💬 NLP

Measuring cross-language intelligibility between Romance languages with computational tools

This paper introduces a novel computational metric based on lexical, surface, and semantic similarity to measure and validate cross-language intelligibility among the five main Romance languages, finding that the results align with human experimental data and linguistic intuitions.

Original authors: Liviu P Dinu, Ana Sabina Uban, Bogdan Iordache, Anca Dinu, Simona Georgescu

Published 2026-02-10

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Liviu P Dinu, Ana Sabina Uban, Bogdan Iordache, Anca Dinu, Simona Georgescu

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

The "Language Mirror" Problem: Can You Understand Your Neighbor?

Imagine you are at a massive international dinner party. You speak Spanish, and the person sitting next to you speaks Portuguese. You both know you’re "cousins" because your languages sound similar, but you aren't quite sure: if they start telling a story, will you catch the plot, or will you just be nodding politely while feeling totally lost?

This paper is essentially a high-tech attempt to measure exactly how much "understanding" exists between the major Romance languages (Spanish, French, Italian, Portuguese, and Romanian) using computers instead of just guessing.

The Core Idea: The Three Pillars of Understanding

The researchers realized that understanding a "cousin" language isn't just about words looking the same on paper. They created a new formula called the DLI (Lexical Intelligibility Index). Think of it like a three-legged stool; if one leg is missing, the whole thing falls over.

The "Look" (Orthography): Does the word look like the one in your language? If you see "tempo" (Italian) and "tiempo" (Spanish), your brain goes, "Aha! I know that!"
The "Sound" (Phonetics): This is the tricky part. A word might look familiar on paper, but if it sounds totally different when spoken, the "mirror" is broken. The paper notes that French is a great example: it might look like its cousins on paper, but once a French person starts speaking, the sounds change so much that the "cousins" often don't recognize them.
The "Meaning" (Semantics): This is the "False Friend" trap. Imagine you and a friend use the word "cool" to mean "temperature," but your friend uses it to mean "fashionable." You’re using the same word, but you aren't actually communicating. The researchers used AI (word embeddings) to make sure the words actually mean the same thing, not just that they sound the same.

The Big Discovery: The "Asymmetry" Glitch

One of the most interesting things the researchers found is that intelligibility is asymmetrical. In science, "symmetrical" means if A understands B, then B understands A. But in language, it’s more like a one-way street.

The Romanian Paradox:
The study found that Romanian is a bit of a "rebel" in the Romance family.

The "One-Way Mirror": Romanian speakers are actually quite good at understanding other Romance languages. However, when other Romance speakers listen to Romanian, they often struggle to understand it.
Why? It’s partly because Romanian has picked up a lot of "Slavic" flavors (words from neighboring Eastern European languages) and has some unique historical twists. It’s like a cousin who moved to a different country and came back with a totally different accent and slang—you can understand them, but they sound like a stranger to you!

Why Does This Matter?

The researchers didn't just want to build a math formula; they wanted to see if a computer could "feel" language the way humans do. They compared their computer scores to real human tests (where people actually tried to fill in missing words in sentences).

The result? The computer was remarkably accurate. It matched the human experience with high correlation.

The Takeaway

By using AI to look at the look, sound, and soul (meaning) of words, we can finally map out the "family tree" of languages. This helps us understand how languages drift apart over centuries and how we might better bridge the gap between people who speak "cousin" languages but still struggle to find common ground.

Technical Summary: Measuring Cross-Language Intelligibility Between Romance Languages with Computational Tools

1. Problem Statement

The paper addresses the challenge of quantifying cross-language intelligibility (or "intercomprehension") within the Romance language family (French, Italian, Portuguese, Spanish, and Romanian). While traditional studies rely on small, hand-selected datasets and human-centric cloze tests, they often fail to capture the full scale of linguistic diversity or the asymmetry of intelligibility (where Speaker A understands Speaker B better than vice versa).

The authors identify three primary linguistic barriers to intelligibility:

Surface Distance: Differences in how words are written (orthography) and spoken (phonetics).
Semantic Divergence: "False friends" or cognates that have evolved to have different meanings.
Contextual Usage: The difficulty of recognizing related words when they appear in natural, flowing discourse rather than isolated lists.

2. Methodology

The researchers propose a novel computational framework to estimate intelligibility by combining lexical, phonetic, and semantic layers.

A. Data Sources:

RoBoCoP Database: A comprehensive database of cognates and borrowings between Romance languages.
Parallel Corpora: Two distinct datasets were used to ensure results weren't domain-specific: RomCro (literary texts) and EuroParl (political/economic proceedings).

B. The Lexical Intelligibility Index ($DLI$):
The core innovation is a mathematical formula that integrates two types of similarity:

Formal Lexical Similarity ( $S_L$ ): Measured using normalized Levenshtein distance applied to both orthographic strings and phonetic representations (generated via the eSpeak library).
Semantic Similarity ( $S_S$ ): Measured using cosine similarity of word embeddings. The study compares static embeddings (FastText trained on Wikipedia) with contextual embeddings (Sentence-BERT) to capture how words function in specific sentences.

The $DLI$ is calculated as a weighted combination of $S_S$ and $S_L$ , ensuring that if words are identical, the score reflects meaning, and if meanings are identical, the score reflects form.

C. Aggregation:
Intelligibility is calculated at the sentence level (weighted average of $DLI$ for related words) and then aggregated to a corpus level to represent the overall relationship between two languages.

3. Key Contributions

Novel Metric: The introduction of the $DLI$ index, which moves beyond simple "cognate counting" by incorporating phonetic and semantic nuances.
Multi-Layered Analysis: The first systematic study to simultaneously account for orthography, phonetics, and semantics in the Romance family.
Large-Scale Contextual Approach: Moving from static word lists (like Swadesh lists) to analyzing related words within authentic, parallel sentence structures.
Addressing Asymmetry: Providing a quantitative method to map the non-reciprocal nature of language comprehension.

4. Results

Confirmation of Asymmetry: The results quantitatively confirm that intelligibility is rarely mutual. For example, Spanish is understood by Portuguese speakers more easily than Portuguese is understood by Spanish speakers.
The "Romanian Gap": Romanian shows the highest levels of asymmetry. Romanian speakers can understand other Romance languages relatively well, but Romanian is significantly harder for others to understand. This is attributed to:
- Semantic Divergence: Words like lume (world) diverging from Latin lumen (light).
- Slavic Influence: A ~10% Slavic lexical component in Romanian.
- Phonetic Divergence: Romanian is highly phonetic, leading to a smaller gap between orthographic and phonetic scores compared to French.
Phonetic vs. Orthographic: Intelligibility scores are consistently lower for phonetic representations than orthographic ones, suggesting that in literate societies, "sight" (reading) facilitates recognition more than "hearing" (listening).
Validation: The $DLI$ scores showed a significant Spearman correlation (0.71) with human cloze test results, validating the metric's ability to mimic human comprehension.

5. Significance

This research provides a scalable, automated benchmark for studying language evolution and contact. By demonstrating that computational tools can accurately reflect human linguistic intuition, the authors offer a method to study language families where large-scale human testing is impractical. Furthermore, the study provides historical and sociolinguistic insights—such as how French influence on Romanian vocabulary affects modern intelligibility—bridging the gap between computational linguistics and historical philology.