Named Entity Linking (NEL)

The initial stage identifies spans of text that refer to entities. In legal documents, this must handle complex noun phrases like 'the Board of Directors of the Delaware Corporation'.

• Surface Form Extraction: Isolates the exact text string used to refer to the entity.
• Boundary Detection: Determines the precise start and end tokens of the mention, critical for multi-word legal terms.
• Nested Entity Handling: Resolves overlapping mentions, such as 'the Court of Appeals for the Second Circuit' containing 'Second Circuit'.
• Domain Adaptation: Requires models fine-tuned on legal corpora to avoid missing citations like 'Id.' or 'supra'.

This phase retrieves a set of possible knowledge base identifiers for each detected mention. Efficiency here is paramount for large legal ontologies.

• Alias Tables: Uses pre-built dictionaries mapping surface forms (e.g., 'Supreme Court') to candidate IDs.
• Fuzzy Matching: Employs edit distance (Levenshtein) and phonetic algorithms to catch OCR errors or misspellings in scanned briefs.
• Acronym Expansion: Resolves context-specific acronyms like 'DOJ' by expanding them against the knowledge base.
• Pruning: Reduces the candidate set using simple heuristics (e.g., entity type filtering) to minimize the load on the disambiguation step.

The core reasoning step selects the single correct entity from the candidate set by analyzing the surrounding textual context.

• Local Context Windows: Analyzes words immediately adjacent to the mention for disambiguating signals.
• Global Coherence: Ensures all linked entities in a document form a semantically consistent set (e.g., linking 'Apple' to the company if 'iPhone' is also mentioned).
• Graph-Based Ranking: Applies algorithms like PageRank over a graph of candidate entities to identify the most central and coherent cluster.
• Neural Scoring: Uses transformer models to encode the mention in context and the candidate entity description into a shared vector space for similarity scoring.

A critical capability for legal knowledge graphs is recognizing when a valid mention has no corresponding entity in the target knowledge base.

• Out-of-KB Detection: Classifies mentions that refer to entities not yet curated in the ontology.
• Thresholding: Applies a confidence score cutoff; if no candidate exceeds the threshold, the system predicts NIL.
• Novel Entity Flagging: Marks NIL predictions for downstream ingestion, triggering a workflow to add the missing entity to the knowledge graph.
• Prevents Forced Errors: Avoids the catastrophic error of linking a new case citation to a random, unrelated existing case.

Assigning a fine-grained semantic type to the linked entity provides essential context for downstream legal reasoning tasks.

• Type Hierarchies: Maps entities to nodes in a legal ontology (e.g., LegalActor > Judge, LegalDocument > Contract).
• Fine-Grained Typing: Goes beyond coarse types (PERSON, ORG) to domain-specific types like Plaintiff, Statute, or Jurisdiction.
• Multi-Label Assignment: Recognizes that a single entity can have multiple valid types (e.g., a company can be both a Corporation and a Litigant).
• Type Constraints: Uses the assigned type to filter candidate entities during disambiguation, improving precision.