LegalSemi: Structured Semantics in Legal AI

• LegalSemi is a framework and benchmark that employs explicit semantic annotation and hybrid lexical-semantic retrieval to enhance legal text modeling.
• It integrates annotated corpora and structured knowledge graphs to improve IRAC automation and legal reasoning in contract law.
• The approach achieves measurable gains in issue spotting (+21.4%), rule retrieval (+17.2%), and application quality, setting new standards for legal AI.

LegalSemi encompasses a mosaic of approaches in legal artificial intelligence that target the structured appreciation and modeling of legal semantics, often via hybrid lexical-semantic retrieval, annotated corpora, knowledge-graph–augmented reasoning, and semi-formal or semi-structured representations. Recent usage emphasizes LegalSemi as a benchmark dataset for IRAC automation in Malaysian contract law, but the term also appears throughout the literature in reference to semantically-rich representations for retrieval, reasoning, and compliance support in legal NLP.

1. Foundations of LegalSemi: Semantic Structuring in Legal AI

LegalSemi refers to frameworks, datasets, and evaluation methodologies that embody a semantically-grounded approach to legal text representation and retrieval. The concept consolidates:

• Explicit semantic annotation (norm/role/entity labeling, rhetorical segmentation)
• Hybrid retrieval pipelines mixing lexical and dense (neural) methods
• Integration of structured or semi-structured legal knowledge bases
• Benchmarks capturing legal reasoning processes (e.g., IRAC, deontic logic, fact-rule chains)
• Evaluation metrics that go beyond surface-level relevance to capture reasoning-fidelity and retrieval adequacy

Crucially, LegalSemi solutions treat legal text not as unstructured bags of words but as repositories of explicit, tractable meaning, policy structure, and context-sensitive logic. This stands in contrast to shallow keyword retrieval or "black-box" end-to-end modeling devoid of legal reasoning transparency (Kang et al., 2024, Mori et al., 15 Jun 2025, Le et al., 19 Jul 2025).

2. LegalSemi as an IRAC Benchmark: Malaysian Contract Law

The most concrete instantiation of LegalSemi in recent literature is as a benchmark for automating IRAC analysis (Issue, Rule, Application, Conclusion) within Malaysian contract law (Kang et al., 2024). Key properties include:

• 54 expertly annotated legal scenarios, covering contract formation, consideration, promissory estoppel, intention, and capacity
• Each scenario decomposed into IRAC subcomponents, linked granularly to statutes (e.g., Malaysian Contracts Act 1950) and judicial precedents
• Cross-checked by qualified annotators, reaching an IRAC-agreement ratio > 0.8
• Annotation schema mediated by a semi-structured knowledge graph (SKG) encoding statutory sections, legal concepts, and layperson interpretations in Neo4j
• Automated pipeline stages for concept identification, issue spotting, rule retrieval (via symbolic and dense methods), conditional application synthesis, and conclusion generation

The empirical pipeline demonstrates that explicit SKG-driven prompting enhances issue identification (+21.4%), rule retrieval F1@5 (+17.2%), and application quality (+18.9%) for LLMs. Both zero-shot and few-shot models benefit, but fine-grained legal term acquisition remains a challenge. The LegalSemi benchmark thus provides the first end-to-end testbed for neuro-symbolic (hybrid logical- and LLM-based) approaches to legal scenario resolution.

3. LegalSemi in Lexical-Semantic Hybrid Retrieval

LegalSemi also denotes two-stage or hybrid retrieval systems that exploit both surface lexical cues (BM25, TF-IDF) and contextual semantic embeddings (e.g., SBERT, LegalBERT) (Mori et al., 15 Jun 2025, Le et al., 19 Jul 2025):

• Stage 1: Lexical models (BM25, BM25Plus) filter top candidates based on term-frequency/inverse-document-frequency, excelling in formulaic, repetitive legal language and long queries
• Stage 2: Dense encoders (fine-tuned BERT-based bi-encoders, LegalSBERT, PhoRanker) re-rank these candidates by cosine similarity in embedding space—critical for handling paraphrase, synonymy, and semantic concept drift
• Semi-hard negative mining is central: negatives are selected from high-similarity candidates (not the hardest), maximizing learning signal while avoiding label noise and unstable loss
• Introduction of the Exist@m metric, measuring whether any gold document enters the candidate pool for downstream re-ranking ("Exist@90" reaches 0.976 post-fine-tuning)
• Lightweight systems (dual 110M parameter encoders) can achieve MRR@10 ≈ 0.79—a ~41% improvement over simplistic setups and competitive with parameter-intensive ensembles (Le et al., 19 Jul 2025)

This duality aligns retrieval with legal practitioner workflows, ensuring both transparency (BM25-driven traceability) and resilience to legal paraphrase (semantic embeddings), and is particularly robust on large, jurisdictionally fragmented corpora.

