How Chatbots Use AI and Fact Databases to Learn and Improve

This patent describes how a chatbot can combine the broad answering abilities of a large language model with a structured knowledge graph to provide more accurate, verifiable, and debuggable responses, especially for repeat questions.

What it covers

This patent describes a method for a chatbot to answer user queries by combining a large language model (LLM) with a knowledge graph. When a query is received, the system first converts it into a 'formal query language' to search the knowledge graph for relevant information. This information from the knowledge graph is then given to the LLM to help it generate a more accurate 'formal language response' to the query. Critically, the system then 'validates an accuracy' of this response. After validation, the formal response is converted back into a natural language answer for the user. Based on this validated response, or user feedback, the knowledge graph is updated by 'modifying,' 'removing,' or 'adding' information, making the chatbot smarter over time. For example, if a user asks about a product's specifications, the chatbot might use the knowledge graph to find basic facts, then send those facts to an LLM to generate a natural-sounding answer, and finally use that answer to correct or expand the product data in the knowledge graph.

What it doesn't cover

• —Does not cover chatbots that rely solely on a large language model without querying a structured knowledge graph for information.
• —Does not cover systems that do not convert natural language queries into a 'formal query language' for a knowledge graph.
• —Does not cover chatbots that fail to validate the accuracy of the large language model's response before updating the knowledge graph.
• —Does not cover systems where the knowledge graph is not updated based on the large language model's responses or user feedback.
• —Does not cover large language models that generate responses without first receiving specific information from a knowledge graph node.

The clever bit

The cleverness lies in using the knowledge graph to both guide the LLM's responses with factual data and to continuously learn from the LLM's validated outputs, creating a self-improving, more reliable system. This combines the LLM's broad understanding with the knowledge graph's structured accuracy.

Why it matters

This patent addresses a key challenge with large language models: their tendency to sometimes produce incorrect or 'hallucinated' information. By integrating a knowledge graph, the system can ground LLM responses in verifiable facts, improving reliability. This approach also allows for easier debugging and the ability to provide 'provenance data,' or sources, for answers, which is crucial for trust in AI systems. This combination helps create more dependable and efficient chatbots for various applications.

Real-world examples

• 1.Amazon Alexa's information retrieval
• 2.Customer service chatbots for e-commerce
• 3.Virtual assistants providing factual answers
• 4.Enterprise search tools enhanced with AI
• 5.AI-powered question-answering systems