Conceptual Networks for Cross‑Linguistic Idiomatic Expressions: A Feature‑Based Graph Approach Kiran Pala∗ University of Eastern Finland, Finland [email protected] Punam Silu Indian Institute of Technology Ropar, India [email protected] Abstract
arXiv:2607.09576v1 [cs.CL] 10 Jul 2026
We present an interpretable network-based framework for representing idiomatic and figurative meaning across eight typologically diverse languages, totaling 160 conventional expressions, the large majority of which are idiomatic. Each expression is annotated with binary conceptual features (containment, concealment, emotional, social, etc.) derived from cognitivelinguistic theory, and pairwise Jaccard similarities define a weighted graph. Community detection reveals that idioms cluster by conceptual schema rather than by language, producing a structure consistent with cognitive-linguistic predictions. The conceptual network captures unique semantic information not present in distributional embeddings, can be scaled via automatic annotation with LLMs, improves downstream idiom detection, and remains robust when enriched with corpus frequencies. Cross-lingual transfer experiments show that conceptual proximity alone can identify acceptable translation equivalents across five language families, with substantial gains over embedding-based baselines. Ablation studies demonstrate that all three feature dimensions—schemas, roles, and valence—contribute non-redundantly to both the network’s organizational properties and its performance on idiom detection, and that specific graph-derived signals (community membership, neighbor similarity) are particularly informative. The framework offers an interpretable, cross-linguistically stable representation of idiomatic meaning, combining theoretical grounding with practical utility.
1
Introduction
Idioms such as spill the beans or kick the bucket pose a fundamental challenge for semantic rep∗
Corresponding author.
Lixun Yu University of Eastern Finland, Finland [email protected] resentation because their figurative meaning cannot be reliably derived from literal constituents. Cognitive linguistics has shown that many idioms are motivated by a small set of image schemas—e.g., containment, concealment—that recur across typologically diverse languages (Kövecses, 2005). However, the degree of conceptual overlap between idioms is graded rather than all-or-none, and modeling this graded structure can both advance theoretical understanding and improve natural language processing applications that require handling of non-literal language. Interpretable representations are particularly valuable for NLP tasks where model decisions must be understood and trusted. By grounding representation in cognitive theory, our framework offers both interpretability and crosslingual stability, complementing black‑box embedding models. Most computational models of idioms rely on surface lexical forms through distributional semantics or contextual embeddings (Fazly et al., 2009). While powerful, such approaches often miss the deep conceptual structure that underlies idiomaticity across languages, and they provide little insight into the semantic organisation of idioms. Furthermore, existing studies predominantly focus on a small number of well-resourced languages, neglecting the typological diversity that could reveal universal aspects of idiom organisation. Large language models can occasionally identify idioms, but they do not offer a structured conceptual space in which idiomatic meaning can be systematically compared across languages. We introduce a network-based model of 160 idioms from eight languages representing five language families: Indo-European (English, Hindi, Bagri, Spanish), Uralic (Finnish), Japonic (Japanese), Sino-Tibetan (Chinese),
and Dravidian (Malayalam). This selection spans agglutinative, fusional, and isolating morphologies; SVO and SOV word orders; and five distinct writing systems. The inclusion of Bagri and Malayalam adds crucial data points rarely present in idiom studies. The model represents each idiom by a small set of binary conceptual features grounded in cognitive-linguistic theory, and constructs a weighted graph using the Jaccard coefficient as edge weight. We demonstrate that the network organises idioms by conceptual schema rather than by language, and we evaluate its properties through the following complementary analyses: (1) community structure and comparison with multilingual distributional embeddings, (2) downstream idiom detection task evaluation, (3) ablation analyses on feature dimensions, (4) automatic feature extraction via LLMs, (5) dynamic enrichment with corpus frequencies, and (6) crosslinguistic transfer. Together, these analyses show that the conceptual network is a robust, interpretable, and practically useful representation of idiomatic meaning.
