When Elote, Choclo and Mazorca are not the Same. Isomorphism-Based Perspective to the Spanish Varieties Divergences

Cristina España-Bonet, Ankur Bhatt, Koel Dutta Chowdhury, Alberto Barrón‐Cedeño2024 · DOI

TL;DR

This paper explores the linguistic diversity within Spanish by analyzing word embeddings from 24 countries. They introduce a data-driven approach using "isomorphism metrics" to quantify the subtle differences between Spanish varieties, revealing that grammatical features like "voseo" have a stronger impact on word embeddings than lexical differences.

What Problem Does This Paper Address?

Spanish is spoken in numerous countries, each with its own unique linguistic variations shaped by historical contact with indigenous languages, immigration, and cultural differences. While these varieties are distinct, they are much more similar to each other than to entirely different languages. This poses a challenge for Natural Language Processing (NLP) techniques, which are often designed for inter-language comparisons. Standard multilingual NLP methods might not be sensitive enough to capture these fine-grained divergences within a single language. The authors aim to address this by developing a method to quantitatively measure and visualize these differences, specifically investigating how well existing "isomorphism metrics" (used for comparing different languages) perform when applied to closely related language varieties.

Methods

The core of the paper involves generating word embeddings for Spanish from 24 different countries using the CEREAL corpus. They then employ three "isomorphism metrics" to quantify the similarity (or divergence) between these embedding spaces:

  • Relational Similarity (RS): Measures the correlation of cosine similarities between word pairs in different embedding spaces.
  • Eigenvector Similarity (EV): Compares the spectral properties (eigenvalues of the Laplacian matrix) of graphs representing the embedding spaces.
  • Gromov-Hausdorff (GH) distance: A metric that measures the distance between two metric spaces, essentially finding the "worst-case" distance between nearest neighbors after an optimal transformation.

To ensure the reliability of these metrics for comparing similar varieties, they first perform a calibration study using Spanish from Spain embeddings under controlled conditions (model and data perturbation). This helps determine the best metric and configuration (e.g., number and type of words) for their main analysis. Finally, they use hierarchical clustering on the divergence scores to build "relational trees" that visualize the relationships between the Spanish varieties.

Datasets

  • CEREAL Corpus: A large, multilingual corpus annotated with country of origin, used for training the Spanish word embeddings.
  • Twitter Embeddings: Publicly available pre-trained embeddings for 22 Spanish varieties, used for comparison.
  • CoNLL-2002 Shared Task: Provided a list of Named Entities (NEs) used in one of the word selection criteria.

Key Insights and Ideas

  • Subtle Divergences: Even within a single language like Spanish, there are measurable differences in word embeddings across countries, reflecting linguistic variations.
  • Grammar Over Lexicon: Grammatical phenomena, particularly "voseo" (the use of "vos" instead of "tú" and its associated verb conjugations), appear to have a stronger and more consistent imprint on word embeddings than lexical differences (e.g., different words for "corn").
  • Metric Sensitivity: The choice of isomorphism metric and the selection of words (e.g., most frequent words vs. random words) significantly impact the results. EV (Eigenvector Similarity) with Most Frequent Words (MFW) emerged as a robust choice.
  • Data Size Matters: The amount of training data significantly influences the performance of metrics like GH, with smaller datasets leading to less reliable results.
  • Geographical vs. Linguistic Clusters: The relational trees derived from embeddings do not always align perfectly with geographical proximity or established linguistic zones, suggesting that linguistic evolution is complex and not solely dictated by borders.

Key Results

  • The study successfully generated word embeddings for Spanish from 24 countries.
  • "Voseo" was identified as the phenomenon leaving the strongest imprint on the embedding spaces, leading to distinct clusters in the relational trees.
  • EV (MFW 100) was found to be the most reliable metric for measuring isomorphism in this context.
  • Relational trees showed that varieties with similar grammatical features (like voseo) tend to cluster together, even if geographically distant.
  • Differences in embedding spaces were observed to be subtle compared to those between different languages, confirming the challenge for standard multilingual NLP.

Limitations and Future Work

  • Data Size Imbalance: The study acknowledges that the amount of training data varies significantly across varieties, which can affect the quality of embeddings and the reliability of metrics. Varieties with less than 15k training sentences were excluded from high-resolution experiments.
  • Domain Differences: The use of CEREAL (web data) and Twitter data highlights potential differences in register and data quality, which can influence results.
  • Future Work: The authors suggest further exploration into how different types of linguistic variation (phonetic, syntactic) manifest in embeddings and how to better account for data size differences in future analyses. They also propose extending this approach to other languages with significant internal variation.

Summary updated

Abstract

Spanish is an official language in 20 countries; in 19 of them, it arrived by means of overseas colonisation. Its close contact with several coexistent languages and the rich regional and cultural diversity has produced varieties which divert from each other. We study these divergences in a data-based approach and according to their qualitative and quantitative effects in word embeddings. We generate embeddings for Spanish in 24 countries and examine the topology of the spaces. Due to the similarities between varieties {---}in contrast to what happens to different languages in bilingual topological studies{---} we first scrutinise the behaviour of three isomorphism measures in (quasi-)isomorphic settings: relational similarity, Eigenvalue similarity and Gromov-Hausdorff distance. We then use the most trustworthy measure to quantify the divergences among varieties. Finally, we use the departures from isomorphism to build relational trees for the Spanish varieties by hierarchical clustering.
BibTeX
@misc{espanabonet2024when,
  author = {Cristina España-Bonet and Ankur Bhatt and Koel Dutta Chowdhury and Alberto Barrón‐Cedeño},
  title = {{When Elote, Choclo and Mazorca are not the Same. Isomorphism-Based Perspective to the Spanish Varieties Divergences}},
  year = {2024},
  doi = {10.18653/v1/2024.vardial-1.5},
}
Figure