AGISystem2 – NL→DSL Grammar-Based Translation

Research Overview

Goal: Extend the NL2DSL grammar engine to handle 95%+ of structured natural language patterns deterministically, without requiring LLM assistance.

Current Status: ~70% coverage of academic benchmark patterns

Target: 95% coverage by Q4 2025

1. Current Architecture

The grammar-based translator uses a multi-stage pipeline:

Natural Language Input
        │
        ▼
┌───────────────────┐
│ Sentence Splitter │  ← Split complex sentences
└───────────────────┘
        │
        ▼
┌───────────────────┐
│ Pattern Matcher   │  ← Match against grammar rules
└───────────────────┘
        │
        ▼
┌───────────────────┐
│ Structure Builder │  ← Build DSL AST
└───────────────────┘
        │
        ▼
┌───────────────────┐
│ DSL Serializer    │  ← Output DSL text
└───────────────────┘

2. Pattern Categories

Category	Examples	Coverage	Priority
IS-A Relations	"X is a Y", "Every X is a Y"	95%	Done
HAS-A Relations	"X has Y", "X contains Y"	90%	Done
Simple Conditionals	"If X then Y"	85%	Done
Quantifiers	"All X", "Some X", "No X"	60%	P1
Temporal	"Before X", "After Y", "During Z"	50%	P1
Comparatives	"X is more than Y", "X is the most Z"	30%	P2
Nested Conditions	"If X and Y then Z unless W"	25%	P2

3. Priority 1: Quantifier Expansion

Challenge: Natural language quantifiers are nuanced and context-dependent.

"All cats are mammals" → Universal quantifier
"Some birds can fly" → Existential quantifier
"Most students passed" → Majority quantifier (not FOL)
"Few exceptions exist" → Minority quantifier

Approach: Implement a quantifier normalizer that:

Identifies quantifier phrases via pattern matching
Maps to formal quantifier types (∀, ∃, majority, etc.)
Generates appropriate DSL with variable bindings

// Input
"All mammals have a heart"

// Parsed structure
{
  quantifier: "universal",
  variable: "?x",
  domain: "Mammal",
  predicate: { op: "has", subject: "?x", object: "heart" }
}

// Output DSL
@r1 Implies (isA ?x Mammal) (has ?x heart)

4. Priority 1: Temporal Expressions

Challenge: Temporal relations require understanding of:

Point-in-time references ("in 1990", "yesterday")
Interval relations ("during the meeting", "throughout the year")
Ordering ("before X", "after Y", "between A and B")
Duration ("for 3 hours", "since Monday")

Approach: Use Allen's interval algebra as the formal foundation:

Extract temporal expressions via regex + NLP patterns
Normalize to interval representation
Map to DSL temporal operators (before, after, during, overlaps)

5. Implementation Plan

Phase	Deliverable	Timeline
Phase 1	Universal/existential quantifiers (∀, ∃)	Q1 2025
Phase 2	Temporal point references	Q1 2025
Phase 3	Temporal intervals (Allen relations)	Q2 2025
Phase 4	Comparative structures	Q2 2025
Phase 5	Nested conditionals with exceptions	Q3 2025
Phase 6	Edge case hardening, 95% target	Q4 2025

6. Evaluation Metrics

Pattern coverage: % of benchmark sentences parsed correctly
Semantic accuracy: % of DSL outputs that preserve meaning
Parse time: Average ms per sentence
Error clarity: % of failures with actionable error messages

7. Related Work

NL2DSL Translation Layer – Current implementation
LLM-Assisted Translation – Complementary approach
External Benchmarks – Evaluation results