---
title: 'Dual-Laws Model: Comparative Law Framework'
url: https://www.emergentmind.com/topics/dual-laws-model-6322014f-cf9a-4aa0-a47b-d76667e7e419
type: topic
---

# Dual-Laws Model: Comparative Law Framework

The Dual-Laws Model is an integrated framework for computational comparative law that represents, aligns, and analyzes pairs of corresponding legal provisions across jurisdictions by unifying two layers: structural interoperability and semantic correspondence detection. In this formulation, Japan’s e-LAWS corpus is converted from the Japanese Legal Standard (JLS) XML schema into Akoma Ntoso (AKN, LegalDocML), and the resulting AKN-encoded provisions are then processed with multilingual semantic representations, semantic textual similarity techniques, FAISS retrieval, and transformer-based Cross-Encoder reranking to generate cross-jurisdictional candidate correspondences and visualize them as exploratory networks [2603.15094].

## 1. Conceptual definition and scope

Within the framework, a “correspondence” is an inter-jurisdictional relation between two provisions—typically articles, but possibly paragraphs or items—that share sufficiently similar legal meaning, function, or doctrinal role, as quantified by a semantic similarity score and bounded by jurisdiction-aware thresholds and quotas [2603.15094]. Correspondence is many-to-many in principle, but the system supports one-to-many linkage per source provision and constructs networks of these edges for exploration.

The model is explicitly dual-layered. Its first layer addresses structural interoperability: Japanese statutes in JLS XML are converted so they can be parsed, identified, and linked in international legislative databases. Its second layer addresses semantic correspondence detection: multilingual embeddings, semantic similarity modeling, efficient nearest-neighbor retrieval, and reranking are combined to surface candidate provision-level matches at scale. This layered design is central to the framework’s name: legal comparison is treated as both a document-structure problem and a semantic alignment problem.

A common misunderstanding is to treat a correspondence as an assertion of exact doctrinal equivalence. The framework does not do this. It defines correspondence operationally through retrieval, reranking, thresholds, and quotas; the resulting links are candidate correspondences for exploratory comparative analysis rather than conclusive statements of identity.

## 2. Structural interoperability through JLS-to-AKN conversion

The structural layer begins from the fact that Japan’s e-LAWS platform publishes all laws and regulations in JLS XML, developed by JaLII. JLS uses hierarchical tags such as `<Law>`, `<Article>`, and `<Paragraph>`, with drafting conventions tailored to Japanese legislation. The target representation is Akoma Ntoso, the OASIS LegalDocML standard, whose core hierarchy is `act → part → chapter → section → article → paragraph → clause/item`, and whose FRBR-based metadata encode Work, Expression, and Manifestation identities together with jurisdiction, language, and version date [2603.15094].

The conversion pipeline is organized into six stages. First, JLS XML is ingested from e-LAWS and parsed. Second, structural analysis determines each element’s logical level and resolves inconsistencies before conversion. Third, an XSLT transformation uses about 50 mapping rules, including `<Law> → <act>`, `<Article> → <article>`, and `<Paragraph> → <paragraph>`. Fourth, FRBR metadata are generated so that Work, Expression, and Manifestation identifiers embed jurisdiction, language, and version date. Fifth, the output is validated against the LegalDocML AKN schema. Sixth, converted documents can be rendered and edited with LIME Editor and Akomantoso Viewer [2603.15094].

Several properties of the conversion are central. Hierarchical ordering is preserved across levels from statute to article to paragraph or item. Article numbering and nested structure are retained. Citations and cross-references present in the text are preserved through conversion, although explicit network modeling of intra-code references is only supported conceptually and is not altered by the conversion pipeline itself. Temporal/versioning information is captured at the FRBR Expression level, while Manifestation identifies the concrete file, enabling timeline-aware comparisons. Stable identifiers and anchors are embedded through AKN IDs, ensuring resolvable URIs within AKN-compliant environments, although the exact URI schema is implementation-dependent.

