Knowledge Injection for Large Language Models

Anonymous

Abstract: Generative Large Language Models (LLMs), such as ChatGPT and GPT-4, offer interactive APIs that can answer common questions at the human-expert level. However, these models often give inaccurate responses when faced with questions requiring domain-specific or professional-specific knowledge not covered in their training corpus. To alleviate this issue, Knowledge Graphs (KGs) have been integrated into LLMs as an additional source of knowledge. However, many state-of-the-art LLMs are not open-source, making it challenging to inject knowledge with model APIs only. In this paper, we propose a novel framework $\texttt{KnowGPT}$, which necessitates the $\textit{knowledge injection}$ for both knowledge retrieval and translation for LLMs. $\texttt{KnowGPT}$ leverages $(i)$ deep reinforcement learning to carefully extract context-aware knowledge from KGs, and $(ii)$ a multi-armed bandit to construct an appropriate prompt format for each question. It significantly outperforms the existing methods on three benchmark datasets. Notably, $\texttt{KnowGPT}$ attains a 91.6\% accuracy on OpenbookQA official leaderboard, which is comparable to human performance. The code will be open-sourced.

Paper Type: long

Research Area: Information Retrieval and Text Mining

Contribution Types: Model analysis & interpretability

Languages Studied: English