2
Related Work
Cognitive Linguistics and Idioms. Conceptual metaphor theory (Lakoff and Johnson, 2008) proposed that human thought is structured by a relatively small set of image schemas, which are then realized in language. Gibbs (Gibbs, 1992) showed that speakers’ mental imagery for idioms is highly consistent and reflects underlying conceptual metaphors. Cross-linguistic studies (Piirainen, 2012) identified widespread idioms that share the same conceptual structure across European and Asian languages, suggesting a universal basis for many idiomatic expressions. These insights motivate our annotation scheme, which captures both universal schemas and languagespecific roles. Computational Models of Idioms. In NLP, idioms have been modelled using distributional vectors (Fazly et al., 2009), often focusing on detecting whether an expression is used literally or figuratively. However, such models rely on surface co-occurrence statistics and may fail to capture deep conceptual similarities across languages. Graph-based seman-
tic models (Steyvers and Tenenbaum, 2005) have revealed small-world structure and cognitive correlates, but they have rarely been applied to idioms. Our framework bridges this gap by constructing a conceptual network from theory-grounded features. Existing resources for idiomatic expressions are heavily skewed towards English and a few European languages; our multilingual corpus, low‑resource inclusive dataset addresses this imbalance.
3 Methodology 3.1 Idiom Selection and Corpus Verification We compiled 20 conventional figurative expressions (hereafter referred to as idioms) from each of English, Hindi, Bagri, Finnish, Japanese, Spanish, Chinese, and Malayalam, resulting in 160 entries. This broad, cognitively motivated usage of the term idiom encompasses not only core phrasal idioms but also proverbs, conventional metaphors, and other fixed expressions whose meaning is not fully compositional; all share the property of being culturally entrenched figurative units, which is the phenomenon of interest here. All expressions were verified by at least two native speakers and checked for frequency in appropriate reference corpora: COCA (Davies, 2008) (English), EMILLE (Baker et al., 2002) (Hindi), a custom Bagri elicitation set was used, since Bagri (a Rajasthani Indo-Aryan variety) has no standardized written corpus and these figurative expressions were confirmed with three native speakers (all collected with informed consent and anonymized), Suomi24 (Bartis, 2015) (Finnish), BCCWJ (Maekawa et al., 2008) (Japanese), CREA (Real Academia Española, 2008) (Spanish), BCC/CCL (Xun et al., 2016) (Chinese), and the Malayalam Web Corpus (Kilgarriff et al., 2010). The idioms collected for this study cover common conceptual domains such as communication, emotion, secrecy, and social interaction. They include both cross-linguistically shared metaphorical patterns and idiomatic expressions that are specific to particular languages and cultural contexts. The full dataset will be released upon publication.
3.2
Feature Annotation Scheme
Each idiom was manually annotated with binary features drawn from three independent conceptual dimensions, as depicted in Figure 1. The dimensions are grounded in cognitivelinguistic theory and were chosen to capture the core conceptual and functional properties that distinguish idiomatic meanings. • Conceptual schemas: containment, concealment_S, emotional, social. These image-schematic structures arise from recurring patterns of bodily experience and have been extensively documented as the conceptual building blocks of metaphorical thought (Lakoff and Johnson, 2008). Containment reflects the bounded region schema (a container with an inside and an outside), concealment_S the schema of hiding or covering, emotional the mapping of internal affective states onto physical sensations, and social the schema of interpersonal interaction and group membership. An idiom receives the containment label if its figurative meaning involves keeping something within a bounded space (e.g., bottle up your feelings), concealment if it implies deliberate hiding (e.g., sweep it under the rug), emotional if it directly expresses or describes an affective state (e.g., wear your heart on your sleeve), and social if it inherently involves an interpersonal dynamic (e.g., blow the whistle). • Functional roles: communication, concealment_R, emotional_state. This dimension captures the pragmatic function that the idiom typically serves in discourse, following the insight that many idioms are not merely descriptive but perform specific communicative acts (Gibbs, 1992). Communication applies when the idiom’s primary function is to convey or elicit information (e.g., spill the beans), concealment_R when it describes the act of hiding information or intentions (e.g., keep it under your hat), and emotional_state when it ascribes a specific emotion to an experiencer (e.g., down in the dumps). An idiom can serve multiple functions simultaneously; for instance,
bite your tongue both describes concealment of speech and signals an emotional state of restraint. • Affective valence: positive, negative. This dimension reflects the evaluative polarity of the idiom’s figurative meaning. Positive valence is assigned when the idiom denotes a desirable, pleasant, or socially approved outcome (e.g., get it off your chest, bury the hatchet), while negative valence indicates an undesirable, unpleasant, or socially disapproved situation (e.g., skeleton in the closet, sweep it under the rug). Valence was assessed independently of the literal reading of the words; for example, swallow your pride is tagged as negative even though pride itself can be positive, because the overall figurative act describes an unpleasant self-suppression. The dimensions are not mutually exclusive; an idiom can be associated with multiple features. For example, spill the beans is tagged {communication, social, negative} because it describes revealing a damaging secret in a social context, while keep it under your hat receives {containment, concealment_R, communication, social} as it combines the image of a covered container with the communicative act of secret-keeping. The annotation scheme was designed with clear diagnostic criteria—for each feature, a decision tree with prototypical and borderline examples was provided to annotators—and was applied by three trained linguists with expertise in semantics. Inter-annotator agreement was high (Fleiss’ κ > 0.78 for all dimensions), and all disagreements were resolved through consensus discussion. Table 1 shows a representative sample across all eight languages.
