LuxBorrow: From Pompier to Pompjee, Tracing Borrowing in Luxembourgish

This paper introduces LuxBorrow, a comprehensive corpus-based analysis of Luxembourgish news from 1999 to 2025 that reveals pervasive multilingual mixing dominated by French loanwords and morphological adaptations, advocating for a shift from document-level code-mixing metrics to token-level borrowing evaluation.

The Gist

Imagine Luxembourgish (the local language of Luxembourg) as a giant, cozy house. For decades, this house has been the main home for everyone. But, because Luxembourg is a tiny country surrounded by big neighbors (Germany, France, and increasingly, the English-speaking world), people constantly bring in furniture, decorations, and tools from those neighbors to make life easier.

This paper, "LuxBorrow," is like a 27-year-long architectural survey of that house. The researchers looked at 259,000 news articles (from 1999 to 2025) to answer a simple question: How much of the "foreign furniture" has actually become part of the house's permanent structure, and how much is just temporary clutter?

Here is the breakdown of their findings, using some everyday analogies:

1. The House is Still the Boss (The "Matrix Language")

Even though the news articles are full of words from French, German, and English, the researchers found that Luxembourgish is still the boss.

• The Analogy: Imagine a dinner party where the host speaks Luxembourgish, but they keep slipping in French words for "wine," German words for "sausage," and English words for "internet." The conversation is still clearly a Luxembourgish dinner party, not a French or German one.
• The Finding: 100% of the articles are fundamentally Luxembourgish. Even the most "mixed" articles are just Luxembourgish with a few foreign sprinkles, not a balanced mix of two languages.

2. The "Borrowing" vs. The "Code-Switching"

The researchers made a crucial distinction between Borrowing and Code-Switching.

• Borrowing (The Renovation): This is when you take a foreign word and remodel it to fit your house. You paint it, change its shape, and make it look like it belongs.
  • Example: Taking the French word pompier (firefighter) and turning it into the Luxembourgish Pompjee. It's now a native resident.
• Code-Switching (The Tourist): This is when you suddenly speak a different language for a sentence or two, like a tourist visiting a museum. It hasn't been remodeled; it's just visiting.
  • Example: Suddenly saying a whole sentence in French in the middle of a Luxembourgish story.
• The Finding: Most of the "foreign stuff" in the news is just tourists (code-switching). However, the researchers focused on the renovations (borrowings) because those are the words that actually change the language forever.

3. The "Remodeling" Rules (How Words Change)

When Luxembourgish "renovates" a foreign word, it follows strict rules, mostly changing the spelling or the grammar endings.

• The Analogy: Think of it like a strict interior designer who refuses to let a French chair stay in its original French shape. They must sand it down and paint it Luxembourgish.
• The Rules:
  • Spelling Changes (The Paint Job): The most common change is turning French sounds into Luxembourgish spelling. For example, changing on to oun or eur to er.
  • Grammar Changes (The Legs): They often add Luxembourgish endings to foreign verbs. For example, turning the French ajuster into ajustéieren.
  • The Result: Out of all the borrowed words they found, 64% were "renovated" with grammar changes, and 36% were just "painted" (spelling changes). Very few (less than 1%) were left completely untouched.

4. The Timeline: A Slow, Steady Influx

The researchers looked at the data over 27 years.

• The Trend: The house is getting slightly more crowded with foreign items every year. The "mixing" has increased by about 37% since 1999.
• The Source:
  • France is the main supplier of "furniture." Almost all the borrowed words come from French.
  • Germany is a distant second, slowly bringing in a few more items.
  • English is barely showing up in the "renovated" section. Even though English is huge globally, in Luxembourgish news, it hasn't yet become a permanent part of the house's structure in the same way French has.

5. Why Does This Matter? (The "Robot" Problem)

Why spend 27 years counting words?

• The Problem: Computers (AI, translation tools, speech recognition) are terrible at understanding this "mixed" house. If a computer sees Pompjee, it might not realize it's a borrowed word for "firefighter" because it looks like a made-up word.
• The Solution: This paper provides a rulebook for computers. By understanding how Luxembourgish remodels foreign words (e.g., "Oh, if it ends in -éiert, it's probably a French verb"), we can teach computers to understand the language better.

Summary

The paper tells us that Luxembourgish is a resilient host. It welcomes guests from France, Germany, and England, but it insists that they change their clothes (spelling and grammar) before they can stay in the living room. While the house is getting a bit more colorful over time, the original structure remains strong, and the "renovations" are happening mostly with French-style furniture.

In short: Luxembourgish news is a multilingual melting pot, but the pot itself is still very much Luxembourgish.

Technical Summary

Here is a detailed technical summary of the paper "LuxBorrow: From Pompier to Pompjee, Tracing Borrowing in Luxembourgish."

1. Problem Statement

Luxembourgish (LU) is a multilingual language in constant contact with German (DE), French (FR), and English (EN). While lexical borrowing (the integration of foreign words into the recipient language) and code-switching (spontaneous alternation between languages) are pervasive in Luxembourgish media, they remain insufficiently studied at scale.

• The Gap: Existing Natural Language Processing (NLP) resources and benchmarks often treat code-switching and borrowing as a monolithic "mixing" phenomenon or focus on high-resource language pairs (e.g., Spanish-English). There is no large-scale, borrowing-labeled corpus for Luxembourgish.
• The Challenge: Distinguishing between entrenched borrowings (lexically listed, morphologically adapted, and integrated into LU grammar) and spontaneous code-switches (unassimilated, syntactically independent insertions) is difficult, especially between closely related languages like LU and DE. Current sentence-level Language Identification (LID) tools often fail to capture token-level nuances or misclassify adapted loanwords as native LU.

