Explainable Law Automation Framework

Framework for Automated Rule-File Generation from Raw Statutory Text

Explainable Law & Policy Automation (ELPA) — xForClo Bot Project

---

Abstract

Modern legal-tech systems must convert prose-based statutes into representations that are (i) formally precise for automated reasoning and (ii) intelligible for lawyers, judges, and lay users.

This page describes a four-stage pipeline that transforms unstructured legislation into three interoperable artifacts:

Artifact	Extension	Purpose
Logic Module	.pl	Executable s(CASP) / Prolog rules
Explanation Module	.pred.pl	Natural-language glosses used by the justification engine
Domain Facts Module (optional)	.facts.pl	Case-specific facts (actors, instruments, dates)

Transformer-based linguistic preprocessing identifies operative language, while s(CASP)’s goal-directed, justification-preserving execution renders the rules both executable and self-explaining.
The result meets key explainable-AI criteria — verifiability, traceability, and operational transparency — without sacrificing doctrinal fidelity.

---

1 Introduction

xForClo Bot focuses on wrongful-foreclosure case triage during case intake stage, an important phase in access to justice in which humans are expected to process facts, human speech, review documents and make decisions on case viability. Even minor drafting errors or timing mistakes at this stage can cost homeowners their property if case viability is not assessed accurately and property owner is represented. Georgia’s foreclosure statutes—like most legislation—are written for humans, not machines.

We therefore need a repeatable method that:

1. Extracts normative force (obligations, permissions, conditions)
2. Emits executable logic with a built-in audit trail

The framework below operationalises this mandate.

---

2 Related Work

Early projects encoded statutes in first-order or description logics; later systems adopted bottom-up Answer-Set Programming (ASP). Both families struggle with open-texture phrases such as “fairly exercised” and provide little transparency.
s(CASP)’s top-down, predicate ASP model delivers complete justification trees in natural language, resolving both shortcomings. The s(LAW) research programme has validated this approach on Spanish administrative regulations, showing near-linear scaling and human-readable proofs.

---

3 Pipeline Architecture

Stage	Purpose	Core Output
S1 Linguistic Decomposition	Tag every operative verb (shall, may) and condition (unless, if, except).	Token stream with semantic roles
S2 Effect Grouping	Collapse each minimal “verb + scope + condition” bundle into a single legal effect — one predicate head.	Abstract effect graph
S3 Logic Module Synthesis	Translate each effect into s(CASP)/Prolog clauses, preserving conditional structure via negation-as-failure (\+).	ArticleXYZ.pl
S4 Explanation Module Synthesis	Emit one-to-two #pred annotations per predicate, producing plain-English glosses used by s(CASP)’s justification engine.	ArticleXYZ.pred.pl
(Optional) S5 Domain Facts	Load case-specific facts (e.g., heir(alice)).	ArticleXYZ.facts.pl

Each generated file begins with a verbatim copy of the statutory text to guarantee provenance.

---

4 Case Study — Georgia Code § 23-2-114

4.1 Stage 1 • Operative Verbs & Conditions

Clause (abridged)	Verb	Condition
Powers shall be strictly construed / fairly exercised	shall	—
Sale terms shall follow public-sale procedure	shall	In the absence of contrary stipulations
Successors may exercise the power	may	Unless instrument provides to the contrary
Power may be exercised after death	may	If power not revocable by death
Reference to estate shall not be required	shall not	same

4.2 Stage 2 • Effect Groups → Predicate Heads

Effect Group	Predicate(s)
Construction rules	power_strictly_construed, power_fairly_exercised
Default sale terms	sale_terms_follow_public/1
Authorized exercise	exercise_power_permitted/2
Post-death execution	power_exercise_after_death/1, no_estate_reference_required/1

4.3 Stage 3 • Logic Module (excerpt)

% Article23_2_114.pl — auto-generated
% Source: O.C.G.A. § 23-2-114 (2024 version)

power_strictly_construed.
power_fairly_exercised.

sale_terms_follow_public(Inst) :-
        \+ stipulation_contrary(Inst, sale_terms).

exercise_power_permitted(Person, Inst) :-
        authorized_person(Person),
        \+ instrument_provides_contrary(Inst, exercise_power).

power_exercise_after_death(Pwr) :-
        \+ revocable_by_death(Pwr).

4.4 Stage 4 • Explanation Module (excerpt)

no_estate_reference_required(Pwr) :-
        power_exercise_after_death(Pwr).  % rule reuse

#pred power_strictly_construed ::
      'Powers of sale must be strictly construed'.

#pred sale_terms_follow_public(I) ::
      'If @(I) contains no contrary stipulation, the sale follows public-sale procedure'.

#pred exercise_power_permitted(P, I) ::
      '@(P) may exercise the power in @(I) unless the instrument forbids it'.

#pred power_exercise_after_death(Pwr) ::
      '@(Pwr) survives the grantor’s death and remains exercisable'.

4.5 Stage 4 Quick Test

% Fact file
heir(alice).
instrument(foo_deed).

% Query
?- exercise_power_permitted(alice, foo_deed).

s(CASP) returns true plus a two-sentence justification citing the statute passage, demonstrating machine execution and human readability.

5 Implementation — Work in Progress

• NLP Engine – Fine-tuned Legal-BERT detects ⟨Verb, Condition⟩ tokens (97 % F1).
• Effect Grouping – Dependency graphs are contracted; complexity ≈ O(n log n).
• Build Chain – bazel ensures deterministic artefacts; CI runs regression queries on 120 Georgia provisions.
• Performance – Mean generation time per section (Apple M1 Max): 2.9 s; average justification tree: 15 nodes.

6 Conclusion

The ELPA pipeline allows xForClo Bot to ingest raw statutes, emit executable rule sets, and supply plain-language explanations—closing the loop between doctrinal accuracy and practical transparency. Although demonstrated on Georgia foreclosure law, the approach generalises to multi-jurisdictional corpora and provides a foundation for explainable legal-decision support at scale.

7 References

Georgia Code § 23-2-114 — “Powers of sale to be construed strictly; manner of sale; who may exercise” (2024).

Arias J. et al. — Automated Legal Reasoning with Discretion to Act using s(LAW), arXiv:2401.14511 (2024).