Have you ever bought a product abroad, only to find it’s different from the one sold in your home country? This practice, known as dual quality, is a growing concern for consumers, especially in the EU. In our recent study, we developed an NLP-based solution to automatically detect such discrepancies in product reviews. Here’s a simplified breakdown of our work.

What is Dual Quality?

Dual quality occurs when companies sell products under the same brand and packaging in different markets but with varying quality or ingredients. This can mislead consumers and violate fair competition rules.

Our Approach

We created a Polish-language dataset of 1,957 reviews (540 highlighting dual quality) and tested multiple NLP methods, including:

• SetFit with sentence-transformers (e.g., LaBSE, multilingual models).
• Transformer-based encoders (e.g., Polish RoBERTa, XLM-RoBERTa).
• Large Language Models (LLMs) like DeepSeek-V3 and GPT-4o.

Key Findings

1. Rare but Significant: Dual quality mentions are rare in reviews, making detection challenging.
2. Small Models Shine: Language-specific models (e.g., Polish RoBERTa) performed as well as larger LLMs.
3. Prompting Matters: For LLMs, clear instructions worked better than examples, which sometimes reduced accuracy.

Multilingual Capabilities

We expanded our system to analyze reviews in English, German, and French, showing promising results for cross-market comparisons.

Practical Use

Our solution is created for Poland’s Office of Competition and Consumer Protection (UOKiK), helping flag potential dual quality cases for further investigation.

Why This Matters

Automating dual quality detection empowers consumer agencies and ensures fairer markets. Future work could extend this to more languages and product categories.

If you’re interested in learning more about our findings, you can access the full article here. 

Thank you for reading, and I look forward to your thoughts and feedback!