Skip to content

Repository on my course Knowledge and Language. This project has the purpose of evaluating how knowledge injection methods like RAG and PRAG affect the hallucination rate of LLMs. This hallucination rate will be evaluated comparing a Gold Standard Knowledge Graph to the one extracted from the LLMs output

Notifications You must be signed in to change notification settings

McDoritos/KL_Knowledge-Injection-Hallucinations

Repository files navigation

kl-project-2025

Process

Each JSON file in the SciERC dataset contains:

  • Documents
  • Sentences belonging to those documents
  • Relations between entities within those sentences

The goal is to generate questions for each document so that the LLM can produce triplets in the same structured format as defined by the SciERC dataset.
These triplets follow the structure:

$$[subject:label, relationship, object:label]$$

Entity and Relation Labels

Both entities and relations are constrained to a predefined set of valid labels.

Entity Labels

  • Method
  • Task
  • Dataset

Relation Labels

  • Used-For
  • Part-Of
  • Compare-With
  • SubClass-Of
  • Synonym-Of
  • Evaluated-With
  • Benchmark-For
  • Trained-With
  • SubTask-Of

Experimental Procedure

During the experiment, the LLM must be prompted with clear and specific instructions to ensure the correct extraction of these triplets for later comparison.

Because the SciERC dataset is quite extensive, only a subset of the available documents from the different JSON files should be used.
This makes the experiment computationally feasible while maintaining representative coverage of entity and relation types.

Five documents from the training dataset were randomly selected and had gold standard KG's generated for them, the questions that will be prompted to the LLM must now representitive enough of this documents in order for the LLM to construct a good KG.


Gold Standard Construction

The gold standard Knowledge Graph (KG), used as the reference for evaluation, must correspond exactly to the same set of documents for which the LLM-generated triplets were produced.

Including any additional documents in the gold standard that were not part of the LLM question set would introduce bias into the comparison process.


About

Repository on my course Knowledge and Language. This project has the purpose of evaluating how knowledge injection methods like RAG and PRAG affect the hallucination rate of LLMs. This hallucination rate will be evaluated comparing a Gold Standard Knowledge Graph to the one extracted from the LLMs output

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published