Hyperlexicon

3 Knowledge Sources

820K+ Total Nodes

1.3M+ Total Edges

10 Languages

0.8752 STSb Spearman ρ

< 80 ms Query p95

MULTI-SOURCE KNOWLEDGE GRAPH

WordNet + ConceptNet + Wiktionary

WORDNET 3.1

Synsets

117,659

Lemmas

155,000+

Surface forms

206,978

Domains

438

Taxonomy, hypernymy, hyponymy, antonymy, domain/register annotations.

CONCEPTNET 5.7

Commonsense edges

500K+

Relation types

relatedTo, usedFor, capableOf, causes, hasProperty, atLocation, isA, partOf, derivedFrom, definedAs, similarTo, synonym, antonym.

WIKTIONARY

Additional senses

200K+

Coverage

Modern vocabulary

Fills WordNet gaps (last updated 2006) with slang, technical terms, and current usage. Sourced from kaikki.org parsed dumps.

SENSE-AUGMENTED EMBEDDINGS

Primary encoder

Mistral-7B (4096-dim)

Fallback encoder

MiniLM (384-dim, 22M params)

Reranker

ms-marco-MiniLM-L-6-v2

ANN index

FAISS HNSW

Structured semantic context (definitions, synonyms, commonsense relations) is injected into text before encoding. The encoder sees knowledge, not just tokens.

STSBENCHMARK RESULTS

Model	Spearman ρ	Time (1,379 pairs)
HyperLexicon + BGE-large (sense-augmented)	0.8752	5.5 s
BGE-large-en-v1.5 (raw, no augmentation)	0.8689	4.9 s
Delta from sense augmentation	+0.0063	+12% latency

Top-3 competitive in parameter class on STSBenchmark. Zero fine-tuning, zero additional training data. Improvement comes purely from structured knowledge injection.

EXECUTIVE BRIEF

Model Performance Envelope & Architecture

MODEL PERFORMANCE ENVELOPE

ARCHITECTURE EFFICIENCY

95–99% of queries resolve through graph and vector pipelines without LLM inference. Compute scales linearly with index size, not token context.

NEURAL OBSERVABILITY • Task-Level Timing Targets

Task	Avg Time	Notes

Operational Comparison

Metric	Hyperlexicon	LLMs (e.g., GPT-4, Claude)

Figures reflect steady-state optimized configurations.

Business Value & Cost Model • Functional Value

QUADRINARY ROUTING ARCHITECTURE

Stage	Handler	Purpose
1. Preprocessing	eLL (Language Layer)	Text normalization, tokenization, syntactic parsing
2. Semantic enrichment	Hyperlexicon	Sense disambiguation, knowledge injection, graph traversal
3. Conditional LLM	Base model (if needed)	Dictionary/thesaurus requests only (1–5% of traffic)
4. Default	Return enriched text	Pass cleaned/enriched output to downstream cognitive pipeline

LLM is a tool, not the engine. Language model inference is reserved for dictionary/thesaurus lookups; all other queries resolve through the graph and vector pipelines.

AI-SPECIFIC FAILURE PATTERNS

Track early and tie to gates: accuracy, tail-latency, cache miss, index staleness.

NEURAL OBSERVABILITY • Quality & Gates

5-minute SLO windows with real-time OOM and degrade triggers.

ML SECURITY MODEL • CONTINUOUS LEARNING OPS

Execution: local-only inference; signed weights/indices; PII-scrubbed vectors; network-gapped nodes.

NODE REGISTRY • BRCA1 PATHWAY

Legend: — = Not applicable (by design), ∅ = Unknown (expected but not found), □ = Withheld (exists but redacted). All timestamps ISO-8601 UTC. Provenance required for all nodes.

OID	LABEL	TYPE	ENSEMBL	PROVENANCE	CREATED	ONTOLOGY	HGVS	UBERON

NODE SCHEMA SPECIFICATION

CORE IDENTIFIERS

oid STRING PRIMARY KEY
label STRING DISPLAY
type ENUM NODE_TYPE
created TIMESTAMP UTC

SEMANTIC REFERENCES

ensemble STRING? ENSEMBL_ID
ontology STRING? GO_TERM
hgvs STRING? HGVS_NOTATION
uberon_like STRING? ANATOMY_REF

PROVENANCE & METADATA

provenance STRING? SOURCE_INFO

TYPE DISTRIBUTION & PATTERNS

BIOLOGICAL ENTITIES

gene ENSEMBLE + GO TERMS
pathway GO ONTOLOGY REF
variant HGVS NOTATION
tissue UBERON REFERENCE

ANALYTICAL NODES

effect FUNCTIONAL IMPACT
classification CLINICAL SIGNIFICANCE
source PROVENANCE TRACKING
output INFERENCE RESULT

GRAPH TOPOLOGY METRICS

CONNECTIVITY PATTERNS

Gene → Pathway 1:N MEMBERSHIP
Variant → Gene N:1 BELONGS_TO
Effect → Variant 1:1 CAUSES
Classification → Effect 1:1 DESCRIBES

INFERENCE CHAINS

Gene → Tissue EXPRESSION_IN
Source → * EVIDENCE_FOR
* → Output GENERATES
Query → Output SEMANTIC_MATCH