Validation is reported conservatively: ten randomly selected laws passed schema validation. This establishes feasibility for standards-compliant transformation, not a claim of exhaustive corpus-level auditing.

## 3. Semantic correspondence detection pipeline

On top of the AKN representation, the framework implements semantic correspondence detection at the provision level. The reported prototype segments provisions at the article level, although AKN permits finer levels such as paragraphs and items, which the prototype can encode and extract. The working languages are Japanese, Korean, and French. No machine translation is required because the prototype uses `multilingual-e5-large`, a multilingual Transformer encoder trained with weakly supervised contrastive pretraining for retrieval-optimized sentence embeddings [2603.15094].

The paper states that `multilingual-e5-large` is used as-is and does not report further fine-tuning on legal corpora. It also does not specify tokenization, exact embedding dimensionality, or special preprocessing for definitions, stopwords, numerals, or references. Articles are embedded directly, and legal numbering and references remain part of the text, with structure preserved by AKN.

Candidate generation uses FAISS approximate nearest-neighbor search over dense embeddings. The paper does not report the FAISS index type or distance metric, although cosine similarity is stated as the standard scoring function:

$$
s(u, v) = \frac{u \cdot v}{\lVert u \rVert \lVert v \rVert}
$$

For each Japanese article, the system retrieves up to 120 nearest-neighbor candidates, retains up to 80 candidates per country, and then applies balanced country-level quotas of 30 for Korea and 30 for France. Reranking is performed by a transformer-based Cross-Encoder that jointly encodes provision pairs in the format `"[CLS] provision_A [SEP] provision_B [SEP]"` and outputs a scalar relevance score $z$. Pairs are reranked by that score; a softmax normalization is described but not mandated:

$$
\operatorname{softmax}(z_i) = \frac{\exp(z_i)}{\sum_j \exp(z_j)}
$$

The final output contains up to 60 correspondences per Japanese article [2603.15094].

| Component | Setting | Value |
|---|---|---|
| Embedding model | Multilingual encoder | `multilingual-e5-large` |
| Candidate retrieval | `topk_cand` | 120 |
| Pre-filter per country | `pre_per_country` | 80 |
| Country quotas | Korea, France | 30, 30 |
| Final output | `topk_final` | 60 |

The architecture therefore separates high-recall dense retrieval from higher-precision pairwise reranking. This suggests a standard two-stage retrieval stack adapted to multilingual legal comparison, but the paper is explicit that the prototype relies on retrieval and reranking rather than on a supervised, legally calibrated equivalence classifier.

## 4. Correspondence graphs and exploratory comparative analysis

Candidate correspondences are incorporated into a network in which provisions are nodes and retained cross-jurisdictional links are edges. In the reported visualization, the graph contains 2,000 nodes in total: 776 Korean nodes, 232 Japanese nodes, and 992 French nodes. The number of edges is 3,545 [2603.15094].

Network inclusion is controlled by country-specific thresholds and per-source quotas. For Japan–Korea, the system retains at most 3 edges per Japanese article with similarity at least 0.95. For Japan–France, it retains at most 3 edges per Japanese article with similarity at least 0.80. These thresholds are explicitly described as balancing relation density across jurisdictions. If an adjacency matrix $A$ is constructed between Japanese provision $i$ and foreign provision $j$, the paper gives the thresholded form

$$
A_{ij} =
\begin{cases}
s_{ij}, & \text{if } s_{ij} \ge \tau_{\text{country}} \\
0, & \text{otherwise}
\end{cases}
$$

with the additional quota condition applied in practice.

Visualization uses a bipartite column layout with Japanese provisions in the center, Korean provisions on the left, and French provisions on the right. Edges depict candidate correspondences. Users can inspect concentrations of alignments by book or chapter and drill down into articles. Rendering is described in relation to AKN viewers, while a more interactive interface is reserved for future work.

