Knowledge Injection to Counter Large Language Model (LLM) Hallucination

Abstract

A shortfall of Large Language Model (LLM) content generation is hallucination, i.e., including false information in the output. This is especially risky for enterprise use cases that require reliable, fact-based, controllable text generation at scale. To mitigate this, we utilize a technique called Knowledge Injection (KI), where contextual data about the entities relevant to a text-generation task is mapped from a knowledge graph to text space for inclusion in an LLM prompt. Using the task of responding to online customer reviews of retail locations as an example, we have found that KI increases the count of correct assertions included in generated text. In a qualitative review, fine-tuned bloom-560m with KI outperformed a non-fine-tuned text-davinci-003 model from OpenAI, though text-davinci-003 has 300 times more parameters. Thus, the KI method can increase enterprise users’ confidence leveraging LLMs to replace tedious manual text generation and enable better performance from smaller, cheaper models.