Epicure: Structured Pattern Extraction Systems

• Epicure is a set of rigorously designed systems that extract structured patterns from high-dimensional data across machine learning, NER, and culinary sciences.
• It employs innovative methodologies including pattern lattices for sequence models, ensemble NER with transformers, and graph-based ingredient embeddings to improve interpretability and precision.
• Epicure demonstrates practical advancements with enhanced accuracy in function naming, anomaly detection, mutation extraction, and culinary ingredient analysis through empirical evaluation.

Epicure refers to a set of distinct, technically rigorous systems and methodologies across machine learning, information extraction, and computational gastronomy. In all usages, "Epicure" denotes an approach that distills high-dimensional, heterogeneous data into human-interpretable structure, whether via pattern lattices in sequence model predictions, neural ensemble architectures for domain NER, or embedding geometry and ingredient curation in recipe-driven food science. The following sections provide a detailed account of these systems, each developed independently but sharing the unifying theme of structured pattern extraction and representation.

1. Pattern Lattices in Sequence Model Output (Epicure (Allamanis et al., 2023))

Epicure is a method for distilling the output of autoregressive sequence models (such as beam search predictions) into concise, abstract regular-expression-like patterns that capture both precise and generalizable aspects of model predictions. Denoting the vocabulary of subtokens as $V$ and the set of predicted concrete sequences as $T = \{t\}$, each model prediction $t = [t_1, ..., t_n]$ is assigned a probability mass $P(t)$. Patterns $s$ are sequences on $V \cup \{\star\}$, where $\star$ ("wildcard") stands for zero or more arbitrary subtokens, and a matching predicate $m(s, t)$ determines if $s$ matches $t$.

Epicure constructs a subsumption lattice over all patterns, with a partial order defined by set inclusion of their match sets:

$T = \{t\}$0

Joins in this lattice (anti-unification) replace mismatched subsequences with a wildcard. Probability mass is aggregated:

$T = \{t\}$1

For applications such as function naming or anomaly detection, Epicure selects a minimal set $T = \{t\}$2 of least-general patterns above a probability threshold $T = \{t\}$3, ensuring mutual compatibility and bounding the false-alarm rate to $T = \{t\}$4.

On naming and anomaly detection benchmarks at 10% FAR, Epicure achieves RegexAcc $T = \{t\}$5 33% versus a 26% baseline (61% relative gain); for anomaly detection, Complete Matches reach 42% versus 26% for the baseline. These gains arise from Epicure’s abstraction over rare sequences—where full matches have low $T = \{t\}$6 but patterns holding for reliable fragments offer higher $T = \{t\}$7—enabling tighter, high-precision predictions in high-entropy settings (Allamanis et al., 2023).

2. Ensemble NER with Pretrained Transformers (EPICURE for Cancer Mutation Extraction (Cao et al., 2021))

EPICURE in the context of biomedical information extraction refers to an ensemble named entity recognition (NER) system designed for extracting cancer-related mutations from literature. The architecture comprises:

• Pretrained Encoder: Utilizes BERT-Large, BioBERT, or SciBERT to transform tokenized sentence input $T = \{t\}$8 into contextual embeddings $T = \{t\}$9 after prepending special tokens.
• CRF Pattern Layer: Casts mutation extraction as a sequence labeling task (BIO or BMEOW) with a conditional random field (CRF) decoding the best label sequence. The emission- and transition-based CRF score is maximized via negative log-likelihood.
• Span Prediction Layer: A pointer network predicts mutation spans by independently modeling start and end positions for each mutation type using token-wise softmax projections.
• Ensemble Voting: Aggregates three token-wise predictions (CRF under BIO and BMEOW, span), mapping all to BIO and applying majority voting.

To address data scarcity, EPICURE augments its training set by merging three annotated corpora (TmVar, EMU, BRONCO) and refining negative sampling using a constructed mutation dictionary. The final model, fine-tuned with Adam ($t = [t_1, ..., t_n]$0, 24 batch size, 100 epochs), achieves strong results across all datasets (e.g., BRONCO $t = [t_1, ..., t_n]$1), with ablation studies attributing gains to both ensemble and data augmentation strategies. EPICURE demonstrates generalization across varying annotation standards and mutation types while avoiding hand-engineered rule dependency (Cao et al., 2021).

3. Food Ingredient Embedding Geometry and Curation (Epicure Systems for Culinary Knowledge, (Radzikowski et al., 21 May 2026, Radzikowski et al., 2 Apr 2026))

Several Epicure systems target the analysis of food ingredient spaces via graph-based neural embeddings, informed by both recipe co-occurrence and food chemistry.