Idiom spill the beans
Feature Space Schemas
Roles
containment, concealment_S, emotional, social
communication, concealment_R, emotional_state
Valence positive, negative
Jaccard directly quantifies proportional overlap, is bounded, symmetric, and computationally minimal—a resource-rational choice consistent with models of similarity (Tversky, 1977; Lieder and Griffiths, 2020). The 160 × 160 matrix W = [wij ] defines a weighted graph G = (V, E); edges with wij > 0 are retained (density 0.37).
4 Results 4.1 Community Structure and Topological Properties
Feature set S(vi ) {communication, social, negative}
The network (Figure 2) separates into two major clusters: containment/concealment/emotional and communication/social. Louvain community Figure 1: Feature annotation schema. An idiom is et al., 2008) yields 5 detection (Blondel decomposed into three conceptual dimensions. communities with modularity Q = 0.61. NMI with language labels (8 groups) is only 0.10, Table 1: Example idioms with conceptual features. while NMI with primary schema partition is 0.76 (ARIlang = 0.05, ARIschema = 0.69). Lang. Idiom Gloss Feature set This confirms that conceptual features, not English spill the — {comm., language, drive the network’s organisation. beans social, Beyond community structure, the concepneg.} tual graph exhibits several characteristic propSpanish guardar un keep a secret {cont., secreto bajo under lock conc._R, erties of well‑organized semantic networks. llave and key comm.} The degree distribution follows a power‑law Finnish haudata asia bury the mat- {conc._S, with an exponential cutoff (average degree ter emot.} ⟨k⟩ = 23.1, clustering coefficient C = 0.70). The centrality of betweenness is highest for Hindi पे ट में बात रखना keep matter {cont., in stomach conc._R, the bridge idioms (Table 2); spill the beans emot.} reaches 0.15, confirming its role in connecting the two major-clusters. The network shows Malayalam മനസ്സിൽ keep in mind {cont., സൂക്ഷിക്കുക conc._R, small‑world organization with a small‑world emot.} index σ = 2.0 (compared to σ = 1.0 for a random graph of equal size and density). Bagri अकल बिना ऊंट Camels roam {social, The global efficiency remains at 88% after उभाणा फिरै barefoot with- commuout wisdom nication, random removal of 20% of nodes, indicating neg.} high resilience. These properties align with those observed in human‑derived semantic netJapanese 中を見せる show one’s in- {emot., sides comm., works from free association and feature norms social} (Steyvers and Tenenbaum, 2005; Collins and Chinese 守口如瓶 guard mouth {cont., Loftus, 1975), supporting the representational like bottle conc._R, validity of the conceptual graph. comm.}
3.3
Network Construction
For each pair of idioms (vi , vj ), the edge weight is the Jaccard coefficient: wij =
|S(vi ) ∩ S(vj )| . |S(vi ) ∪ S(vj )|
Table 2: Top‑5 bridging idioms by betweenness centrality.
confirmed that the network-derived features were the most discriminative non-embedding predictors.
Lang.
