AGISystem2 – NL2DSL Translation

1. Architecture

The NL2DSL pipeline has three main stages:

Natural Language Input
        │
        ▼
┌───────────────────┐
│ 1. Sentence Split │  Split text into individual sentences
└───────────────────┘
        │
        ▼
┌───────────────────┐
│ 2. Pattern Match  │  Match against rule/fact/copula/relation patterns
└───────────────────┘
        │
        ▼
┌───────────────────┐
│ 3. DSL Emit       │  Generate DSL statements with references
└───────────────────┘
        │
        ▼
    DSL Output

2. Supported Patterns

2.1 Rule Patterns Rules

Rules express conditional logic (IF-THEN, universal quantification).

Input:  "If something is a mammal then it is warm-blooded"
  →
Output: @r1 isA ?x Mammal
        @r2 hasProperty ?x warm_blooded
        Implies $r1 $r2

2.2 Fact Patterns Facts

Facts assert ground truths about specific entities.

Input:  "Marie Curie is a scientist"
  →
Output: isA Marie_Curie Scientist

2.3 Copula Patterns Copula

Copula patterns handle "is/are/was/were" constructions with various complements.

Input:  "The ball is in the box"
  →
Output: in Ball Box

2.4 Relation Patterns Relations

Relation patterns handle verb-based relationships between entities.

3. Compound Structures

3.1 Coordination (AND/OR)

Input:  "Tom is a cat and Jerry is a mouse"
  →
Output: isA Tom Cat
        isA Jerry Mouse

Input:  "If X is red and X is round then X is colorful"
  →
Output: @r1 hasProperty ?x red
        @r2 hasProperty ?x round
        @r3 And $r1 $r2
        @r4 hasProperty ?x colorful
        Implies $r3 $r4

3.2 Negation

Input:  "Penguins are not able to fly"
  →
Output: Not (hasProperty Penguin able_to_fly)

3.3 Quantified Subjects

Input:  "Every red thing is colorful"
  →
Output: @r1 hasProperty ?x red
        @r2 hasProperty ?x colorful
        Implies $r1 $r2

4. Entity Normalization

NL2DSL automatically normalizes entities for consistent representation:

5. Operator Auto-Declaration

Input:  "John supervises Mary"  (unknown operator "supervise")
  →
Output: @supervise:supervise __Relation   # Auto-declared
        supervise John Mary

Enable with: { autoDeclareUnknownOperators: true }

6. Fallback: Opaque Statements

When a sentence cannot be parsed, NL2DSL can create an opaque reference (hash-based) to preserve the information without losing it:

Input:  "The quantum entanglement defies classical intuition"  (unparseable)
  →
Output: hasProperty KB opaque_ctx_a1b2c3d4e5  # Hash-based reference

Enable with: { fallbackOpaqueStatements: true }

7. API Usage

import { translateContextWithGrammar } from './src/nlp/nl2dsl/grammar.mjs';

const text = `
  All cats are mammals.
  All mammals are animals.
  Tom is a cat.
`;

const result = translateContextWithGrammar(text, {
  autoDeclareUnknownOperators: true,
  fallbackOpaqueStatements: false
});

console.log(result.dsl);
// Output:
// @r1 isA ?x Cat
// @r2 isA ?x Mammal
// Implies $r1 $r2
// @r3 isA ?x Mammal
// @r4 isA ?x Animal
// Implies $r3 $r4
// isA Tom Cat

console.log(result.stats);
// { sentencesTotal: 3, sentencesParsed: 3, sentencesOpaque: 0, autoDeclaredOperators: 0 }

console.log(result.errors);
// [] (empty if all parsed successfully)

8. Known Limitations

9. Pattern Priority

Patterns are matched in this priority order:

1. Rule patterns (IF-THEN, quantifiers) – checked first
2. Fact patterns (ground assertions) – checked second
3. Copula patterns (is/are constructions) – within fact parsing
4. Relation patterns (verb-based) – fallback
5. Opaque fallback (if enabled) – last resort

10. Performance