Attribute-Aware Controlled Product Generation with LLMs for E-commerce

TL;DR

The paper addresses the lack of high quality labeled data for e commerce product attribute extraction by proposing an attribute aware LLM based pipeline that generates synthetic product data through controlled modifications of existing records. It introduces three generation strategies and a structured prompt design to ensure attribute fidelity and semantic coherence across product text fields. Extensive human evaluation on 2,000 synthetic products demonstrates high realism and attribute validity, while downstream attribute extraction with synthetic data matches real data performance and improves with hybrid training, showing practical utility for low resource settings. The work provides a scalable, cost effective approach to augment e commerce datasets, enabling robust search filtering and recommendations despite limited labeled data.

Abstract

Product information extraction is crucial for e-commerce services, but obtaining high-quality labeled datasets remains challenging. We present a systematic approach for generating synthetic e-commerce product data using Large Language Models (LLMs), introducing a controlled modification framework with three strategies: attribute-preserving modification, controlled negative example generation, and systematic attribute removal. Using a state-of-the-art LLM with attribute-aware prompts, we enforce store constraints while maintaining product coherence. Human evaluation of 2000 synthetic products demonstrates high effectiveness, with 99.6% rated as natural, 96.5% containing valid attribute values, and over 90% showing consistent attribute usage. On the public MAVE dataset, our synthetic data achieves 60.5% accuracy, performing on par with real training data (60.8%) and significantly improving upon the 13.4% zero-shot baseline. Hybrid configurations combining synthetic and real data further improve performance, reaching 68.8% accuracy. Our framework provides a practical solution for augmenting e-commerce datasets, particularly valuable for low-resource scenarios.

Figures (5)

• Figure 1: Synthetic product generation pipeline architecture. Given a base product $p$ with text fields $T(p)$ and structured attributes $S(p)$, the system selects a category-appropriate attribute $s \in S(p)$. The Value Provider LLM generates a new value $v$, and a generation strategy is sampled using probabilities $\pi$ to create positive, negative, or unknown examples. The Generation LLM then produces a synthetic product while maintaining product coherence.
• Figure 2: Synthetic product examples from MAVE dataset. Colors indicate original text (red), synthetic text (green), and incorrect attributes (orange). Brand/model names replaced with placeholders for confidentiality.
• Figure 3: Distribution of products across categories
• Figure 4: Additional examples of synthetic product generation showing various attribute modifications. The first two are "correct" examples, the second two are "incorrect" examples, while the last is an "unknown" example
• Figure 5: Auditing interface for human evaluation of synthetic product data. The interface shows the original and synthetic products side by side, with modifications highlighted in color, followed by evaluation questions about validity, consistency, and quality.