Word Predictability on Code-switching Points in Cantonese–English Discourse

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, short

Summary

A study published in the Proceedings of the Society for Computation in Linguistics 2026 investigates how word predictability affects code-switching probability in Cantonese–English discourse. Researchers analyzed 1,010 code-switched instances from naturalistic interviews with 41 bilinguals across homeland, immersed, and heritage groups. They operationalized predictability as surprisal, estimated using pretrained transformer-based language models, including autoregressive and masked types, with varying contextual information. Mixed-effects logistic regression analyses consistently showed that less predictable words are more likely to be code-switched. These findings, presented in San Diego, CA, highlight predictability's significant role in bilingual speech production across diverse language experiences.

Key takeaway

For research scientists modeling bilingual language use, this study indicates that word predictability is a robust factor in predicting code-switching events. You should consider incorporating surprisal metrics from transformer-based models into your analyses to better understand and predict bilingual speech patterns. This approach offers new insights into the cognitive processes underlying language alternation across diverse speaker populations.

Key insights

Less predictable words significantly increase the likelihood of code-switching in bilingual speech production.

Principles

• Less predictable words are more prone to code-switching.
• Predictability influences bilingual speech production.
• Effects are consistent across model types and bilingual groups.

Method

Surprisal, estimated via pretrained transformer-based language models (autoregressive and masked) with varying context, was used to quantify predictability. Mixed-effects logistic regression then analyzed its influence on code-switching.

Topics

• Code-switching
• Bilingualism
• Word Predictability
• Transformer Models
• Language Modeling
• Sociolinguistics

Best for: NLP Engineer, AI Scientist, Research Scientist

Related on AIssential

• LBC[W]: Predictive Testing of the Resource-Vector Hypothesis (LBC(w)) in Language Evolution Author… — Language evolution is predictable based on a resource-vector balance of legitimacy, cost, benefit, and contextual weight.
• Beyond Bilingual Transfer: Multilingual Code-Switching in Instruction Tuning — Multilingual code-switching instruction tuning improves LLM performance beyond bilingual settings.
• What Makes Words Hard? Sakura at BEA 2026 Shared Task on Vocabulary Difficulty Prediction — LLM fine-tuning with soft-target loss effectively predicts vocabulary difficulty while explainable models offer insights.
• Measuring the entropy of English — Shannon's entropy estimation for English revealed that true compressibility requires understanding human prediction, not just statistical analysis.
• Machine learning and digital pragmatics: Which word category influences emoji use most? — MARBERT model shows promise in predicting Arabic emoji use, but needs further refinement for dialectal nuances.
• When Better Codebooks Are Not Enough: Predictive Performance and Behavioral Reliability in LLM Political Event Coding — LLM accuracy in codebook-guided tasks does not guarantee behavioral reliability or adherence to underlying coding logic.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.