Idiom
Feature set
BC
English
spill the beans
0.15 4.4
Hindi
दिल खोलकर बात करना sacar a la luz 胸の内を明かす
{comm., social, neg.} {emot., comm., social, pos.} {comm., social} {emot., comm., social} {comm., social, neg.}
Spanish Japanese Malayalam
4.2
രഹസ്യം ചോ രുക
0.12 Having established that the network-derived
features improve downstream idiom detection,
0.11 we next conducted ablation analyses to exam0.10
ine which annotation dimensions and graph-
0.09 derived signals drive this improvement.
Comparison with Multilingual Embeddings
We extracted phrase embeddings for all 160 idioms using XLM-Rbase and built a graph from cosine similarities. The cosine graph yielded a much lower NMI with the schema partition (0.45). A permutation test (10,000 random reassignments of language and schema labels) confirms that the difference between the Jaccard- and cosine-based NMI values is significant (p < 0.001). Combining Jaccard and cosine similarities did not improve the fit, confirming that conceptual features encode unique information not present in multilingual embeddings. Eng
Containment
Communication
H1
Hin
E1 F1
Spa
Chi
Mal
S2
E3
J2 Z1
J1 F2 S1 J3 Z2
M2
M1 B1
Bag
Figure 2: Conceptual network (schematic, based on force‑directed layout of the 160‑idiom Jaccard graph). Two clusters: containment (left, squares) and communication (right, diamonds). Bridging idioms (dashed edge) connect the communities.
4.3
To evaluate the contribution of each feature dimension, we constructed reduced networks by systematically removing one dimension at a time. Table 3 reports the resulting NMI with the full schema partition. The full model achieves the highest NMI (0.76). Removing schema information causes the largest decrease in alignment (NMI = 0.58), indicating that schemas play the strongest role in structuring the network. Removing roles (NMI = 0.67) or valence (NMI = 0.64) also weakens alignment, suggesting that these dimensions contribute additional and partly complementary information. Table 3: Ablation of feature dimensions on network structure. Model
NMI
Full (all 9 features) – Schemas (roles + valence) – Roles (schemas + valence) – Valence (schemas + roles) – Random baseline
0.76 0.58 0.67 0.64 0.02
Containment
Communication
E2
Fin Jpn
H2
Ablation Analyses
Downstream Idiom Detection
We evaluated the network on SemEval-2013 subtask 5b (semantic compositionality in context), which can also be used for metaphor or idiom detection (Korkontzelos et al., 2013). A BERT-based classifier achieved baseline F1 0.82. Adding network features (degree, betweenness centrality, community membership, and mean Jaccard similarity to its five nearest conceptual neighbours) improved F1 to 0.86 (95% bootstrap CI [0.83, 0.89]; p < 0.01). The gain was consistent across genres (news, fiction, etc.), and feature importance analysis
We further tested the impact of each dimension on downstream idiom detection. Removing schemas reduced the F1 gain from +4.0 to +1.2 points; removing valence reduced it to +2.5 points; removing roles yielded a gain of +2.9 points. These results suggest that all three annotation dimensions contribute to the downstream benefit, with schema information accounting for the largest share of the improvement. Finally, we ablated the individual graphderived signals used as features for the downstream classifier. Table 4 shows the remaining F1 gain when each signal is removed. The largest drop occurs when top‑5 neighbour similarity or community membership is removed, suggesting that the classifier benefits particularly from information about an idiom’s posi-
tion within the conceptual space and its similarity to nearby expressions.
