Dictionary Graphs: Models & Applications

• Dictionary graphs are graph-based models that represent words and definitions as nodes connected by directed edges, capturing complex lexical relationships.
• They employ techniques like sink-pruning, SCC extraction, and MinSet computation to reveal essential structures such as the kernel, core, and satellites within dictionaries.
• Applications range from semantic search and reverse dictionary retrieval to computational linguistics and cognitive insights into language acquisition.

A dictionary graph is a mathematical and algorithmic framework that models dictionaries as graphs—typically directed, weighted, multi-partite, or semantic graphs—capturing the structural and functional relationships among words, senses, or features. Dictionary graphs play a central role in the study of the latent structure and algorithms underlying lexical resources, representation learning, graph signal processing, and linguistic grounding. Their analysis reveals fundamental properties about language, cognition, and structured data.

1. Dictionary Graphs: Definition and Basic Models

The canonical model of a dictionary as a graph is a directed graph (digraph) where each node is a word (often restricted to content words), and there is a directed edge from word $u$ to word $v$ if $u$ appears in the definition of $v$ (formally: $G=(V, E)$, $V= \{w_1,\ldots, w_n\}$, $E = \{(u \to v ) : u \text{ in def}(v)\}$). This simple construction, when recursively analyzed, reveals nontrivial topological properties and allows one to probe the minimal structures necessary for the definitional system to function without circularity or redundancy (Vincent-Lamarre et al., 2014, Picard et al., 2013).

Extensions of dictionary graphs appear in several contexts:

• Sense graphs: Nodes correspond to word senses, and edges encode definitional, translational, or semantic relations (including disambiguation in multilingual lexica) (Flati et al., 2014).
• Graph signal models: Nodes represent variables (e.g., sensors, regions, attributes) and edges encode structural relations inferred or imposed on the data; dictionary graphs here represent basis elements or "atoms" for graph signal decomposition (Chen et al., 2016, Cappelletti et al., 2024).
• Semantic knowledge graphs: Nodes include lexical units, semantic roles, or gloss chunks, and edges encode fine-grained semantic or ontological relations (Silva et al., 2018).

2. Latent Structure: Kernel, Core, Satellites, and Feedback Sets

Directed dictionary graphs exhibit a distinctive stratified structure emerging from the patterns of definitional dependencies:

• Kernel ($K$): The unique minimal subgraph obtained by recursively removing "sink" nodes (words that never define others). $K$ is typically $\approx$8–12% the size of the entire dictionary. All words outside $K$ can be defined from it by chains of definitions (Vincent-Lamarre et al., 2014, Picard et al., 2013).
• Core ($C$): The largest strongly connected component (SCC) within $K$, typically 65–90% of $K$ (i.e., $\approx$6.5–9% of the dictionary). Every pair of words in $C$ is mutually reachable via definitional paths.
• Satellites ($S$): The remaining much smaller SCCs in $K$ (10–35% of $K$) surround the core and serve as additional definitional resources.
• Minimum Feedback Vertex Sets (MinSets)/Minimal Grounding Sets (MGS): Minimal sets of nodes whose removal breaks all directed cycles, making the graph acyclic. MinSets are the smallest sets of grounded words required to define all others and are $\approx$1–4% of the dictionary (about 15–48% of the kernel), with elements drawing from both core and satellites (Vincent-Lamarre et al., 2014, Picard et al., 2013).

These structures are algorithmically extracted via sink-pruning and SCC decomposition (using linear-time algorithms such as Tarjan's or Kosaraju's), while MinSet extraction reduces to the NP-hard minimum feedback vertex set problem, commonly solved via ILP or combinatorial algorithms.

3. Algorithmic Frameworks and Models

Several algorithmic paradigms for dictionary graphs have been developed:

Dictionary Structure Mining

• Sink-pruning (to compute the kernel): recurses by removing nodes with out-degree zero; $O(|V|+|E|)$.
• SCC extraction: identifies strongly connected structure; also $O(|V|+|E|)$.
• MinSet/MGS computation: formulated as feedback vertex set optimization; solved with ILP or reductions (Picard et al., 2013).

Signal Processing and Dictionary Learning on Graphs

• Graph Dictionary Signal Model: Models data as mixtures of multivariate signals defined by linear or convex combinations of graph Laplacians (“atoms”); algorithms infer both the graph atoms and their mixture coefficients via convex optimization, notably via generalized primal-dual splitting (Cappelletti et al., 2024).
• Piecewise-constant dictionary learning: Decomposes signals on a known graph into a sparse linear combination of adaptive, piecewise-constant patterns, each supported on a “dictionary graph” (subgraph), using alternating minimization and localization algorithms (Chen et al., 2016).
• Online Graph Dictionary Learning: Models unregistered (possibly heterogeneously-sized) graphs as sparse mixtures of learned graph atoms using Gromov-Wasserstein divergence, enabling subspace tracking, clustering, and kernelization (Vincent-Cuaz et al., 2021).

