US20250111192
2025-04-03
Physics
G06N3/042
The patent application introduces a system for generating knowledge graphs using large language models (LLMs) to enhance chatbot functionality. This system allows chatbots to answer user queries by utilizing LLMs for novel questions and knowledge graphs for repeat inquiries. The integration of knowledge graphs provides efficiencies such as improved debugging, provenance tracking, and augmentation with additional data sources. This approach aims to address the limitations of LLMs, such as inaccuracies and resource-intensive fine-tuning, while maintaining their ability to handle complex queries across various domains.
Chatbots have evolved from rule-based systems, which relied on predefined responses, to more advanced AI-driven models capable of handling complex tasks. Large language models (LLMs) are a type of generative AI that excels in understanding and generating human-like text responses. These models have gained popularity due to their ability to process a wide range of queries. However, despite their capabilities, LLMs face challenges like producing incorrect answers and lacking transparency regarding the source of information.
Knowledge graphs are structured data representations that organize information using nodes, edges, attributes, and labels. They enable chatbots to analyze relationships and entities, providing a straightforward mechanism for updating and modifying responses. While knowledge graphs are effective for certain tasks, they typically require extensive datasets and may struggle with unexpected or complex queries. By leveraging LLMs to build these graphs, the system can overcome these limitations and enhance chatbot reliability.
The proposed system uses a combination of knowledge graphs and LLMs to answer user queries. Initially, the knowledge graph is queried for an answer; if none is found, the system prompts the LLM for a response. The LLM's answer may be validated through secondary LLMs or external data sources before being added to the knowledge graph. This process builds a robust knowledge base while ensuring accuracy and transparency in chatbot responses.
The system offers users additional features such as provenance data for answers, detailing the source or origin of information provided by the chatbot. This transparency allows users to trust and verify the answers independently. By storing provenance information within the knowledge graph, chatbots can easily retrieve and present this data upon request, enhancing user confidence in the chatbot's responses.