3.1 Data Curation and Canonicalization

Epicure normalizes millions of raw recipe ingredient strings into a closed canonical vocabulary (1,790 or 1,032 entries), using a multi-stage pipeline: deterministic LLM classification, semantic clustering (using gemini-embedding-001), and manual curation steps. Each canonical ingredient is mapped to FlavorDB and, where possible, USDA FoodData Central.

3.2 Graph and Embedding Construction

• Graphs: The system constructs a weighted undirected ingredient-ingredient (I-I) graph using Normalized PMI over recipes, and a typed heterogeneous ingredient-compound (I-C) graph using FlavorDB (with 2,247 compound nodes across 15 categories).
• Metapath2Vec Variants: Three sibling models share skip-gram/negative-sampling architecture ($t = [t_1, ..., t_n]$2), differing in random-walk schemas:
  • Cooc: Walks only the I-I co-occurrence graph.
  • Chem: Walks only metapaths through typed compound nodes.
  • Core: Blends both with controlled walk-injection, providing explicit spectrum control (chemistry vs. co-occurrence).

3.3 Emergent Structure and Dimensionality

These embeddings encode rich culinary semantics, revealed via isotropy measures (participation ratio, mean cosine), label recovery (NMI and $t = [t_1, ..., t_n]$3NN purity for food groups/cuisines), and supervised probes against chemical, sensory, macro-nutrient, and cultural axes. Embedding geometry is further analyzed using FastICA and GMM to reveal “modes” corresponding to interpretable culinary clusters (e.g., “South-Asian spice blends,” “sweet baking”). Notably, Core achieves maximal mode coherence (0.833) and competitive label-recovery across food groups and cuisines (Radzikowski et al., 21 May 2026).

4. Explicit Culinary Dimensions and Applications (Radzikowski et al., 2 Apr 2026)

The Epicure framework systematically extracts explicit, independently classifiable axes from food ingredient embedding spaces, using LLM-augmented curation to maximize structure recovery:

• Dimensions Identified: At least fifteen axes—including taste (sweet, salty, sour, bitter, umami, Scoville heat), texture (hardness, viscosity, crunchiness, chewiness, moisture, fattiness), geography (climate latitude), processing (NOVA score), and culture (7-way regional cuisine clusters)—each verified against hand- or lab-coded benchmarks.
• Empirical Validation: Axes correlate with laboratory measurements (USDA FoodData Central, FooDB); e.g., sweetness vs. sugars ($t = [t_1, ..., t_n]$4), saltiness vs. sodium ($t = [t_1, ..., t_n]$5), umami vs. glutamic acid ($t = [t_1, ..., t_n]$6). Cross-validated projections confirm robust separability and predictive utility.
• Applications: The system enables multi-dimensional ingredient substitution, dietary accommodation, cuisine profiling, texture calibration, fusion recipe development, and “clean-label” reformulation by matching ingredients along interpretable axes validated by chemical/nutritional data.

5. Comparative Synthesis and Theoretical Insights

Epicure approaches differ in domain but share methodological convergence:

• Pattern Lattices: The sequence model Epicure and the pattern-ensemble NER system both formalize complex, high-entropy outputs via pattern extraction, lattice structures, or ensemble hypothesis selection, with explicit treatment of precision–recall and generality–specificity trade-offs.
• Embedding Geometry: The embedding-based Epicure systems treat model navigation as movement in a high-dimensional, semantically structured space. Explicit axes and clustering reveal that culinary, sensory, and cultural regularities are reliably embedded—often recoverable by curation-driven axis construction and dimensionality reduction.
• Data Quality: Across all Epicure systems, a bottleneck is identified in curatorial rigor, not solely modeling architecture. LLM-augmented pipelines substantially improve recoverable structure, application fidelity, and utility for downstream tasks.

6. Limitations and Future Directions

• The pattern extraction Epicure’s effectiveness is mediated by the entropy and structure of sequence model outputs; rare, high-quality sequences benefit most from its abstraction strategy (Allamanis et al., 2023).
• NER performance in EPICURE is constrained by label imbalance, especially for rare mutation types; future upgrades could involve class-balanced losses and more principled negative sampling (Cao et al., 2021).
• In culinary embedding Epicure, corpus imbalance and LLM-dependence in curation and labeling limit generalization; enhancing data diversity, introducing continuous mixing in random-walk schemas, and cross-modal alignment are plausible extensions (Radzikowski et al., 21 May 2026).
• Across domains, statistical significance tests and multi-task or curriculum learning remain underexplored avenues for improvement.

Epicure, in its various specialized realizations, exemplifies a principled, data-driven approach to extracting latent, interpretable structure from complex prediction, information extraction, and culinary representation domains, grounded in rigorous methodology and empirical validation.