semantica/docs/reference/reasoning.md

Reasoning

Advanced inference engine supporting Rule-based, SPARQL, Abductive, and Deductive reasoning strategies.

---

🎯 Overview

• :material-graph-outline:{ .lg .middle } Rule-Based Inference

  ---

  Forward and Backward chaining with Rete algorithm optimization

• :material-database-search:{ .lg .middle } SPARQL Reasoning

  ---

  Query expansion and property chain inference

• :material-lightbulb-question:{ .lg .middle } Abductive Reasoning

  ---

  Generate hypotheses to explain observations (Sherlock Holmes style)

• :material-check-decagram:{ .lg .middle } Deductive Reasoning

  ---

  Logical proof generation and theorem proving

• :material-text-box-search:{ .lg .middle } Explanation

  ---

  Generate natural language explanations for inferred facts

• :material-flash:{ .lg .middle } Rete Algorithm

  ---

  High-performance pattern matching for large rule sets

!!! tip "When to Use" - Inference: Deriving new facts from existing data (e.g., Parent(A,B) & Parent(B,C) -> Grandparent(A,C)) - Query Expansion: Finding results that aren't explicitly stored but implied - Hypothesis Generation: Finding potential causes for an observed event - Validation: Checking logical consistency of the knowledge graph

---

⚙️ Algorithms Used

Rule-Based Inference

• Forward Chaining: Data-driven. Apply rules to facts to derive new facts until saturation.
• Backward Chaining: Goal-driven. Start from a goal and work backward to find supporting facts.
• Bidirectional Chaining: Meet-in-the-middle strategy for complex paths.

Rete Algorithm

• Alpha Nodes: Filter facts by single attributes (e.g., type=Person).
• Beta Nodes: Join results from Alpha nodes (e.g., Person.id == Parent.child_id).
• Memory: Stores partial matches to avoid re-computation.
• Conflict Resolution: Priority-based selection when multiple rules match.

SPARQL Reasoning

• Query Rewriting: Modifying queries to include inferred patterns.
• Property Paths: Handling transitive relationships (foaf:knows+).
• Materialization: Pre-computing inferred triplets for fast read performance.

Abductive Reasoning

• Hypothesis Generation: Finding rules where the conclusion matches the observation.
• Ranking: Scoring hypotheses by Simplicity, Plausibility, and Coverage.
• Consistency Check: Ensuring hypotheses don't contradict known facts.

---

Main Classes

InferenceEngine

General-purpose inference engine supporting multiple strategies.

Methods:

Method	Description	Algorithm
infer(facts, rules)	Derive new facts	Forward Chaining
query(goal, rules)	Check if goal is true	Backward Chaining

Example:

from semantica.reasoning import InferenceEngine, Rule

rules = [
    Rule("Grandparent", "Parent(x, y) & Parent(y, z) -> Grandparent(x, z)")
]
facts = ["Parent(Alice, Bob)", "Parent(Bob, Charlie)"]

engine = InferenceEngine()
inferred = engine.infer(facts, rules)
# Result: ["Grandparent(Alice, Charlie)"]

ReteEngine

High-performance pattern matching engine.

Methods:

Method	Description
build_network(rules)	Compile rule into network
add_fact(fact)	Propagate fact through network
match_patterns()	Get triggered rules

SPARQLReasoner

SPARQL-based reasoner for RDF graphs.

Methods:

Method	Description
expand_query(query)	Rewrite query with inference
infer_results(result)	Add inferred triplets to result

AbductiveReasoner

Generates explanations for observations.

Methods:

Method	Description
generate_hypotheses(observations)	Generate hypotheses
rank_hypotheses(hyps)	Score and sort

Example:

from semantica.reasoning import AbductiveReasoner

reasoner = AbductiveReasoner(rules)
hypotheses = reasoner.generate_hypotheses(["Pavement is wet"])
# Result: ["It rained", "Sprinkler was on"]

ExplanationGenerator

Explains why a fact was inferred.

Methods:

Method	Description
generate_explanation(fact)	Generate reasoning trace
show_reasoning_path(trace)	Graph visualization

---

Convenience Functions

from semantica.reasoning import forward_chain, backward_chain, generate_explanation

# Quick inference
new_facts = forward_chain(facts, rules)

# Explain result
explanation = generate_explanation(new_facts[0], rules)
print(explanation.text)

---

Configuration

Environment Variables

export REASONING_MAX_ITERATIONS=100
export REASONING_STRATEGY=rete
export REASONING_TIMEOUT=30

YAML Configuration

reasoning:
  default_strategy: rete
  max_depth: 10
  
  rete:
    node_sharing: true
    
  abductive:
    max_hypotheses: 5

---

Integration Examples

Knowledge Graph Enrichment

from semantica.reasoning import InferenceEngine, Rule
from semantica.kg import KnowledgeGraph

# 1. Define Ontology Rules
rules = [
    Rule("SymmetricSibling", "Sibling(x, y) -> Sibling(y, x)"),
    Rule("TransitiveAncestor", "Ancestor(x, y) & Ancestor(y, z) -> Ancestor(x, z)")
]

# 2. Load Graph
kg = KnowledgeGraph()
facts = kg.get_all_triplets()

# 3. Run Inference
engine = InferenceEngine()
inferred_triplets = engine.infer(facts, rules)

# 4. Update Graph
kg.add_triplets(inferred_triplets)

---

Best Practices

1. Limit Recursion: Be careful with recursive rules (e.g., A(x,y) -> A(y,x)) which can cause infinite loops in naive implementations.
2. Use Rete for Scale: For >100 rules or >10k facts, always use the Rete engine.
3. Materialize vs. Query: Materialize (pre-compute) for read-heavy workloads; Query-rewrite for write-heavy workloads.
4. Validate Rules: Ensure rules are logically consistent to avoid exploding the fact space.

---

See Also

• Ontology Module - Source of schema-based rules
• Triplet Store Module - Backend for SPARQL reasoning
• Modules Guide - Consistency checking overview

Cookbook

• Reasoning and Inference