Knowledge Engineering, Ontologies & Semantic Reasoning Suite

Comprehensive Suite Spanning Macroeconomic Ontologies (Neo4j/OWL), Multi-Hop LLM Graph Reasoning (GPT-4.1), and Lexical Semantic Engineering (FarsNet & NetworkX).

Overview

Knowledge Representation and Reasoning (KRR) forms the bedrock of symbolic artificial intelligence, bridging unstructured natural language data and deterministic relational logic. Constructing machine-interpretable schemas enables deep multi-hop reasoning, transparent inference path verification, and context grounding that mitigates hallucinations in generative Large Language Models (LLMs).

This project consolidates an advanced Knowledge Engineering and Semantic Reasoning Suite spanning three interconnected domains: macroeconomic ontology modeling in Neo4j and W3C OWL/RDF, empirical evaluation of LLM graph reasoning (GPT-4.1) across multi-hop biomedical and scientific taxonomies, and algorithmic traversal of the FarsNet Persian lexical semantic database using NetworkX.

Core Systems & Modular Breakdown

1. Macroeconomic Ontology & Geopolitical Knowledge Graph

  • Domain Modeling: Captures the complex geopolitical and macroeconomic determinants governing global gold price fluctuations from the outbreak of the Russia-Ukraine conflict onward.
  • Graph Architecture: Formulates an interconnected graph schema consisting of 48 conceptual/instance nodes and 64 directed relationship edges (e.g., positivelyCorrelatedWith, causedBy, sanctionedBy).
  • Automated Cypher Pipeline: Implements custom parser scripts (json_to_cypher.py) transforming Neo4j Arrows vector exports into transactional, idempotent MERGE Cypher queries.
  • Semantic Web Export: Converts graph stores into standard W3C ontologies (.owl / .rdf) via rdflib (csv_to_owl.py), ensuring SPARQL query compliance and formal Description Logic (DL) expressivity.
  • Competency Queries: Traces multi-variable paths assessing the systemic economic impacts of the 2024 US Presidential Election, crude oil shocks, inflation surges, and regional conflicts.

2. Multi-Hop Knowledge Graph Reasoning & LLM Evaluation

  • Evaluation Framework: Rigorously benchmarks the structural inference capabilities of GPT-4.1 over structured knowledge graphs across four distinct domains: CSO (Computer Science), FMA (Human Anatomy), Pleiades (Ancient Geography), and RxNorm (Pharmacology).
  • Triplet Extraction: Automates the mining and partitioning of (Subject, Predicate, Object) triples into strict taxonomic hierarchies (is_a) and non-taxonomic relational properties (other).
  • Chain-of-Reasoning Benchmarking: Evaluates balanced verification questions across multiple reasoning depths:
    • Single-Hop: Direct taxonomic subclass verification.
    • Multi-Hop (Long Chain): Deep compositional paths spanning 4 to 7 transitive hops.
    • Non-Taxonomic Relations: Multi-predicate cross-entity relations.
  • Empirical Prompting Analysis: Demonstrates that while zero-shot parametric prompting degrades toward random-guess performance (~50%) on deep transitive chains, injecting explicit sub-graph context into prompts elevates retrieval and verification accuracy to ~100%.

3. Computational Lexical Semantics on FarsNet

  • Lexical Taxonomy Traversal: Implements a computational semantic network engine over FarsNet (the Persian WordNet lexical ontology), querying hypernymy, hyponymy, meronymy, and domain associative linkages.
  • Graph Traversal & Search: Utilizes NetworkX to construct directed graph representations from structured relational fact triples (farsnet_facts.tsv).
  • Automated Inference: Implements shortest-path algorithms, transitive closure computation, and taxonomic distance metrics to verify lexical semantic entailment.

Key Technical Highlights

  • Graph Database Orchestration: Native Neo4j administration, Cypher index optimization, and binary dump restoration (hwii.dump).
  • Semantic Web Standards: Construction of serialized RDF/XML and OWL schemas adhering to W3C Semantic Web compliance.
  • Parametric vs. Non-Parametric LLM Profiling: Systematic benchmarking demonstrating the limitations of internal model weights on long-chain graph deductions without external knowledge augmentation.
  • Network Algorithms: Graph construction, cycle detection, transitive reduction, and topological pathfinding across dense lexical knowledge bases.

Technical Stack

  • Languages & Core: Python 3.8+, Pandas, NumPy
  • Graph Databases & Querying: Neo4j, Cypher Query Language, Arrows.app
  • Semantic Web & Ontologies: RDFLib, OWL (Web Ontology Language), RDF/XML
  • Graph Analytics: NetworkX, Matplotlib
  • LLM Benchmarking & Evaluation: GPT-4.1, Automated Prompt Templating, Jupyter Notebooks
  • Domain Focus: Knowledge Representation & Reasoning (KRR), Knowledge Graphs, Lexical Semantics, LLM Reasoning