OntoAgent is our cognitive architecture supporting the development of LEIAs (Language, Embodiment, Intelligence, Agency). It provides a comprehensive framework for building agents capable of interpreting multiple input modalities, reasoning about goals and plans, and acting in real or simulated environments.
At its core, OntoAgent emphasizes content-centric cognitive modeling—the idea that intelligent behavior requires deep, semantically-grounded knowledge of the world, language, and self.
Architecture Overview
Perception
Receiving and interpreting stimuli from the environment—speech, text, vision, and other modalities.
Memory
The storehouse of all agent knowledge: ontology, lexicon, episodic memory, and agenda.
Action
Realizing change through effectors—physical actions, speech generation, and internal reasoning.
The architecture supports the full cycle of intelligent behavior: perceiving the world, updating beliefs, forming intentions, making plans, and executing actions—all grounded in explicit, inspectable knowledge representations.
OntoSem: Semantic Analysis
OntoSem (Ontological Semantics) is our approach to deep language understanding. Unlike surface-level NLP, OntoSem produces rich meaning representations grounded in a comprehensive ontology of world knowledge.
Processing Pipeline
- Basic Syntax — Initial parsing and morphological analysis
- OntoSyntax — Adapting syntax to ontological categories
- Semantic Mapping — Producing Text Meaning Representations (TMRs)
- Reference Resolution — Linking to episodic memory and world knowledge
- Situational Reasoning — Contextual interpretation and inference
Key Features
Ontologically-Grounded
All meanings are expressed using concepts from a rich ontology, enabling deep reasoning and cross-modal integration.
Disambiguation
Systematic resolution of lexical, syntactic, and semantic ambiguities through constraint satisfaction and world knowledge.
Incremental Processing
Multi-stage analysis that builds understanding progressively, allowing early commitment and revision.
Context-Sensitive
Deep integration of discourse context, speaker models, and situational knowledge.
OntoGen: Language Generation
OntoGen is our natural language generation component, producing fluent text from semantic meaning representations. It reverses the OntoSem pipeline, transforming TMRs into natural language.
Generation Process
- Content Selection — Deciding what information to convey
- Text Planning — Organizing content into coherent structure
- Lexicalization — Selecting words to express concepts
- Surface Realization — Producing grammatical output
Applications
Explanation Generation
Producing natural language explanations of agent reasoning and decisions.
Dialogue Response
Generating contextually appropriate responses in conversation.
Report Generation
Summarizing observations and plans for human collaborators.
Paraphrase
Expressing the same meaning in different linguistic forms.
Design Principles
Explainability First
Every agent decision is traceable to explicit knowledge and reasoning steps that can be explained to humans.
Hybrid Integration
Symbolic reasoning augmented by statistical methods where appropriate—LLMs as tools, not foundations.
Lifelong Learning
Agents designed to acquire knowledge continuously through reading and interaction.
Metacognition
Agents that know what they know and can reason about their own capabilities and limitations.
Learn More
Explore our publications for detailed technical descriptions, or visit the demos page to see OntoAgent in action.
View Publications →