Addressing Hallucinations in Language Models with Knowledge Graph Embeddings as an Additional Modality

TL;DR

This paper introduces a method to reduce hallucinations in large language models by integrating knowledge graph embeddings via adapters, improving factual accuracy without retraining the entire model.

Contribution

The authors propose a novel approach that incorporates knowledge graph embeddings into language models using adapters, avoiding fine-tuning and external retrieval.

Findings

Improved performance on HaluEval, True-False, and FEVER datasets.

Created WikiEntities dataset with 3 million annotated Wikipedia texts.

Demonstrated reduced hallucinations in multiple LLMs.

Abstract

In this paper we present an approach to reduce hallucinations in Large Language Models (LLMs) by incorporating Knowledge Graphs (KGs) as an additional modality. Our method involves transforming input text into a set of KG embeddings and using an adapter to integrate these embeddings into the language model space, without relying on external retrieval processes. To facilitate this, we created WikiEntities, a dataset containing over 3 million Wikipedia texts annotated with entities from Wikidata and their corresponding embeddings from PyTorch-BigGraph. This dataset serves as a valuable resource for training Entity Linking models and adapting the described method to various LLMs using specialized adapters. Our method does not require fine-tuning of the language models themselves; instead, we only train the adapter. This ensures that the model's performance on other tasks is not affected. We trained an adapter for the Mistral 7B, LLaMA 2-7B (chat), and LLaMA 3-8B (instruct) models using this dataset and demonstrated that our approach improves performance on the HaluEval, True-False benchmarks and FEVER dataset. The results indicate that incorporating KGs as a new modality can effectively reduce hallucinations and improve the factual accuracy of language models, all without the need for external retrieval.