Protoknowledge Shapes Behaviour of LLMs in Downstream Tasks: Memorization and Generalization with Knowledge Graphs

TL;DR

This paper introduces the concept of protoknowledge to analyze how Large Language Models memorize and utilize knowledge Graphs during pretraining and inference, impacting their generalization and task performance.

Contribution

It formalizes protoknowledge, categorizes its types, and develops Knowledge Activation Tasks to measure and analyze how LLMs leverage knowledge Graphs in downstream tasks.

Findings

Protoknowledge can be categorized into lexical, hierarchical, and topological forms.

Knowledge Activation Tasks effectively measure knowledge utilization in LLMs.

Prompting strategies influence the activation and use of protoknowledge in task performance.

Abstract

We introduce the concept of protoknowledge to formalize and measure how sequences of tokens encoding Knowledge Graphs are internalized during pretraining and utilized at inference time by Large Language Models (LLMs). Indeed, LLMs have demonstrated the ability to memorize vast amounts of token sequences during pretraining, and a central open question is how they leverage this memorization as reusable knowledge through generalization. We then categorize protoknowledge into lexical, hierarchical, and topological forms, varying on the type of knowledge that needs to be activated. We measure protoknowledge through Knowledge Activation Tasks (KATs), analyzing its general properties such as semantic bias. We then investigate the impact of protoknowledge on Text-to-SPARQL performance by varying prompting strategies depending on input conditions. To this end, we adopt a novel analysis framework that assesses whether model predictions align with the successful activation of the relevant protoknowledge for each query. This methodology provides a practical tool to explore Semantic-Level Data Contamination and serves as an effective strategy for Closed-Pretraining models.