Klareco

Experimental

An Esperanto conversational-AI research project that prioritizes deterministic, explainable processing. It encodes regular grammar programmatically and reserves learned models for semantic reasoning.

🔗 GitHub Repository

Overview

Modern large language models are impressive but opaque. They learn everything—including grammar—from data, making their behavior difficult to predict or explain. Klareco asks: what if we separated the learnable from the deterministic?

Esperanto, with its completely regular grammar (16 rules, no exceptions), is the perfect testbed. We can encode grammar directly in code and reserve machine learning for the genuinely hard part: meaning.

Features

  • Esperanto's 16 grammar rules encoded directly in the parser, not learned
  • Deterministic parser achieves 91.8% parse rate on Esperanto text
  • Compositional lexicon targets 18,928 Esperanto roots through root and affix embeddings
  • Corpus index spans 5.3M Esperanto sentences
  • Working RAG query system combines AST-aware search with neural reranking
  • ~5.5M learned parameters support the current semantic pipeline; grammar remains programmatic
  • Research status - the parser and RAG system work, while vocabulary and semantic-model limitations remain active research work

Architecture

Input Text
    ↓
[Deterministic Tokenizer]
    ↓
[Rule-Based Parser] ← Esperanto's 16 rules (91.8% parse rate on 4.2M sentences)
    ↓
Abstract Syntax Tree
    ↓
[Compositional Embeddings + Reranking] ← Learned (~5.5M params)
    ↓
Semantic Representation
    ↓
[Deterministic Generator]
    ↓
Output Text

Why Esperanto?

Natural languages are messy. English alone has:

  • Irregular verbs (go/went, be/was/were)
  • Context-dependent parsing (time flies like an arrow)
  • Exceptions to every rule

Esperanto was designed to be regular:

  • All verbs conjugate identically
  • Word class is marked by ending (-o noun, -a adjective, -e adverb, -i verb infinitive)
  • No irregular forms

This regularity means a hand-coded parser can handle 100% of Esperanto grammar—something impossible for natural languages.

Research Questions

  1. How much of language model capability comes from grammar vs. semantics?
  2. Can explicit structure reduce the parameters needed for competent language use?
  3. Does deterministic grammar improve explainability without sacrificing capability?

Klareco investigates these questions through a working implementation that separates learned semantic reasoning from programmatic linguistic structure.