The graph is therefore not merely a display artifact. It is the model’s exploratory surface: structural normalization makes provisions addressable and comparable, while semantic scoring turns inter-jurisdictional proximity into weighted edges. A plausible implication is that the network is intended less as a final adjudication of doctrinal similarity than as an instrument for expert-guided navigation through large legislative corpora.

## 5. Case studies, evaluation, and observed failure modes

The target corpora are AKN-encoded Civil Codes and Commercial Codes for Japan, Korea, and France, with Germany also prepared. The reported experiment focuses on the Civil Codes of Japan, Korea, and France. The paper states that high-score correspondences cluster around doctrinally similar sections, but it does not publish specific article pairs [2603.15094].

Two earlier validation settings are reported. A domestic alignment experiment compared the Japanese Electricity Business Act with the Gas Business Act using a manually prepared gold standard and obtained $F1 = 0.768$. A cross-jurisdiction alignment experiment compared family-law sections of the Japanese Civil Code and the German Civil Code using pairs derived from a commentary’s comparison sections and obtained $F1 = 0.348$. These results are presented as feasibility evidence, while also underscoring the higher difficulty of cross-lingual, cross-system legal alignment.

The evaluation status of the JP–KR–FR network remains exploratory. Large-scale expert-annotated gold standards and comprehensive metrics such as precision, recall, MRR, and nDCG are planned but not yet reported. This is important for interpreting the system’s status. The prototype demonstrates operational viability, but it does not yet claim benchmark-grade validation across the full multilingual comparative-law task.

The paper also identifies several qualitative failure modes. Linguistic differences and translation variation reduce precision across jurisdictions. Conceptual divergence between legal traditions leads to semantically close but functionally distinct matches. Structural disparities in placement and granularity can mislead similarity at article boundaries. Context-sensitive definitions and renvoi or citations complicate pure text-based matching. These limitations directly constrain the interpretation of similarity scores: high semantic proximity does not guarantee doctrinal equivalence, and low proximity does not necessarily imply absence of functional analogy.

## 6. Limitations, future work, and place in computational comparative law

The framework’s stated limitations fall into five broad categories. First, semantic nuance and context sensitivity remain difficult because statutory language depends on definitions, exceptions, and renvoi, while sentence-level embeddings may miss dependencies across paragraphs or items. Second, multilingual alignment is complicated by domain-specific terminology and doctrinal framing across Japanese, Korean, French, and German corpora. Third, structural disparities mean that functional analogs may reside in different titles or chapters and at different granularities. Fourth, citations and cross-references are preserved but not modeled semantically in the matching score. Fifth, although FRBR Expression captures version date, longitudinal drift in legal meaning is not yet modeled, so correspondence validity may be time-bounded [2603.15094].

The future work agenda follows directly from those constraints. The paper proposes integrating legal ontologies and knowledge graphs to represent defined terms, renvoi, and cross-references as structured features. It proposes domain-specific multilingual legal encoders, described conceptually as “World-Legal BERT,” trained or fine-tuned on large AKN corpora. It proposes supervised alignment using expert-annotated correspondences to train Cross-Encoders and calibrate thresholds. It also proposes broader language coverage, richer mapping rules for citations and defined terms, standardized URI schemes, temporal/version-aware comparison, and interactive search and explanation modules with LLM integration.

In that sense, the Dual-Laws Model is best understood as an infrastructural design for comparative-law computation rather than as a closed benchmark system. It operationalizes computational comparative law by coupling JLS→AKN structural interoperability with multilingual semantic correspondence detection, and it does so in a way that makes the outputs inspectable as cross-jurisdictional networks. The framework’s significance lies in this coupling: without structural normalization, provisions are difficult to identify and link consistently; without semantic retrieval and reranking, interoperable documents remain only formally aligned. The model joins those two requirements into a single workflow and thereby defines a concrete architecture for scalable, exploratory comparative analysis.

Source: https://www.emergentmind.com/topics/dual-laws-model-6322014f-cf9a-4aa0-a47b-d76667e7e419