Multilayer Lexical Networks

Updated 9 May 2026

Multilayer lexical networks are formal frameworks that represent words or concepts as nodes across multiple interacting layers, each reflecting different linguistic relations.
They use mathematical structures like supra-adjacency tensors and block matrices to model intra- and inter-layer connections, enabling computation of metrics such as centrality and clustering.
These networks support applications ranging from mental lexicon modeling and word acquisition to document summarization and neural representation analysis in language processing.

A multilayer lexical network is a formal framework in which lexical units (words, sentences, or concepts) are represented as nodes structured across multiple interacting layers, each capturing distinct types of linguistic, cognitive, or communicative relationships. This architecture supports simultaneous modeling of semantic, phonological, syntactic, morphological, discourse, or social relations—extending beyond traditional, single-relation lexical graphs. Multilayer lexical networks are mathematically grounded in tensor and supra-adjacency matrix representations, support custom inter- and intra-layer couplings, and enable the analysis of connectivity, robustness, growth, acquisition, evolution, and processing across the complete stratification of language and cognition.

1. Formal Definitions and Mathematical Structure

Let $V = \{1, ..., N\}$ denote a set of $N$ lexical units (words, sentences, or linguistically defined entities), and let $\mathcal{L} = \{1, ..., M\}$ refer to $M$ distinct layers. In canonical multilayer (multiplex) lexical networks, each layer $\alpha \in \mathcal{L}$ is a graph $G_\alpha = (V, E^{(\alpha)})$ , with adjacency matrix $A^{(\alpha)} \in \{0, 1\}^{N \times N}$ . The full structure is encoded by a fourth-order supra-adjacency tensor $A_{i\alpha, j\beta}$ :

$A_{i\alpha, j\beta} = \begin{cases} A^{(\alpha)}_{ij}, & \text{if } \alpha = \beta \ C^{(\alpha\beta)}_{ij}, & \text{if } \alpha \ne \beta \end{cases}$

where intra-layer adjacency is specified by $A^{(\alpha)}$ and inter-layer coupling by $N$ 0. In the multiplex case, $N$ 1 for a coupling parameter $N$ 2 and Kronecker delta $N$ 3.

The corresponding supra-adjacency block matrix $N$ 4 has block elements $N$ 5, supporting standard and multiplex-specific network measures. Supra-Laplacians, random walk propagators, and related tools are constructed accordingly (Stella et al., 2022, Stella et al., 2016, Margan et al., 2015).

2. Typology of Layers and Construction Principles

Different instantiations model layers at varying linguistic or cognitive granularity:

Lexical association layers: Free association, synonymy, taxonomic hierarchy, feature sharing, and semantic similarity graphs (e.g., WordNet, association norms) (Stella et al., 2016, Stella et al., 2017).
Phonological layers: Edit-distance-1 graphs over word phoneme strings; phonological neighborhood structures (Stella, 2020, Stella et al., 2016).
Syntactic and morphological layers: Syntactic dependencies, co-occurrence within specified window sizes, and morpheme sharing (Margan et al., 2015).
Discourse/document layers: Sentences or documents as nodes, with intra- and inter-unit edges based on similarity or overlap (Tohalino et al., 2017).
Social and communicative layers: Social networks vs. media/broadcast networks, supporting active and passive information transfer (Javarone, 2013).
Multimodal or multi-language layers: Visual inputs, orthographic-to-lexical couplings, cross-linguistic mapping (e.g., for multilingual mental lexica) (Huynh et al., 7 Nov 2025).

Mathematical construction depends on the data source and linguistic subsystem, with systematic preprocessing (lemmatization, segmentation, frequency thresholding), and often symmetric or directed, weighted or unweighted, binary or continuous edge representations. For example, in sentence multilayer networks, edge weights are typically cosine similarities of tf–idf vectors, and inter-layer weights are scaled by a tunable parameter $N$ 6 to reflect cross-document salience (Tohalino et al., 2017).

3. Key Multilayer Metrics and Network Phenomena

The multilayer formalism enables the computation of both layer-specific and cross-layer network descriptors:

Degree and strength: $N$ 7, summed or vectorial across layers; $N$ 8.
Participation coefficient: $N$ 9 quantifies layerwise diversity (Stella et al., 2016, Stella et al., 2016).
Multiplex clustering coefficients: Focus on cross-layer triangles (mutual neighbors in multiple layers) (Stella et al., 2017, Margan et al., 2015).
Centralities: Multiplex PageRank, random-walk-based accessibility, eigenvector centrality, and community detection are generalized to operate on the supra-adjacency structure (Stella et al., 2022, Stella, 2020, Tohalino et al., 2017).
Viability/Largest Viable Cluster (LVC): The set $\mathcal{L} = \{1, ..., M\}$ 0 such that, for all layers, each induced subgraph $\mathcal{L} = \{1, ..., M\}$ 1 is connected; captures the intersectional connectivity essential for robust lexical access (Stella et al., 2017, Stella et al., 2022).
Robustness profiles: $\mathcal{L} = \{1, ..., M\}$ 2 and $\mathcal{L} = \{1, ..., M\}$ 3 track the sizes of the giant component and viable kernel under progressive node removal, revealing catastrophic phase transitions under multi-centrality attacks but robustness under random or single-layer attacks (Stella, 2020).
Layer overlap: Metrics such as Jaccard index, preserved weighted ratios, and motif correlations quantify structural affinities among layers (Margan et al., 2015).