4. Semantic Annotation and Norm Recognition

A critical element of LegalSemi is fine-grained legal semantic labeling, notably in systems for semantic norm recognition (Duarte et al., 2022, Sleimi et al., 2020). Key facets:

• Span classification over legal texts, annotating norm types (obligation, right, definition, etc.), named entities, and semantic roles (experiencer, condition, purpose, effect)
• Transformer (BERTimbau) + BiLSTM + biaffine span scorer architecture for multi-layer annotation, with partial norm dependency flows to inform deep legal role prediction
• Benchmark results on Portuguese consumer law: macro-averaged F₁ = 81.44%, with per-class peaks for LEFFECT (100% F₁) and OBLIG (88.37% F₁)
• Downstream application: norm-labeled segments enable automatic QA pair generation, driving a boost to top-1 retrieval accuracy from 44% to 51% (QA-fine-tuned retrieval)
• Annotation noise and span-boundary ambiguity (10–15% F₁ loss in fine-grained roles) persists, partly due to the intricate, layered structure of legal text

This suggests that semantic norm annotation, whether at the statement or phrase level, is integral to both retrieval relevance and compliance analysis, especially across jurisdictional boundaries and document types.

5. LegalSemi in Benchmarking, Evaluation, and Best Practices

Large-scale evaluation platforms such as the Massive Legal Embedding Benchmark (MLEB) provide an ecosystem for quantifying the efficacy of LegalSemi-inspired models (Butler et al., 22 Oct 2025):

• Ten expert-annotated datasets cover retrieval, classification, QA across US, UK, EU, Australia, Ireland, Singapore, and various document types
• Metrics include cosine/dot-product similarity, InfoNCE/triplet loss for training, and ranked evaluation via NDCG@k, MRR, MAP
• Domain-adapted dense models (e.g., Kanon 2 Embedder) achieve NDCG@10 up to 86.03; transferability is reduced without multi-jurisdictional pretraining (e.g., 8-point NDCG drop when US-specialized models are used for UK law)
• Best practices involve chunking statutes to 512–1024 tokens, hybrid retrieval (embedding+BM25 yielding +5pp NDCG@10 on Bar Exam QA), clause normalization, and temperature-tuned re-ranking

LegalSemi systems must therefore be rigorously evaluated not only for retrieval metrics, but for their ability to accurately surface precise logical and policy-relevant information in diverse, high-stakes contexts.

6. Semi-Structured and Semi-Formal Representations

LegalSemi methodologies frequently employ semi-structured or semi-formal representations to bridge legal expertise and knowledge engineering. SBVR Structured English (SE), as formalized in KR4IPLaw, illustrates this (Ramakrishna et al., 2014):

• SE grammar constrains legal statements to NounConcepts, VerbConcepts, and deontic RuleTemplates (e.g., “It is obligatory that Examiner rejects Claim if Claim is_rejected_under EssentialSubjectMatterRequirement”)
• Each construct is mapped to an OWL class/property or LegalRuleML norm, enabling transformation to ontological or rule-based systems
• Annotated SE forms are human-authored by legal experts, parsed, and transformed to LegalRuleML (XML) and OWL2 ontologies
• Patent-law examples verify the approach, but expressivity is limited for procedural/temporal and argumentation-heavy rules

A plausible implication is that semi-formal representations enable bi-directional transfer between natural-language legal knowledge and formal compliance systems, but generalization remains constrained by vocabulary curation and complex cross-referencing needs.

7. Outlook and Open Directions

LegalSemi, in its various manifestations, establishes a robust paradigm for legal text understanding, retrieval, and reasoning that foregrounds semantically explicit, linguistically and logically structured representations. Its impact is evidenced by:

• Tangible gains in retrieval effectiveness and knowledge transfer when hybrid pipelines and explicit semantic annotations are deployed
• Domain adaptation and negative mining strategies that support lightweight but high-performing systems
• Benchmarks (LegalSemi IRAC, MLEB) setting new standards for reproducibility, comparative evaluation, and task-difficulty stratification

Continued research is warranted in extending annotation schemas, dynamically updating structured knowledge bases, integrating symbolic-neural inference for advanced reasoning, and generalizing LegalSemi constructs across legal domains and languages. The convergence of semi-structured knowledge, hybrid retrieval, and logic-aware generation defines the cutting edge for legal informatics.