Standardizing Food Delivery Dish Names: Knowledge Graph Construction and Applications

Food‑delivery dish names are highly personalized, which creates difficulties for operational analysis, recall ranking, and backend management. This article, the second in a series on food‑delivery knowledge graphs, describes the end‑to‑end process of building a standardized dish‑name system, covering NLP techniques (entity extraction, text matching, relation classification) and CV techniques (image matching).

1. Background and Goals

Dish items are core to the delivery transaction and affect supply‑demand matching. Millions of dishes have inconsistent naming (e.g., many variants of “tomato scrambled eggs”), making aggregation impossible with simple keyword matching. The goal is to create a standard name taxonomy that enables accurate aggregation for supply‑demand analysis, personalized recall, and merchant tag generation.

2. Industry Survey

The design is inspired by Taobao’s SPU model, which standardizes products by key attributes and bound attributes. For food, the focus shifts from attribute standardization to dish‑name standardization.

3. Problem Analysis and Challenges

Challenges include high personalization, non‑standard entry, lack of industry‑level granularity, and limited culinary knowledge.

4. Solution

4.1 Name Aggregation

Two methods are used to discover synonym relations:

• Rule‑based matching: NER parses dish names into ingredient, cooking method, etc., and a synonym dictionary links equivalent terms (e.g., “potato” ↔ “洋芋”).

• Semantic matching: A BERT‑DSSM model trained on millions of positive/negative pairs expands synonym coverage beyond rule‑based limits.

4.2 Matching & Mapping

Standard name mapping combines text and image signals.

• Text matching: After cleaning, 2‑gram overlap recalls candidate standard names, Jaccard distance filters the top‑20, and BERT embeddings rank them using a combined Jaccard‑plus‑Cosine score.

• Image matching: An EfficientNet‑based multi‑class classifier predicts the top‑level standard name from dish images, with iterative noise‑filtering and active learning to improve robustness.

4.3 Hierarchical Construction

Both rule‑based and model‑based approaches discover parent‑child relations among standard names, producing a multi‑level hierarchy that supports fine‑grained aggregation for recommendation and ranking scenarios.

5. Business Applications

The standard name layer improves product aggregation, powers food‑ranking lists, and enables interactive recommendation by linking user‑viewed items to related dishes.

6. Summary and Outlook

The system now covers the majority of online dishes, delivering measurable business gains. Future work includes continuous synonym and hierarchy refinement, focusing on high‑impact core standard names, and deeper integration with downstream recommendation models.