The limits of Spanglish?

Barbara Bullock, Wally Guzmán, Almeida Jacqueline ToribioNAACL 2019

TL;DR

This paper investigates the linguistic boundaries of "Spanglish" by analyzing code-switching patterns in literary texts and comparing them to natural speech. The authors introduce a novel metric to normalize code-switching complexity across corpora with varying language distributions, revealing that author Ilan Stavans exhibits unusually complex and potentially "unnatural" switching patterns compared to other literary and conversational examples.

What Problem Does This Paper Address?

Code-switching (C-S), the practice of alternating between two or more languages within a single conversation, is a common phenomenon in bilingual communities. When authors incorporate C-S into literature, it becomes an artistic choice that can reflect real-world bilingual interactions. However, there's a question of where the "limits" of this literary C-S lie. Some critics view certain "Spanglish" texts, particularly those by Ilan Stavans, as deviating from established norms of C-S observed in natural speech. This paper aims to quantify these differences and understand what makes Stavans's writing distinct, potentially pushing the boundaries of what is considered typical or acceptable literary code-switching. This is particularly relevant for natural language processing (NLP) tasks, as C-S complicates standard language identification, POS tagging, and language modeling.

Methods

The researchers employ a multi-faceted approach to model and compare C-S patterns:

  • Language Identification: They use a word-level language identification system for Spanish-English to tag each word in the corpora.
  • Metric Development: A key contribution is the introduction of a Normalized I-Index (I²). This metric addresses a limitation of previous I-Indices by accounting for the underlying language distribution of a corpus. This allows for more accurate and comparable measurements of switching complexity across different texts, even if they have vastly different proportions of each language.
  • Feature Extraction: They extract linguistic features, specifically focusing on:
  • Out-of-vocabulary (OOV) items: Words that are not standard in either language.
  • Syntactic Transitions: Analyzing switches at specific grammatical junctures, such as between determiners and nouns (DET-N) and between pronouns and verbs (PRON-V).
  • Manual Review: Extracted features are manually reviewed for grammatical analysis to ensure accuracy and distinguish genuine C-S from errors.
  • Quantitative Analysis: They apply established metrics like the M-Index (language balance), I-Index (switching probability), and Burstiness (span of language use before a switch) alongside their new I² metric.

Datasets

The study utilizes a combination of literary and conversational datasets:

  • Literary Corpora:
    • Excerpts from Ilan Stavans's works (totaling 10,051 words).
  • Killer Crónicas: Bilingual Memories by Susana Chávez-Silverman.
  • Yo-Yo Boing! by Giannina Braschi.
  • Conversational Corpora:
    • A Spanish-English transcription of a bilingual conversation in Texas (S7).
  • A conversation from the Miami Corpus (M40).

These datasets represent both naturally occurring C-S and literary adaptations, allowing for a comparative analysis.

Key Insights and Ideas

  • Literary C-S is More Complex: Literary texts generally exhibit more complex C-S patterns than natural speech, suggesting authors leverage C-S for artistic effect beyond everyday bilingual interactions.
  • Stavans Pushes Boundaries: Ilan Stavans's writing stands out for its unusually high rate and complexity of code-switching, including word-internal switches and switches at tight syntactic boundaries (like DET-N and PRON-V).
  • The Need for Normalized Metrics: The development of the Normalized I-Index (I²) is crucial because it allows for fair comparisons of C-S complexity across corpora with different language proportions, a limitation of previous metrics.
  • Qualitative Differences in Switching: Beyond just the amount of switching, the types of switches differ. Stavans's work shows a higher frequency of switches at junctures that are typically avoided in natural speech and are often ruled out by linguistic constraints.
  • "Unnatural" Switching: The specific patterns observed in Stavans's work (e.g., word-internal switches, switches after determiners and pronouns) are precisely what lead some critics to label his "Spanglish" as "unnatural" or "grotesque."

Key Results

  • M-Index: Literary corpora (Stavans, Killer Crónicas, Yo-Yo Boing!) show a more balanced language distribution (higher M-Index) compared to conversational corpora (M40, S7), where one language often predominates.
  • I² Metric: The Normalized I-Index (I²) reveals that Stavans's excerpts exhibit extreme switching relative to other datasets, even more so than Killer Crónicas, which has a very high M-Index.
  • Burstiness: Stavans and Killer Crónicas show more regular mixing throughout the text (less "bursty"), while Yo-Yo Boing!, M40, and S7 display more sporadic C-S.
  • Word-Internal C-S: Stavans shows a significantly higher frequency of word-internal C-S (0.635%) compared to Killer Crónicas (0.009%).
  • DET-N Switching: Stavans exhibits a much higher mixing rate at the Determiner-Noun juncture (.36) compared to Killer Crónicas (.17).
  • PRON-V Switching: Stavans switches after a subject pronoun at a rate of about 13%, particularly if the pronoun is Spanish, a rate significantly higher than in Killer Crónicas.

Limitations and Future Work

  • Tentative Conclusions: The authors acknowledge that their conclusions are tentative due to the reliance on small corpora and the complexity of the phenomenon.
  • Defining "Outlier" C-S: A key future direction is to determine expected constants of C-S for "Spanglish" literature versus natural speech. This will help establish objective criteria for identifying when a C-S contour is an outlier.
  • Language Models for Comparison: The authors suggest that developing language models capable of direct statistical distribution comparisons across corpora would be beneficial for establishing the limits of mixed vernaculars like "Spanglish."
  • Understanding Criticisms: While the study quantifies the differences, it notes that it's still unclear whether the frequency of switching, the decision to switch at specific functional elements and within words, or a combination of these factors is primarily responsible for critical reactions to Stavans's work.

Summary updated

BibTeX
@inproceedings{bullock2019limits,
  author = {Barbara Bullock and Wally Guzmán and Almeida Jacqueline Toribio},
  title = {{The limits of Spanglish?}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
}
Figure