2. Methodology

The authors present LuxBorrow, a pipeline designed to analyze 27 years of Luxembourgish news (1999–2025) from RTL Luxembourg, comprising 259,305 articles and 43.7 million tokens.

A. Data Collection & Preprocessing

• Corpus: Longitudinal dataset spanning six socio-historical periods (1999–2025), including the Financial Crisis, the 2020 Orthography Reform, and the AI era.
• Hybrid LID Pipeline:
  1. Sentence-Level Gate: Uses OpenLID (FastText-based) to identify the matrix language. A length-adaptive threshold prevents short sentences from being misclassified. Only sentences labeled as LU proceed to the next stage.
  2. Token-Level Detection: Within LU sentences, tokens are lemmatized and normalized using spellux. A custom detector matches tokens against a curated loanword registry (derived from the Lëtzebuergeschen Online Dictionnaire - LOD) and applies morphological/orthographic rules.

B. Borrowing vs. Code-Switching Distinction

The system operationalizes the "Simple View" of borrowing:

• Borrowings: Identified via morphological adaptation (e.g., French pompier $\to$ LU Pompjee), orthographic shifts (e.g., on $\to$ oun), or lexical listing in dictionaries. These are treated as integrated LU items.
• Code-Switches: Identified as longer contiguous spans of foreign language tokens with donor-language syntax, lacking LU morphological integration.
• Disambiguation Logic: The pipeline uses contextual features (foreign run-length, local LU token density) and external knowledge (Wiktionary) to resolve:
  • Parallel Borrowing: Determining if a word is a direct loan or a shared inheritance (e.g., distinguishing between a direct French loan vs. a word derived from a common Latin root).
  • LU/DE Proximity: Reclassifying words as non-loans if they can be traced to a common Old High German root rather than a modern German loan.

C. Statistical Metrics

The study moves beyond simple mixing indices to compute:

• Borrowed Token/Type Rates: Proportion of borrowed items.
• Donor Entropy: Diversity of source languages among borrowed items.
• Assimilation Ratio: Share of borrowed tokens showing morphological/orthographic integration.
• Code-Mixing Index (CMI), Shannon Entropy, and M-Index: Calculated at document and section levels to measure mixing intensity and diversity.

3. Key Contributions

1. LuxBorrow Corpus & Registry: The first token-level, borrowing-centric corpus for Luxembourgish, containing a curated registry of ~7,400 loanword entries with donor tags (FR, DE, EN) and adaptation patterns.
2. Novel Detection Pipeline: A hybrid approach combining sentence-level LID with rule-based, lexicon-guided token resolution, specifically designed to separate integrated borrowings from code-switches in low-resource, high-contact settings.
3. Diachronic Analysis: A 27-year longitudinal study tracking the evolution of borrowing and code-switching, revealing trends in integration and donor language dominance.
4. Methodological Framework: Proposes "borrowing-first" diagnostics (assimilation ratios, donor entropy) as more actionable metrics for NLP tasks than headline document-level mixing indices.

4. Key Results

A. Document-Level Mixing (RQ2)

• Matrix Language Stability: LU remains the dominant matrix language in 100% of documents.
• Prevalence vs. Intensity: While 77.1% of articles contain at least one donor language and 65.4% use three or four languages, the mixing intensity is low.
  • Median Code-Mixing Index (CMI) rises only modestly with language breadth: 3.90 (LU+1) $\to$ 7.00 (LU+3).
  • This indicates localized insertions rather than balanced bilingual text.
• Donor Dominance: French (FR) is the primary donor (91.1% of multilingual documents), followed by German (DE) and English (EN).

B. Borrowing Adaptation Patterns (RQ3)

• Total Instances: 25,444 identified borrowing instances.
• Adaptation Profile:
  • Morphological: 63.8% (Collectively dominant, e.g., er $\to$ éieren).
  • Orthographic: 35.9% (Highest individual counts, e.g., on $\to$ oun, eur $\to$ er).
  • Lexical (Unadapted): 0.3% (Rare).
• Top Rules: The most frequent single rules are orthographic (on $\to$ oun, eur $\to$ er), but morphological rules drive the collective volume.

C. Diachronic Trends (RQ4)

• Intensification: Code-switching rates have increased by 37.4% from 1999–2004 to 2020–2025 (from ~5.76% to ~7.92%).
• Borrowing Growth: While the share of morphologically adapted borrowings remains small (~0.73%), the absolute volume has grown 13.2x (from 483 to 6,398 instances) over the same period.
• Source Shift: French overwhelmingly supplies adapted items (~97–99%). German shows modest growth (0.8% $\to$ 2.7%), while English remains negligible in morphological adaptation.

5. Significance and Implications

• NLP Robustness: The study highlights that standard LID fails to distinguish integrated loanwords from native LU, leading to errors in downstream tasks (ASR, MT, NER). A borrowing-aware pipeline is essential for robust processing.
• Linguistic Theory: The findings support the "Simple View" of borrowing, showing that in edited news, the matrix language (LU) retains structural control, with foreign elements appearing as localized, often adapted, insertions rather than full code-switching.
• Resource Development: The release of the loanword registry and adaptation patterns provides a foundation for training sequence taggers and normalizing Luxembourgish text, addressing the scarcity of resources for low-resource languages.
• Future Directions: The authors advocate for "borrowing-first" evaluation metrics and plan to expand the corpus to include conversational and social media data to capture rarer, non-standardized borrowing patterns.

Conclusion

LuxBorrow demonstrates that while Luxembourgish media is highly multilingual in terms of breadth (many languages present), it is low-intensity in terms of mixing (LU remains dominant). The study successfully quantifies the "borrowing" component of this mix, revealing a strong trend toward French-sourced morphological adaptation and providing a critical toolkit for improving NLP systems in multilingual, low-resource environments.