ADD LLM TO Knowledge-Graph: NEW GIVE Method (Berkeley)

Graph Inspired Veracity Extrapolation (GIVE) is a novel reasoning framework designed to enhance the performance of large language models (LLMs) in knowledge-intensive tasks by integrating sparse external knowledge graphs (KGs) with the LLM's internal knowledge. The main insight of GIVE is that even when working with incomplete or limited KGs, it's possible to improve the reasoning capabilities of LLMs by using the structure of the KG to inspire the model to infer and extrapolate potential relationships between concepts. This approach facilitates a more logical, step-by-step reasoning process akin to expert problem-solving, rather than relying solely on direct fact retrieval from dense knowledge bases. The GIVE framework operates in several key steps. First, it prompts the LLM to decompose the query into crucial concepts and attributes, extracting key entities and relations relevant to the question. It then constructs entity groups by retrieving entities from the KG that are semantically similar to these key concepts. Within these groups, GIVE induces intra-group connections using the LLM's internal knowledge to explore relationships among similar entities. For inter-group reasoning, it identifies potential relationships between entities across different groups by considering both the relations mentioned in the query and those present in the KG. Additionally, GIVE introduces intermediate node groups to facilitate multi-hop reasoning necessary for complex questions, effectively bridging gaps in sparse KGs. By prompting the LLM to assess and reason about these possible relationships—including counterfactual reasoning where the model considers both the presence and absence of certain relations—GIVE builds an augmented reasoning chain. This chain combines factual knowledge from the KG with extrapolated inferences from the LLM, enabling the generation of more accurate and faithful responses even when the available external knowledge is limited. Great insights by @UCBerkeley and @penn 00:00 Integrate LLM and Knowledge Graphs 01:06 Think on Graph (ToG) 05:58 ToG GitHub code repo 06:30 GIVE Graph Inspired Veracity Extrapolation 09:33 GIVE vs Harvard Knowledge Graph Agent 10:33 Why RAG fails in Knowledge Graphs 11:40 Example of GIVE in detail 16:16 Compare ToG to GIVE All rights w/ authors: THINK-ON-GRAPH: DEEP AND RESPONSIBLE REASON- ING OF LARGE LANGUAGE MODEL ON KNOWLEDGE GRAPH https://arxiv.org/pdf/2307.07697 GIVE: STRUCTURED REASONING WITH KNOWLEDGE GRAPH INSPIRED VERACITY EXTRAPOLATION https://arxiv.org/pdf/2410.08475v1 Towards Trustworthy Knowledge Graph Reasoning: An Uncertainty Aware Perspective https://arxiv.org/pdf/2410.08985 #airesearch #aiagents #harvarduniversity #berkeley #knowledge #llm