Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Scaling Ontologies with Language Models and Math

Scaling Ontologies with Language Models and Math

Slides presented at the ONTOCHAIN Summit for Trustworthy Internet by Marc Green, Software Architect in Cybersecurity


June 01, 2022

More Decks by ONTOCHAIN

Other Decks in Technology


  1. Problems • Humanity’s corpus of scientific knowledge is not yet

    robustly linked • Ontology engineering is hard, time-consuming, and expensive [email protected] Marc Green • Ologs provide mathematical robustness to ontologies • Language Models increase efficiency and accessibility of knowledge workflows Solutions
  2. Intro to Ologs (ontology logs) [email protected] - “Rigorous mathematical framework

    for knowledge representation, construction of scientific models, and data storage” (Spivak, 2011) - More scalable than traditional semantic web standards like RDF Marc Green (Spivak, 2011) (Spivak, 2011)
  3. Intro to Language Models (LMs) - Enable a new paradigm

    of Human-Computer Interaction - B2C for closed domains (chatbots), B2B for open domains (ai assistants) [email protected] Marc Green https://huggingface.co/blog/large-language-models
  4. Ontochain application idea: Ontology AI Assistant - Public service to

    increase access to and utility of shared understandings - Key Ideas: - Language Models increase accessibility and improve efficiency of working with text - Ologs provide rigorous and scalable mechanism for building shared understandings - Potential features: - Natural Language UI for interacting with ontologies (querying, editing, aligning, etc) - SotA NLP for systematic generation and manipulation of ontologies (eg from existing literature) - Integrates with decentralized network, knowledge graph and knowledge/asset market - Modular architecture allows app to scale as LM tech improves [email protected] Marc Green
  5. Other References and Related Work • Spivak, 2011: Ologs: a

    categorical framework for knowledge representation (arXiv:1102.1889) • “23% searching …”: https://topos.institute/networked-mathematics • Training and Application of Neural-Network Language Model for Ontology Population (https://www.researchgate.net/publication/347861342_Training_and_Applicati on_of_Neural-Network_Language_Model_for_Ontology_Population) • Ontology-Based Model Abstraction (https://www.researchgate.net/publication/332290797_Ontology-Based_Mode l_Abstraction) • Semantic Scholar by Allen Institute of AI • MathFoldR by Topos Institute [email protected] Marc Green