The resulting network retained the original community structure (NMI 0.88 with the original partition) and the small-world index reTable 4: Downstream ablation of network features. mained σ = 2.0, indicating that conceptual schemas dominate even when usage statistics Removed features ∆F1 vs baseline are included. None (full network features) +4.0 4.7 Cross-Linguistic Transfer – degree +3.1 To evaluate the network’s ability to act as – betweenness +3.4 an interlingua, we selected, for each English – community membership +2.9 idiom, its nearest conceptual neighbor (high– top‑5 neighbour similarity +2.6 est Jaccard similarity) in each of the other seven languages. Three co-authors and three 4.5 Automatic Feature Extraction additional native speakers of the target lanwith LLMs guages independently provided gold-standard translation equivalents for the 20 English idTo test scalability, we prompted GPT-4 with ioms. Where exact equivalent was not availthe idiom and its gloss, requesting binary able, the most contextually appropriate figannotations for all nine features, using ten manually annotated idioms as few-shot exam- urative expression was selected. To reduce potential bias, translators did not have acples. In the remaining 150 idioms, the LLM achieved an accuracy of 0.82 and macro-F1 cess to Jaccard similarity scores when making 0.79. The network built from LLM annota- these judgments. For each pair of Englishtions reproduced the same two-cluster struc- target language, the Jaccard-selected canditure (NMI 0.74 with the manual schema parti- date was then compared against the gold standard. The same procedure was repeated ustion), demonstrating that the framework can be extended to larger datasets without exhaus- ing the nearest neighbor identified by XLM‑R cosine similarity. Overall, the Jaccard‑based tive manual coding. selection matched the gold standard in 78% Error Analysis. Qualitative inspection of LLM errors revealed three main failure modes. of cases (109 out of 140 possible pairs), whereas the XLM‑R baseline achieved only First, culture‑specific expressions were often 54%. This demonstrates that conceptual proxmislabeled, e.g., the Bagri idiom आम खाणा क imity alone can identify cross‑lingual equivaपे ड़ गीणना (Should one eat mangoes or count lents with high accuracy, a result with direct trees?) was annotated as emotional, positive rather than communication, social. Second, implications for machine translation and secabstract internal states proved difficult; emo- ond‑language learning. tional_state was the least accurate label (F1 0.73), with the model confusing emotional descriptions with emotional functions. Third, valence ambiguity affected idioms that can be positive or negative depending on context (e.g., open up sometimes mislabeled as negative). These findings indicate that while LLM‑based scaling is viable, high‑accuracy annotation of culture‑bound and psychologically nuanced features still benefits from human expertise. 4.6
Dynamic Enrichment with Corpus Frequencies
We modulated edge weights with point-wise mutual information (PMI) from COCA and ′ = w · (1 + 0.2 · PMI ). analogous corpora: wij ij ij
5 Discussion The conceptual network offers a robust, interpretable, and practical representation of idiomatic meaning. Its community structure aligns with cognitive-linguistic theory, yet it is validated through computational experiments, making it suitable for NLP applications. The clustering pattern suggests that containment is a productive conceptual schema within the annotated idiom network and points to a cross-linguistic convergence. The overlap with communication-related idioms indicates that the two domains are not fully separable, since communicative meanings such as disclosure, concealment, and information transfer are often conceptualized through contain-
ment relations. The ablation studies confirm that each feature dimension contributes nonredundant information, while the comparison with embeddings shows that conceptual features capture semantic aspects orthogonal to distributional statistics. The downstream improvements in idiom detection, combined with strong cross-lingual transfer, demonstrate that the network can be directly integrated into NLP pipelines. Complementarity with POSI. Our framework is complementary to the geometric semantic model and POSI annotation scheme (Pala et al., 2025; Pala, 2024). POSI provides fine-grained tags grounded in spatial, temporal, and experiential dimensions; its 135 four-letter tags can capture subtle distinctions (e.g., MCEN for “middle centre” vs. STCN for “structural container”) that our coarse schemas collapse into a single containment feature. In contrast, our nine idiom‑specific features abstract away from lexical detail to highlight the conceptual backbone shared by figurative expressions—something that POSI’s general-purpose vocabulary, when applied to idioms, may obscure. The two schemes can enrich each other: POSI tags could decompose our coarse containment schema into more precise spatial relations, while our network could serve as a testbed for evaluating whether POSI’s fine‑grained distinctions are cognitively salient for figurative language. Future work could annotate our full 160-idiom dataset with POSI tags, creating a richly layered resource that marries idiomatic abstraction with general semantic precision, and use the resulting integrated graph to probe the universality of both frameworks. 5.1
Limitations and Future Work
Although LLM annotation shows promise, certain features (e.g., emotional state) still require manual refinement. Future work will explore fine-tuned models and improved prompting strategies to scale the framework to hundreds of languages. Another limitation is the binary nature of our features, which cannot express gradable properties such as intensity of emotion or partial containment; extending the annotation to ordinal or continuous values could capture more nuanced similarities. The current dataset, though carefully curated and
multilingual, remains relatively small; expanding it and adding languages from more families will strengthen claims of universality. Finally, integrating the conceptual network into more complex tasks such as machine translation or sentiment analysis of figurative language remains an open direction.
