Hypothesis
Many hypothesis have been done in this project (by me at least). I'd like you to comment on them here and in your reports.
H1. At the start, I thought that data cleaning was required before product grouping and merging. I assumed that once done, product grouping could be done based on brand and product name alone. Now it doesn't seem to be true. Could you comment please.
H2. By mid summer, I then thought that product grouping couldn't be done without the help of LLMs or the Reference DB. Do you think that it's true or not ?
H3. Early in the project, I assumed that product merging would be difficult without LLMs (which was called plan B). Could you tell us what you think about it.
H4. It was implicitly assumed in this project that it was realistic to group and merge products in Open Food Facts. But maybe it's not... What do you think about it.
Hi Louis,
H1: Data cleaning is a necessary step before product grouping/merging because it ensures that the products are consistent and complete. But the experiments showed that brand+product_name alone is not sufficient. Because there are many products that have reordered words (Smooth Peanut Butter, Peanut Butter Smooth) and are incomplete (Instant Oatmeal, Instant Oatmeal Maple Brown Sugar). In those cases, we would need additional evidence like nutriments, ingredients to support product grouping/merging.
H2: It is true. We wouldn't need LLMs to group every product. Deterministic models (RapidFuzz, TF-IDF, Splink) helped us in most cases. But there are a few ambiguous cases (mainly when product names are incomplete), where these models struggle to prove that the products can be grouped. Only such cases are sent to LLMs for review.
H3: Same as H2. The same logic applies to product merging. Splink comparisons focus on this logic. In cases where all the deterministic models failed, LLMs have given us good results. So, this hybrid model can give us promising results for both product grouping and product merging.
H4: It is probably unrealistic to expect a system to automatically produce a perfectly deduplicated global product catalogue from all Open Food Facts records with no human review, because the source contains incomplete names, inconsistent nutrition, reformulations, regional variants, package variations, and sometimes insufficient information to determine identity at all. But it is realistic to build a system that confidently groups a subset of records, leaves uncertain cases unresolved, and progressively improves coverage using reference data, LLM reasoning, or human review. In other words, the target should not be “every record must be merged”. It should be high-precision grouping where evidence is sufficient, explicit uncertainty where it is not, and safe enrichment only within validated groups.