Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Federico Ranaldi, Andrea Zugarini, Leonardo Ranaldi, Fabio Massimo Zanzotto

TL;DR

This work formalizes KG protoknowledge as the memorized, reusable knowledge within LLMs, decomposing it into lexical, hierarchical, and topological forms. It introduces Knowledge Activation Tasks (KATs) to probe each form and links protoknowledge absorption to downstream Text-to-SPARQL performance under different prompting scenarios. Across experiments with GPT-4, GPT-3.5-Turbo, and Llama variants on Wikidata and DBpedia, topological protoknowledge emerges as the strongest predictor of correct SPARQL generation, while lexical and hierarchical forms also contribute, especially under low-context prompts. The study highlights a persistent semantic bias shaped by pretraining data distribution, underscores risks of semantic-level data contamination, and provides a practical framework for evaluating and leveraging protoknowledge in Closed-Pretraining settings.

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.

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

TL;DR

This work formalizes KG protoknowledge as the memorized, reusable knowledge within LLMs, decomposing it into lexical, hierarchical, and topological forms. It introduces Knowledge Activation Tasks (KATs) to probe each form and links protoknowledge absorption to downstream Text-to-SPARQL performance under different prompting scenarios. Across experiments with GPT-4, GPT-3.5-Turbo, and Llama variants on Wikidata and DBpedia, topological protoknowledge emerges as the strongest predictor of correct SPARQL generation, while lexical and hierarchical forms also contribute, especially under low-context prompts. The study highlights a persistent semantic bias shaped by pretraining data distribution, underscores risks of semantic-level data contamination, and provides a practical framework for evaluating and leveraging protoknowledge in Closed-Pretraining settings.

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.

Paper Structure

This paper contains 50 sections, 1 equation, 14 figures, 13 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Protoknowledge on Knowledge Graphs
  4. Forms of Protoknowledge
  5. Lexical Protoknowledge.
  6. Hierarchical Protoknowledge.
  7. Topological Protoknowledge.
  8. Knowledge Activation Tasks
  9. Measuring Protoknowledge
  10. Models and Analysis Design
  11. Lexical Protoknowledge
  12. KATs:
  13. Test Set.
  14. Results.
  15. Hierarchical Protoknowledge
  16. ...and 35 more sections

Figures (14)

  • Figure 1: Protoknowledge Impact: LLMs acquire three protoknowledge forms from Knowledge Graphs, measured via Knowledge Activation Tasks (KATs). The degree of absorption of protoknowledge measured by KATs correlates with the Text-To-SPARQL performances of LLMs showing that, when protoknowledge is acquired, it is also positively used.
  • Figure 2: SPS scores for Wikidata and DBpedia.
  • Figure 3: NO-URI vs Original Agreement and Disagreement. Above bar is reported Positive Agreement Ratio %.
  • Figure 4: URI Recognition Task
  • Figure 5: Inverse Subsumption Task
  • ...and 9 more figures