4. Cognitive and Algorithmic Applications

Multilayer lexical networks provide a foundation for modeling cognitive processes and linguistic tasks:

Mental lexicon modeling: Developmental, clinical, and cognitive interpretation of lexicon structure as multiplex networks, revealing the emergence of the LVC around age 7–8 and its alignment with polysemy-rich, high-frequency core vocabularies (Stella et al., 2017, Stella et al., 2022).
Word learning and acquisition: Age-of-acquisition trajectories are best modeled using multiplex centralities or aggregate measures, rather than single-layer statistics. Distinct phases of acquisition are revealed by the shifting importance of phonological, semantic, and usage-based layers (Stella et al., 2016, Stella et al., 2016, Citraro et al., 2022).
Summarization and document representation: For multi-document extractive summarization, sentence-level multilayer networks with explicit intra/inter-document distinction improve ROUGE-1 recall, with tuning of inter-layer weights boosting informativeness and diversity of selected sentences (Tohalino et al., 2017).
Lexical innovation and diffusion: The spread of new words or meanings in social-media multiplexes demonstrates how structural coupling mediates consensus, misunderstanding, and persistent polysemy, depending on active/passive channel ratios and network topology (Javarone, 2013).
Neural network modeling: Deep CNNs trained for either phonological or semantic objectives instantiate separate "multilayer lexical networks" in the sense of layerwise feature subspaces aligned with dorsal (articulatory) and ventral (semantic) pathways, supporting the emergence of dual lexica via architectural pressures (Avcu et al., 2021).
Lexical semantics in neural LLMs: Layerwise probing of transformer models (LLMs) reveals the trajectory of lexical semantic encoding; lower to middle layers optimize sense discrimination, whereas higher layers shift toward next-token prediction, with direct parallels to multilayer lexical network stratification (Liu et al., 2024).

5. Empirical Findings and Quantitative Validation

Recent studies provide robust empirical support for the multilayer lexical network paradigm:

Explosive transitions: Acquisition simulations with multiplex networks demonstrate abrupt increases in the LVC at specific developmental timepoints. For example, per (Stella et al., 2017), the LVC emerges at $\mathcal{L} = \{1, ..., M\}$ 4 years, with a sudden addition of roughly 420 words, an effect sensitive to polysemy-degree correlations and not reproduced by null models.
Robustness and core identification: Removal of high multiplex-degree or PageRank nodes leads to catastrophic collapse of connectedness at $\mathcal{L} = \{1, ..., M\}$ 5 (GCC) and of viability at $\mathcal{L} = \{1, ..., M\}$ 6 (LVC), while random or single-layer attacks do not induce such collapse (Stella, 2020).
Language learning predictions: Multiplex centralities (especially closeness) and conformity-aware random-walk models achieve up to 75% accuracy in predicting CDI-based word learning stages in toddlers, outperforming structure-only or feature-only baselines (Citraro et al., 2022, Stella et al., 2016).
Structural coupling and motif profiles: Analyses on Croatian and English corpora reveal that word-level layers are power-law distributed and highly overlapping, while subword layers (syllables, graphemes) encode system-specific features, as quantified by overlap and triad significance profiles (Margan et al., 2015).

6. Theoretical and Computational Limitations

Several challenges and open questions persist in multilayer lexical network research:

Parameter estimation and model selection: Psychological and corpus data seldom provide direct measures for inter-layer coupling strengths $\mathcal{L} = \{1, ..., M\}$ 7, requiring principled methods for parameter inference (Stella et al., 2022).
Sparsity vs. redundancy: Adding layers can introduce spurious or redundant links. Structural reducibility analyses (e.g., Von Neumann entropy maximization) are essential for maximal interpretability (Stella et al., 2022).
Beyond pairwise relations: Many phenomena (morphology, polysemy, construction grammar) require higher-order modeling—hypergraphs or simplicial complexes extend the accessible representational space (Stella et al., 2022).
Scalability: The supra-adjacency representation for $\mathcal{L} = \{1, ..., M\}$ 8 and $\mathcal{L} = \{1, ..., M\}$ 9 can be computationally prohibitive. Efficient sparse and parallel algorithms for centrality, community detection, and percolation analysis are under active development (Stella et al., 2022, Margan et al., 2015).
Empirical coverage: Some proposed models (e.g., multilingual multimodal multiplex structures) have only conceptual validation and await large-scale experimental deployment (Huynh et al., 7 Nov 2025).

7. Broader Impact and Future Directions

Multilayer lexical networks provide a principled bridge between cognitive, computational, and linguistic sciences:

Cognitive kernels: The existence and role of LVCs support theories of robust core vocabulary, facilitate modeling of aphasia/degradation, and inform clinical lexical restoration strategies (Stella, 2020, Stella et al., 2017).
Behavioral and neurocognitive integration: Design of architectures that link neuroimaging-derived brain networks to lexical multiplexes is an emerging research focus (Stella et al., 2022).
Multimodality and multilinguality: Architectures that integrate visual modalities and cross-linguistic layers model the co-activation characteristic of bi- and multilingual individuals, and can be mapped to behavioral and neural outcomes (Huynh et al., 7 Nov 2025).
Extensions in NLP and AI: Layerwise analysis of representation learning in LLMs recapitulates key multilayer lexical network dynamics, suggesting direct applicability for interpretability and testable predictions in artificial neural systems (Liu et al., 2024).
Dynamic and evolutionary modeling: Increasing attention is paid to the simulation of lexicon growth, innovation diffusion, and diachronic stability under multi-network constraints (Stella et al., 2016, Javarone, 2013).

The multilayer lexical network formalism is now established as a foundational paradigm for capturing, analyzing, and predicting the structural dynamics of language in both human and artificial systems, with significant implications for lexical acquisition, mental representation, cognitive resilience, and linguistic theory.