KNOWLEDGE GRAPH ASSISTED LARGE LANGUAGE MODELS

Smart overview of the Invention

The patent application outlines a system that integrates knowledge graphs with large language models (LLMs) to enhance chatbot functionality. This approach addresses the limitations of LLMs by using them to build and expand knowledge graphs, which can then be utilized to answer repeated queries more efficiently. The knowledge graph can be debugged, improved, and augmented with other data sources, enhancing the reliability and accuracy of chatbot responses. By correcting inaccuracies in LLM-generated answers within the knowledge graph, the system offers a robust solution for providing accurate information across various domains.

Chatbots have evolved from rule-based systems to more advanced models capable of handling complex queries using artificial intelligence (AI). Traditional chatbots relied on predefined rules and decision trees, limiting their effectiveness in dealing with unexpected or complex questions. The advent of generative AI, particularly LLMs, has significantly improved chatbots' capabilities by enabling them to understand context and generate human-like responses. Despite these advancements, LLM-based chatbots face challenges such as resource-intensive fine-tuning, inaccurate responses, and difficulty tracing the source of information.

The described system leverages LLMs to construct knowledge graphs that grow with each new query answered by the LLM. These graphs offer a structured way to store information, making it easier to debug and enhance responses. Knowledge graphs consist of nodes representing entities or concepts, edges denoting relationships, attributes providing metadata, and labels indicating the type of entity or relationship. This structured format allows for straightforward updates and corrections, ensuring that chatbots provide accurate and reliable answers.

While LLMs excel at handling complex and novel queries due to their extensive training data and contextual understanding capabilities, they are prone to errors and inefficiencies. The integration with knowledge graphs mitigates these issues by providing a mechanism for validating and storing accurate answers. When a query cannot be resolved using the existing knowledge graph, the system consults the LLM for an answer, which is then validated through secondary LLMs or external data sources before being added to the graph.

The system also enhances user experience by offering provenance data for answers provided by the chatbot. This data includes information about the source or origin of answers, enabling users to assess the trustworthiness of the information and verify it independently if needed. By combining the strengths of LLMs with the structured reliability of knowledge graphs, this system offers an innovative solution for improving chatbot performance across diverse subject areas.