amazon Review pages using a predefined set of attributes.
• Rigorous filtering and post-mapping are applied to ensure high-quality, clean attribute-value pairs, resulting in 3 million annotations across 2.2 million products.
• Multi-source product representations are aggregated from title, description, features, specifications, and key-value pairs to enhance context and coverage.
• A novel sequence tagging model, MAVEQA, is proposed to jointly model multi-source inputs and handle long input sequences effectively.
• The model is fine-tuned on the MAVE dataset using a pre-trained ETC model, which is optimized for long sequences.
• Zero-shot and few-shot evaluation protocols are established to assess generalization, with systematic ablation on sequence length and model capacity.

Experimental results

Research questions

• RQ1 Can a large-scale, multi-source product dataset significantly improve zero-shot attribute value extraction performance?
• RQ2 How does multi-source input representation affect the accuracy and robustness of attribute value extraction models?
• RQ3 To what extent can models generalize to unseen attributes without any fine-tuning examples?
• RQ4 How does model performance degrade with increasing input sequence length, and can models be designed to handle long product profiles effectively?
• RQ5 What is the impact of pre-trained model choice on few-shot and zero-shot attribute extraction?

Key findings

• MAVE is the largest publicly available attribute value extraction dataset, with 3 million annotations across 2.2 million products and 1257 categories.
• The MAVEQA model achieves significantly better performance than AVEQA on long sequences, with stable precision and recall even beyond model max sequence length.
• On zero-shot attributes, MAVEQA outperforms AVEQA on Resolution and Number of Sinks, but both models fail completely on Compatibility, indicating a true zero-shot challenge.
• For few-shot learning, F1 scores on most attributes exceed 90% with 100 training examples, demonstrating strong sample efficiency.
• The model's performance remains flat on Food Processor Capacity, suggesting potential data quality issues with missing ground-truth annotations in the evaluation set.
• The dataset enables effective pretraining for downstream tasks such as product search, recommendation, and question answering.