Introduction
When writing in a second language, speakers often exhibit "interference" from their native tongue. While this is commonly observed in phonology (accents) or syntax (grammatical errors), this research explores a more subtle lexical layer: the subconscious preference for words whose etymological roots mirror those of the speaker's native language. For instance, a French speaker might prefer "flower" (Latinate) while a German speaker prefers "bloom" (Germanic).
Methodology
The researchers analyzed essays from the International Corpus of Learner English (ICLE), focusing on 11 Indo-European languages. They mapped English words to their historical ancestors using an etymological dictionary derived from Wiktionary. Each essay was then represented as a vector of proportions based on these ancestral roots (e.g., percentage of words derived from Latin vs. Old Norse).
To test the strength of this "etymological fingerprint," the team employed two main approaches:
- Classification: Using Support Vector Machines (SVM) to predict a writer's native language based on their etymological word choices.
- Clustering: Applying the UPGMA algorithm to reconstruct the Indo-European language family tree based on the distances between these fingerprints.
Results: Reconstructing the Language Tree
The most compelling finding was the ability to reconstruct historical language relationships from modern English essays. The generated tree closely mirrored the established "gold standard" for Indo-European languages, correctly identifying the Romance, Slavic, and Germanic branches.
One notable deviation was the placement of Czech within the Germanic branch rather than the Slavic branch. The authors suggest this reflects the deep historical influence of German on the Czech lexicon, which manifests as a preference for Germanic-rooted English words among Czech speakers.
Results: Native Language Identification
Quantitatively, etymological features proved to be a significant signal for identifying a writer's native language. While not as strong as traditional features like Part-of-Speech (POS) n-grams on their own, they provided complementary information that improved overall model performance.
As shown in the results table, the etymology-only model achieved 31.7% accuracy—well above the 10% random baseline. Combining etymological features with standard patterns pushed the overall accuracy from 54.2% to 56.3%.
Conclusion
This study demonstrates that our choice of vocabulary in a second language is deeply tied to the history of our first. These "etymological fingerprints" are robust enough to not only identify a speaker's background but also to trace the evolutionary history of languages themselves. Future work could expand this analysis to non-Indo-European language families or apply these insights to develop more personalized language learning tools.
Summary updated