💬 NLP

Conceptual Networks for Cross-Linguistic Idiomatic Expressions:A Feature-Based Graph Approach

This paper introduces an interpretable, feature-based graph framework that represents idiomatic expressions across eight diverse languages through cognitive-linguistic conceptual features, demonstrating that such networks cluster by semantic schema rather than language and outperform distributional embeddings in cross-lingual translation and idiom detection tasks.

Original authors: Kiran Pala, Punam Silu, Lixun Yu

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kiran Pala, Punam Silu, Lixun Yu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, messy closet full of 160 different idioms—those tricky phrases like "spill the beans" or "kick the bucket"—from eight very different languages, including English, Hindi, Japanese, and even Bagri (a language from Rajasthan, India). Usually, if you try to sort these idioms, you'd expect them to group by language: all the English ones together, all the Japanese ones together, like sorting socks by color.

But this paper suggests something much more interesting. The researchers built a special map, or a "conceptual network," where each idiom is a dot. Instead of sorting by language, they sorted them by the hidden ideas inside them. They asked: Does this phrase involve hiding something? Is it about feelings? Is it a social interaction?

The Big Discovery: Ideas Over Languages
When they connected the dots based on these hidden ideas, the map didn't look like a language closet. It looked like a theme park with two main zones. One big cluster was all about containment and hiding (like keeping a secret in a box), and the other was about communication and socializing (like talking or revealing secrets).

The paper shows that a Japanese idiom about hiding a secret is actually "closer" to an English idiom about hiding a secret than it is to another Japanese idiom about, say, being happy. The network proved that conceptual features drive the organization, not language. In fact, when they tried to sort the idioms by language, the map barely recognized the groups (a score of only 0.10), but when they sorted by the hidden ideas, it was a perfect match (a score of 0.76).

Why Old Maps Failed
The researchers tested this against the "black box" maps that computers usually use (called distributional embeddings). These old maps try to guess meaning by looking at which words appear near each other in text. The paper argues that these old maps miss the deep, structural soul of an idiom. When they compared the new idea-based map to the old word-neighborhood map, the new one was significantly better at finding the true conceptual groups. The old maps just couldn't see the "hiding" or "feeling" patterns that the new map spotted clearly.

The "Bridge" Idioms
Some idioms act like bridges between the two zones. The paper highlights "spill the beans" as a superstar bridge. It has a high "betweenness centrality" score of 0.15, meaning it's the main path connecting the "hiding" world to the "talking" world. It's the perfect example of taking something hidden and making it public.

Does It Actually Help Computers?
The paper doesn't just make a pretty picture; it tests if this map helps computers do better jobs.

Idiom Detection: When they gave a computer program a standard test to spot idioms, it got an F1 score of 0.82. But when they added this new map's data (like "how many neighbors does this idiom have?" or "which group does it belong to?"), the score jumped to 0.86. That's a real, measurable improvement.
Translation: They tried to use the map to find the right translation for an English idiom in seven other languages. The map picked the correct equivalent 78% of the time. The old computer method only got it right 54% of the time. This suggests that if you want to translate idioms, looking at the concept is way better than just looking at the words.

The "Secret Sauce" and the Limits
The researchers broke down their map to see which parts mattered most. They found that all three parts of their system—schemas (the big ideas like "hiding"), roles (what the idiom does, like "communicating"), and valence (whether it's positive or negative)—are necessary. If you remove the "schemas," the map falls apart the most.

They also tried to see if a super-smart AI (an LLM) could draw this map automatically instead of humans doing it. The AI did a pretty good job, getting 82% accuracy, but it still struggled with tricky, culture-specific jokes or very subtle emotional states. The paper suggests that while AI can scale this up to hundreds of languages, human experts are still needed for the really nuanced stuff.

What This Isn't
The paper is careful to say this isn't a magic bullet that solves everything. The features they used are binary (yes/no), so they can't capture how strong an emotion is, just that it exists. Also, the dataset, while diverse, is still relatively small (160 idioms), so we can't say for sure that this works for every idiom in the world yet.

The Takeaway
In short, this paper proposes that idioms from different languages are like different species of birds that all build nests in the same way. If you sort them by their DNA (the language), they look different. But if you sort them by how they build their nests (the conceptual ideas of hiding, feeling, or talking), they line up perfectly. This new "nest-building" map is more accurate, more helpful for computers, and more honest about how human brains actually organize these tricky phrases.

Technical Summary: Conceptual Networks for Cross-Linguistic Idiomatic Expressions

Problem Statement

Idiomatic expressions pose a fundamental challenge for semantic representation because their figurative meanings cannot be reliably derived from their literal constituents. While cognitive linguistics posits that many idioms are motivated by a small set of recurring image schemas (e.g., containment, concealment) across typologically diverse languages, computational models have largely relied on surface-level distributional semantics or contextual embeddings. These existing approaches often fail to capture the deep conceptual structure underlying idiomaticity, provide limited insight into the semantic organization of idioms, and predominantly focus on well-resourced languages, neglecting typological diversity. Furthermore, current Large Language Models (LLMs) lack a structured conceptual space for systematically comparing idiomatic meaning across languages.

