Detecting Unassimilated Borrowings in Spanish: An Annotated Corpus and Approaches to Modeling

Elena Álvarez Mellado, Constantine LignosACL 2022 · DOI

When a Spanish speaker says they are going to a reunión, they are using a standard Spanish word. But if they say they have a meeting, they are using a lexical borrowing—specifically an "unassimilated" one, meaning it hasn't changed its spelling to fit Spanish rules (unlike fútbol from "football").

For Natural Language Processing (NLP) systems, these words are a nightmare. They are often "out-of-vocabulary" (OOV) and can confuse parsers or translation engines. In a new paper presented at ACL 2022, researchers Elena Álvarez-Mellado and Constantine Lignos introduce a robust new dataset and modeling approach to tackle this linguistic challenge.

TL;DR

The authors present COALAS, a large-scale, high-quality annotated corpus for detecting unassimilated borrowings in Spanish. They show that while traditional models struggle with unseen words, a BiLSTM-CRF architecture combining contextualized embeddings with subword information (like characters and BPE) achieves state-of-the-art performance, significantly outperforming standard multilingual BERT models.

What Problem Does This Paper Address?

Identifying borrowings is essential for lexicographers tracking language change and for improving NLP tools. However, previous Spanish datasets were small, limited to news headlines, and—crucially—had too much overlap between training and test data. This meant models could "cheat" by memorizing specific words rather than learning the patterns of what a borrowing looks like.

The authors focus on unassimilated borrowings: words that enter a language without orthographic adaptation. These are particularly tricky because they often appear in specific contexts (tech, fashion, sports) and frequently involve new, "trendy" terms that a model might never have seen during training.

Datasets: The COALAS Corpus

The primary contribution is the COALAS (Corpus of Anglicisms in the Spanish Press) dataset. It is designed to be diverse and challenging, moving beyond simple news headlines to capture a wide array of topics.

The COALAS dataset draws from 21 media sources across various topics, ensuring a diverse representation of Spanish text.
  • Size: 370,000 tokens across 21 different Spanish media sources.
  • Diversity: Covers topics from politics and economy to "gossip," videogames, and parenthood.
  • Difficulty: The test set was intentionally designed to be "OOV-rich." 92% of the borrowings in the test set were not seen during training, forcing models to generalize rather than memorize.
  • Availability: The dataset is publicly available on GitHub.

Key Insights and Ideas

  • The "Memorization" Trap: The CRF baseline, which scored an F1 of ~86 on older, easier datasets, plummeted to an F1 of 55.44 on COALAS. This highlights how much previous benchmarks over-estimated model performance by not accounting for OOV words.
  • Subwords are the Secret Sauce: For unassimilated borrowings, the internal structure of the word is a massive clue. English-style letter combinations (like "-sh" or "-ck") are rare in Spanish. Models that "looked" at characters and subword units (BPE) were much better at spotting new borrowings.
  • Context Matters, but so does Origin: The best-performing models didn't just use Spanish context; they benefited from combining Spanish embeddings (BETO) with English BERT embeddings. This suggests that "knowing" what a word looks like in its donor language helps identify it when it "borrows" into a new one.

Key Results

The champion model was a BiLSTM-CRF using a combination of BETO + English BERT + BPE + character embeddings.

Performance of the top-performing BiLSTM-CRF models, showing the impact of combining Spanish and English BERT with subword information.
  • Best F1 Score: 84.22 on the test set.
  • Comparison: This outperformed the standard mBERT (82.02) and the monolingual BETO (80.71).
  • Recall Boost: Adding character-level information was the single most effective way to improve recall, helping the model find those 92% of borrowings it had never seen before.

Limitations and Future Work

The authors acknowledge several areas for growth, including geographic bias (focusing on European Spanish) and genre constraints (newswire focus). Future directions include exploring "character-level perplexity"—measuring how "weird" a word looks compared to standard Spanish—as a potential feature for even better detection.

Summary updated

BibTeX
@inproceedings{mellado2022detecting,
  author = {Elena Álvarez Mellado and Constantine Lignos},
  title = {{Detecting Unassimilated Borrowings in Spanish: An Annotated Corpus and Approaches to Modeling}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.acl-long.268},
}
Figure