Lexical and Semantic Enhancement

• CQC Algorithm: Uses cycles and quasi-cycles in the bilingual sense graph to disambiguate translation edges, semantically enrich dictionaries, and extract synonyms (Flati et al., 2014).
• Reverse Dictionary via Node-Graph Architecture: Constructs a sparse, directed graph based on word-definition links, and ranks retrieval candidates using shortest-path distances modulated by inverse term frequency (Thorat et al., 2016).
• Knowledge graph construction from definitions: Segments definitions into semantic roles, transforms them into RDF triples, and enables rich SPARQL querying and interpretable entailment reasoning (Silva et al., 2018).
• Local Graph-based Dictionary Expansion (LGDE): Builds a word similarity manifold using continuous k-NN on embedding spaces, detects local semantic neighborhoods via graph diffusion, and reconstructs expanded dictionaries with principled community structure (Schindler et al., 2024).

4. Theoretical and Psycholinguistic Insights

Quantitative analyses across several English dictionaries have established the following empirical gradients (Vincent-Lamarre et al., 2014, Picard et al., 2013):

Structure	% Dictionary	Word properties
Kernel (K)	8–12%	Essential definitional “backbone”
Core (C)	6.5–9%	Earliest learned, most frequent, less concrete
Satellites (S)	1–3%	Intermediate age/frequency, more concrete
MinSet/MGS	1–4%	Composed of core and satellite words
Rest	88–94%	Least frequent, learned last, moderately concrete

Core words are typically associated with higher usage frequency, earlier age of acquisition, and lower concreteness. Satellite words exhibit a psycholinguistic gradient: the further from the core, the more concrete, less frequent, and later-learned. Words outside the kernel are learned last.

A principal implication is that the structure of dictionary graphs reflects deep cognitive and semantic regularities, suggesting a strong alignment between definitional topology and psycholinguistic characteristics.

5. Applications in NLP, Knowledge Representation, and Learning

Dictionary graph frameworks enable several practical applications:

• Grounding lexical meaning: Only a small MinSet/MGS needs to be directly grounded (sensorimotor experience or perceptual learning), allowing the rest of the lexicon to be defined symbolically via acyclic traversal (Vincent-Lamarre et al., 2014).
• Controlled vocabulary expansion: LGDE leverages graph manifold structure to discover new, relevant keywords for specialized corpora (e.g., hate speech, conspiracy terminology), outperforming naive cosine- or co-occurrence-based expansions (Schindler et al., 2024).
• Semantic enrichment and curation: Cycle-search algorithms (e.g., CQC) identify translation sense alignments, synonyms, and dictionary errors, directly guiding lexicographic enhancement (Flati et al., 2014).
• Reverse dictionary search: Node-graph similarity architectures enable efficient tip-of-the-tongue lexical retrieval, achieving high accuracy in semantic retrieval benchmarks (Thorat et al., 2016).
• Signal representation and classification: Graph dictionary learning enables interpretable, sparse decompositions of multivariate data and improved supervised classification in neuroimaging and time-varying graph data (Chen et al., 2016, Cappelletti et al., 2024).

6. Methodological Variants and Extensions

Beyond word–definition graphs, dictionary graph paradigms generalize to a broad class of signals and structures:

• Graph Dictionary Embedding (GDE): Uses prototype graph atoms and variational adaptation to obtain input-specific dictionaries, enabling expressive, structure-sensitive representations for graph classification via optimal transport (Liu et al., 2023).
• Unregistered and labeled graph embedding: Online GDL and Fused Gromov-Wasserstein models enable sparse dictionary representations for graphs with variable size/label, with provable approximation bounds and practical fast kernelization (Vincent-Cuaz et al., 2021).
• Manifold-based community detection: LGDE utilizes the geometry of embedding spaces and diffusion processes for nonlinear semantic expansion (Schindler et al., 2024).

These methods are unified by their reliance on sparse coding, optimal transport, and community detection on graphs, adapted to different data modalities and problem regimes.

7. Cognitive and Theoretical Significance

Dictionary graphs provide a powerful analytical framework for understanding symbol grounding, lexicon acquisition, semantic network formation, and the informational structure of language. Core and MinSet words approximate a "definitional basis" for the lexicon, offering an answer to infinite regress in definition chains by privileging a minimal, psycholinguistically-adapted anchor set (Vincent-Lamarre et al., 2014, Picard et al., 2013). The linkage to classic models such as Paivio’s dual-code theory further connects dictionary graphs to the cognitive architecture of human language.

The spectrum of dictionary graph research—from minimal feedback sets in definition graphs, to adaptive dictionaries for graph signals, to diffusion-based semantic manifolds—demonstrates the centrality of graph-theoretic principles in lexicography, representation learning, and computational semantics.