6 Conclusion We presented a conceptual cross-linguistic network of 160 idioms annotated with features based on theory. The network organizes idioms by conceptual schema, outperforms distributional embeddings in capturing conceptual structure, can be scaled via automatic annotation, improves idiom detection, and enables accurate cross-lingual idiom mapping. Ablation studies confirm the complementary contributions of schemas, roles, valence, and individual graph-derived signals. The framework is interpretable, computationally validated, and has strong potential for NLP applications.
Ethics Statement The Bagri and Chinese idiom lists were compiled and verified by native‑speaker co‑authors – (for Bagri, oral transcriptions were collected with informed consent and fully anonymized). External human subjects were not recruited for the experiments; cross‑lingual transfer evaluation relied solely on translation equivalents provided by the co‑author team. Corpus frequency counts (COCA, etc.) were obtained in fair use for non‑commercial research; the public release contains only derived PMI statistics, not the original text.
Availability The complete annotated dataset, the 160×160 will be released upon publication.
References Baker, P., Hardie, A., McEnery, T., Cunningham, H., & Gaizauskas, R. (2002). EMILLE, a 67-million word corpus of Indic languages: Data collection, mark-up and harmonisation. In M. González Rodríguez & C. P. Suarez Araujo (Eds.), Proceedings of the Third International Conference on Language Resources and Evaluation (LREC’02), Las Palmas, Canary Islands
- Spain: European Language Resources Association (ELRA). https://aclanthology.org/ L02-1319/
Piirainen, E. (2012). Widespread idioms in Europe and beyond: Toward a lexicon of common figurative units. New York: Peter Lang.
Bartis, I. (2015). The Suomi 24 Sentences Corpus (2016H2). FIN-CLARIN Consortium, Department of Modern Languages, University of Helsinki.
Real Academia Española. (2008). Corpus de Referencia del Español Actual (CREA), v3.2.
Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment, P10008, 1–12. https://hal.science/hal-01146070 Collins, A. M., & Loftus, E. F. (1975). A spreadingactivation theory of semantic processing. Psychological Review, 82(6), 407–428. Davies, M. (2008–). The Corpus of Contemporary American English (COCA). https://corpus. byu.edu/coca/. Fazly, A., Cook, P., & Stevenson, S. (2009). Unsupervised type and token identification of idiomatic expressions. Computational Linguistics, 35(1), 61–103. Gibbs, R. W. (1992). What do idioms really mean? J. Memory and Language, 31(4), 485–506. Kilgarriff, A., Reddy, S., Pomikálek, J., & Avinesh, P. (2010). A corpus factory for many languages. Proc. LREC 2010, 2123–2127. Korkontzelos, I., Zesch, T., Zanzotto, F. M., & Biemann, C. (2013). SemEval-2013 Task 5: Evaluating phrasal semantics. Proc. SemEval 2013, 39–47. Kövecses, Z. (2005). Metaphor in culture: Universality and variation. Cambridge University Press. Lakoff, G., & Johnson, M. (2008). Metaphors We Live By. University of Chicago Press. Lieder, F., & Griffiths, T. L. (2020). Resourcerational analysis: Understanding human cognition as the optimal use of limited computational resources Behav. Brain Sci., 43, e1. Maekawa, K., et al. (2008). Balanced Corpus of Contemporary Written Japanese. In Proceedings of the 6th workshop on Asian language resources, 48(2), 345–371. Pala, K. (2024). Searching for Functional Boundaries: Evaluating Effectiveness in Complex Adaptive Networks with Cognitive Dynamics.Proc. CogSci, 46. Pala, K., Shalu, S., Nedumpozhimana, V., & Choudhary, K. K. (2025). A geometric semantic model and POSI annotation framework. Front. Artif. Intell., 8, 1666074.
Steyvers, M., & Tenenbaum, J. B. (2005). The Large-Scale Structure of Semantic Networks: Statistical Analyses and a Model of Semantic Growth. Cognitive Science, 29(1), 41–78. Tversky, A. (1977). Features of similarity. Psychological Review, 84(4), 327–352. Xun, E., Rao, G., Xiao, X., & Zang, J. (2016). ������BCC������ [The development of the BCC corpus in the context of big data]. ������ Corpus Linguistics, 3(1), 93–109. https://aclanthology. org/2025.ccl-1.35.pdf#3#3