Methodology

The authors propose an interpretable, network-based framework representing 160 conventional figurative expressions (idioms) from eight typologically diverse languages spanning five language families: Indo-European (English, Hindi, Bagri, Spanish), Uralic (Finnish), Japonic (Japanese), Sino-Tibetan (Chinese), and Dravidian (Malayalam).

1. Data Collection and Verification

The dataset comprises 20 idioms per language. Expressions were verified by native speakers and checked against reference corpora (e.g., COCA, EMILLE, BCCWJ). For Bagri, a Rajasthani Indo-Aryan variety lacking a standardized written corpus, expressions were confirmed via oral elicitation with native speakers.

2. Feature Annotation Scheme

Each idiom is annotated with binary features derived from three independent conceptual dimensions grounded in cognitive-linguistic theory:

Conceptual Schemas: Containment (bounded region), Concealment_S (hiding/covering), Emotional (mapping affect to physical sensation), and Social (interpersonal interaction).
Functional Roles: Communication (conveying information), Concealment_R (hiding information), and Emotional_State (ascribing emotion).
Affective Valence: Positive or Negative based on the figurative outcome.

Annotations were performed by three trained linguists with high inter-annotator agreement (Fleiss' $\kappa > 0.78$ ).

3. Network Construction

A weighted graph $G = (V, E)$ is constructed where nodes represent idioms. The edge weight $w_{ij}$ between two idioms is calculated using the Jaccard coefficient of their feature sets $S(v_i)$ and $S(v_j)$ :
$w_{ij} = \frac{|S(v_i) \cap S(v_j)|}{|S(v_i) \cup S(v_j)|}$
This approach quantifies proportional conceptual overlap, resulting in a 160x160 matrix with a density of 0.37.

Key Contributions and Results

1. Network Topology and Community Structure

Community detection using the Louvain algorithm reveals that the network organizes idioms by conceptual schema rather than by language.

Clustering: The network separates into two major clusters: one combining containment/concealment/emotional schemas and another combining communication/social schemas.
Language Independence: The Normalized Mutual Information (NMI) between language labels and community structure is low (0.10), whereas the NMI with the primary schema partition is high (0.76).
Small-World Properties: The graph exhibits small-world organization ( $\sigma = 2.0$ ), high clustering ( $C = 0.70$ ), and high resilience (88% global efficiency after 20% node removal), aligning with properties of human-derived semantic networks.

2. Comparison with Distributional Embeddings

When compared to a graph built from XLM-R base cosine similarities, the Jaccard-based conceptual graph demonstrates superior alignment with schema partitions (NMI 0.76 vs. 0.45). A permutation test confirms this difference is significant ( $p < 0.001$ ), indicating that conceptual features encode unique semantic information not captured by multilingual embeddings.

3. Downstream Idiom Detection

Integrating network-derived features (degree, betweenness centrality, community membership, and neighbor similarity) into a BERT-based classifier for the SemEval-2013 idiom detection task improved the F1 score from a baseline of 0.82 to 0.86. Ablation studies confirm that all three feature dimensions (schemas, roles, valence) contribute non-redundantly, with schema information providing the largest performance gain.

4. Cross-Linguistic Transfer

The framework was tested as an interlingua for identifying translation equivalents. Selecting the nearest conceptual neighbor (highest Jaccard similarity) for English idioms in other languages matched gold-standard translations in 78% of cases, significantly outperforming the XLM-R cosine baseline (54%).

5. Scalability and Robustness

LLM Annotation: Automatic feature extraction using GPT-4 achieved 82% accuracy and reproduced the two-cluster structure (NMI 0.74), suggesting the framework can scale to larger datasets.
Dynamic Enrichment: Modulating edge weights with corpus frequencies (PMI) preserved the original community structure, confirming that conceptual schemas dominate even when usage statistics are included.

Significance and Claims

The paper claims that this framework offers a robust, interpretable, and cross-linguistically stable representation of idiomatic meaning. Its primary significance lies in bridging the gap between cognitive-linguistic theory and computational NLP:

Interpretability: Unlike black-box embedding models, the network provides a transparent structure where semantic relationships are defined by explicit, theory-grounded features.
Cross-Lingual Stability: The organization of idioms by conceptual schema rather than language supports the hypothesis of universal cognitive structures underlying figurative language.
Practical Utility: The framework improves downstream tasks (idiom detection) and facilitates cross-lingual transfer, offering a viable alternative or complement to distributional semantics for handling non-literal language.
Complementarity: The authors position their coarse-grained schema approach as complementary to fine-grained geometric models (like POSI), suggesting a path toward richer, layered semantic resources.

The authors remain modest regarding limitations, noting that binary features cannot express gradable properties (e.g., intensity) and that certain culture-specific or psychologically nuanced features still benefit from human expertise despite promising LLM results. Future work is directed toward expanding the dataset to more languages and integrating the network into complex tasks like machine translation.