Concept-based human-machine knowledge system

Abstract

Despite recent advancements, Large Language Models (LLMs) face difficulties in providing reliable and verifiable information, especially in critical domains. It is challenging to find possible hallucinations or manage the intrinsic knowledge of modern models. This paper presents an approach leveraging human-machine cooperation inspired by LLM technologies and Knowledge Graph (KG) structures. The proposed framework enables users to create a hierarchical network of concepts from textual input. Concepts range from characters (low-level) through words (medium-level) to word chains (high-level). The system generates relationships mainly on word co-occurrences within sentences, while leveraging Natural Language Processing (NLP) techniques and statistics for providing recommendations for entities and relations that users then validate and curate through an interactive interface. This cooperation between human and machine embraces the strengths of computational accuracy and storage capability with human cognitive skills. The idea aims to empower users to effectively organise complex information and navigate the information space with ease through the help of strongly connected concepts.