Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BambooKG: A Neurobiologically-inspired Frequency-Weight Knowledge Graph

Vanya Arikutharam, Arkadiy Ukolov

TL;DR

BambooKG tackles the challenge of robust, multi-hop reasoning across documents by replacing rigid triplet-centric graphs with a frequency-weighted, non-triplet associative memory graph built from chunk-level tags. It separates memorisation (chunking, tagging, and building a global tag-based graph) from recall (query-tag extraction, subgraph retrieval, and context construction) to form episodic, query-relevant context for an LLM. Empirical results on HotPotQA and MuSiQue show BambooKG achieves higher data recall and faster retrieval than RAG, OpenIE, GraphRAG, and KGGen, due to stronger cross-document connectivity and a retrieval pipeline that avoids heavy embedding-based or triplet-centric constraints. The work presents a neurobiologically inspired alternative for scalable, multi-hop knowledge retrieval, with practical implications for reducing hallucinations and ageing in retrieval-augmented generation.

Abstract

Retrieval-Augmented Generation allows LLMs to access external knowledge, reducing hallucinations and ageing-data issues. However, it treats retrieved chunks independently and struggles with multi-hop or relational reasoning, especially across documents. Knowledge graphs enhance this by capturing the relationships between entities using triplets, enabling structured, multi-chunk reasoning. However, these tend to miss information that fails to conform to the triplet structure. We introduce BambooKG, a knowledge graph with frequency-based weights on non-triplet edges which reflect link strength, drawing on the Hebbian principle of "fire together, wire together". This decreases information loss and results in improved performance on single- and multi-hop reasoning, outperforming the existing solutions.

BambooKG: A Neurobiologically-inspired Frequency-Weight Knowledge Graph

TL;DR

BambooKG tackles the challenge of robust, multi-hop reasoning across documents by replacing rigid triplet-centric graphs with a frequency-weighted, non-triplet associative memory graph built from chunk-level tags. It separates memorisation (chunking, tagging, and building a global tag-based graph) from recall (query-tag extraction, subgraph retrieval, and context construction) to form episodic, query-relevant context for an LLM. Empirical results on HotPotQA and MuSiQue show BambooKG achieves higher data recall and faster retrieval than RAG, OpenIE, GraphRAG, and KGGen, due to stronger cross-document connectivity and a retrieval pipeline that avoids heavy embedding-based or triplet-centric constraints. The work presents a neurobiologically inspired alternative for scalable, multi-hop knowledge retrieval, with practical implications for reducing hallucinations and ageing in retrieval-augmented generation.

Abstract

Retrieval-Augmented Generation allows LLMs to access external knowledge, reducing hallucinations and ageing-data issues. However, it treats retrieved chunks independently and struggles with multi-hop or relational reasoning, especially across documents. Knowledge graphs enhance this by capturing the relationships between entities using triplets, enabling structured, multi-chunk reasoning. However, these tend to miss information that fails to conform to the triplet structure. We introduce BambooKG, a knowledge graph with frequency-based weights on non-triplet edges which reflect link strength, drawing on the Hebbian principle of "fire together, wire together". This decreases information loss and results in improved performance on single- and multi-hop reasoning, outperforming the existing solutions.

Paper Structure

This paper contains 18 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Existing Models
  3. RAG
  4. OpenIE
  5. GraphRAG
  6. KGGen
  7. BambooKG: Architecture and Pipelines
  8. Memorisation Pipeline
  9. Chunking Stage
  10. Tag Generation Stage
  11. Knowledge Graph Creation Stage
  12. Recall Pipeline
  13. Query Tag Extraction
  14. Subgraph Retrieval
  15. Context Construction
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Overall memory pipeline: (a) Chunking, (b) Tag Generation, and (c) Knowledge Graph Creation stages (left - tag to chunk mapping KG and right - tag frequency weight KG)
  • Figure 2: Query Subgraph Generation. For illustrative purposes we take the top 2 first degree neighbours and the top 2 second degree neighbour for each query node. Note: some connections are not shown to ensure the readability of the diagram
  • Figure 3: Stage 2 of the Recall Pipeline: (a) Chunk extraction from query subgraph based on tags obtained from Stage 1 (b) Final Context