Knowledge Graph: A Brief Overview and Creation Guide

A knowledge graph is a structured representation of entities and their relationships within a system or domain.

In this post, I want to briefly talk about how to create a knowledge graph. At DoorDash, understanding products and their attributes is crucial for customer recommendations and brand advertisements. For example, when a new product like Coke is added, we need to determine its relationships, such as:

(coca-cola company) -[manufactures]-> (coke)

How to Extract Entities?

This process is called Named-Entity Recognition (NER).

• Brute Force (Manual Annotation)
  • Manually labeling entities in the dataset.
  • Cons:
    • Not scalable.
    • Costly and inefficient.
    • Not recommended at DoorDash, which processes large amounts of input data daily.
• String Matching + Classifier
  • Uses string matching combined with a classifier to verify if an entity passes a defined threshold.
  • Pros:
    • High precision.
    • Decent coverage.
  • Cons:
    • Only work for obvious patterns.

There are some more classic ways to perform entity extraction, but what if we leverage LLMs to assist us? LLMs have contextual and semantic understanding that can improve the process.

• Using LLMs to Identify Attributes
  • Suitable for simple and common attributes such as identifying a product’s brand.

What if we need additional attributes with deeper meanings that are not explicitly stated on product labels or descriptions, or specialized attributes unique to this product?

• Few-Shot Prompting via Embedding Similarity
  • Retrieve similar but labeled products using embedding similarity.
  • Usefew-shot prompting to guide the LLM based on known examples and labeled attributes.
• LLM Reasoning
  • Make the LLM “think” and deduce results.
  • Example: An LLM could analyze a product’s packaging photo and infer whether it is organic.
• LLM Agent
  • A specialized software system with an LLM core and Pre-defined goals and characteristics.
  • We can use it to autonomously perform specific tasks.

Reference

Building DoorDash’s Product Knowledge Graph with Large Language Models