Entity Mapping Boosts Google Knowledge Graph, Not LLMs

Image credit: Search Engine Journal

Entity mapping effectively enhances Google‘s curated Knowledge Graph, but its direct impact on the parametric memory of large language models (LLMs) like ChatGPT remains negligible, according to industry analysis released Tuesday.

This distinction highlights fundamental differences in how these advanced AI systems acquire and process factual information, influencing their reliability and the strategies for optimizing their knowledge bases.

Google’s Knowledge Graph, launched in 2012, relies on structured data and third-party verification, such as information from Wikidata, to build authoritative entity representations, Duane Forrester, an industry expert, reported. This structured approach allows entity mapping to directly feed and refine the graph’s nodes and relationships.

The search giant’s artificial intelligence answers also leverage this same Knowledge Graph, extending the influence of precise entity mapping beyond traditional search results, Forrester said.

In contrast, large language models like ChatGPT do not possess a similar graph structure that can be directly modified through entity mapping, Forrester explained. Their factual understanding is an emergent property, encoded parametrically during extensive training on vast text corpuses.

Schema markup or ‘sameAs’ links, which are crucial for entity mapping on Google, do not directly influence the internal factual knowledge of LLMs, according to the analysis. The models learn associations and facts from the statistical patterns within their training data.

Forrester indicated that effective ‘entity work’ for LLMs involves a slower, more indirect process. This includes generating citations and mentions from a wide array of trusted external sources, which gradually reinforces information within the models’ training data over time.

This method, while more durable, requires a different approach than the on-site schema audits and direct data feeds that prove highly effective for Google’s Knowledge Graph, experts noted.

Source: Search Engine Journal