EvoRM: Evolvable Neuro-Symbolic Reasoning Framework

EvoRM is an evolvable neuro-symbolic reasoning framework for entity matching and schema matching. It combines the reasoning capabilities of LLMs with the efficiency of symbolic rules, achieving high accuracy with reduced LLM calls.

Architecture

EvoRM consists of four core components:

1. RuleEncoding — Distills LLM decisions into structured FOL (First-Order Logic) rules
2. HypergraphStorage — Weighted hypergraph for rules + entity context indexing
3. TwoStageInferenceController — Stage 1: symbolic filtering (fast), Stage 2: LLM judgment (accurate)
4. RuleMaintenance — Dynamic confidence tracking + periodic rule optimization

Plus an optional MLP Gate (gϕ) for neural routing between stages.

Key Features

• Plug-and-play: Integrates with any LLM-based entity matching baseline
• Two-stage inference: Rules route high-confidence decisions, saving LLM calls
• Continual learning: Rules evolve with each inference, building a knowledge base
• Confidence boost: Rules gain confidence as they are triggered correctly

Installation

pip install -r requirements.txt

Quick Start

from evorm_core import EvoRMPlugin

# Initialize EvoRM plugin
plugin = EvoRMPlugin(
    client=openai_client,
    persistence_dir="./evorm_state",
)

# Use with your entity matching pipeline
decision, triggered, candidates = plugin.stage1(eid_a, eid_b, es_ctx, et_ctx)
if decision is not None:
    # Stage 1 routed (high confidence)
    pass
else:
    # Stage 2: LLM judgment
    result = plugin.stage2_simple(eid_a, eid_b, es_ctx, et_ctx, triggered, prompt)
    decision = result['decision']

Supported Baselines

EvoRM provides wrappers for 6 baseline methods:

Baseline	Task	Wrapper
MatchGPT	Entity Resolution	baselines/evorm_wrappers/matchgpt.py
Anymatch	Entity Resolution	baselines/evorm_wrappers/anymatch.py
LELA	Entity Linking	baselines/evorm_wrappers/lela_el.py
ChatEA	Entity Alignment	baselines/evorm_wrappers/chat_ea.py
Matchmaker	Schema Matching	baselines/evorm_wrappers/cohard.py
AdaCoAgentEA	Entity Alignment	baselines/evorm_wrappers/zero_cot.py

Datasets

The framework supports 14 datasets across 4 tasks:

• Entity Resolution (ER): Amazon-Google, Beer, DBLP-ACM, DBLP-GoogleScholar, Fodors-Zagats, iTunes-Amazon, Walmart-Amazon, Abt-Buy
• Entity Linking (EL): ZESHEL (FR, Lego, ST, YG)
• Schema Matching (SM): MIMIC, SYNTHEA
• Entity Alignment (EA): DBP-WIKI-V1, DBP-WIKI-V2

Results (V5)

Entity Resolution (MatchGPT)

Dataset	Baseline F1	+EvoRM F1	Δ	S1 Rate
DBLP-ACM	1.0000	1.0000	0.00	16.7%
DBLP-GoogleScholar	0.9474	0.9474	0.00	13.3%
Walmart-Amazon	0.9714	0.9714	0.00	13.3%
Amazon-Google	0.9756	0.9756	0.00	0%

Schema Matching

Dataset	Baseline	Baseline F1	+EvoRM F1	S1 Rate
MIMIC	matchmaker	0.9697	1.0000	40%
SYNTHEA	matchmaker	0.9697	1.0000	80%
SYNTHEA	rematch	0.9697	1.0000	66.7%

Configuration

from evorm_core import EvoRMConfig

config = EvoRMConfig(
    theta_hi=0.65,        # High-confidence match threshold
    theta_prune=0.75,     # Conservative non-match threshold
    route_nonmatch=True,  # Allow S1 to route non-match decisions
    alpha=0.6,            # Rule encoding weight
    beta=0.4,             # Context similarity weight
)

Citation

@article{evorm2024,
  title={EvoRM: Evolvable Neuro-Symbolic Reasoning for Entity Matching},
  journal={IEEE TKDE},
  year={2024}
}

License

MIT License