Proceedings of the 7th Workshop on Computational Approaches to Linguistic Code-Switching

Genta Indra Winata, Sudipta Kar, Marina Zhukova, Thamar Solorio, Xi Ai, Injy Hamed, Mahardika Krisna Krisna Ihsani, Derry Tanti Wijaya, Garry KuwantoNAACL 2025

TL;DR

This paper presents the proceedings of CALCS 2025, the seventh workshop on computational approaches to linguistic code-switching. It highlights recent advancements in understanding and modeling how multilingual speakers switch between languages, addressing the challenges posed by this complex phenomenon in natural language processing.

What Problem Does This Paper Address?

Code-switching, the practice of alternating between two or more languages or dialects within a single conversation, is a widespread phenomenon among multilingual speakers. While it's a natural part of communication for many, it poses significant challenges for computational linguistics and natural language processing (NLP) systems. Traditional NLP models are often trained on monolingual data and struggle to accurately process or generate code-switched text. This workshop aims to bridge this gap by bringing together researchers to discuss new methods, datasets, and insights for tackling code-switching, fostering collaboration between speech and text researchers, and encouraging new research in this area.

Methods

The workshop proceedings showcase a variety of computational approaches to linguistic code-switching. While specific methods are detailed within individual papers, the overarching themes include:

  • Corpus Development: Creating new, naturally sourced datasets specifically designed to capture code-switching phenomena.
  • Model Fine-tuning: Adapting existing monolingual models (e.g., Automatic Speech Recognition - ASR) to handle code-switched inputs.
  • Data Quality Analysis: Investigating how the quality of synthetic code-switched data impacts downstream NLP tasks like Machine Translation (MT) and ASR.
  • Language Mixing Analysis: Studying the patterns and mechanisms of language mixing in specific language pairs and contexts.
  • Large Language Model (LLM) Safety: Exploring methods to improve the safety and robustness of LLMs when dealing with code-switched inputs, potentially through novel learning approaches.
  • Preference Prediction: Developing models to predict human preferences in generated code-switched text, aiming for more natural and contextually appropriate outputs.

Datasets

The workshop features research utilizing and introducing several datasets:

  • EuskañolDS: A newly introduced naturally sourced corpus for Basque-Spanish code-switching.
  • Yoruba-English Code-Switching Data: Used for fine-tuning ASR models.
  • Spanish-Guaraní Code-Switching Data: Used for studying language mixing patterns in Paraguay.
  • Generated Code-Switched Text: Used for evaluating models that predict human preferences.

Key Insights and Ideas

  • The importance of naturally sourced data: While synthetic data can be useful, naturally occurring code-switched data is crucial for building robust models that reflect real-world language use.
  • Task-dependent impact of synthetic data quality: The effectiveness of synthetic code-switched data varies significantly depending on the specific NLP task (e.g., MT vs. ASR).
  • Fine-tuning for code-switching is effective: Adapting monolingual models through fine-tuning can significantly improve their performance on code-switched inputs, even for challenging tasks like ASR.
  • LLMs need specific attention for code-switching safety: Standard LLMs may exhibit vulnerabilities or biases when encountering code-switched text, necessitating targeted research into safety mechanisms.
  • Understanding "how" and "where" languages mix is key: Detailed linguistic analysis of code-switching patterns can inform the development of more accurate computational models.

Key Results

  • The EuskañolDS corpus provides a valuable resource for Basque-Spanish code-switching research.
  • Fine-tuning monolingual ASR models on Yoruba-English code-switched data shows promising improvements in performance.
  • Research on Spanish-Guaraní code-switching offers insights into the linguistic strategies employed by speakers.
  • Studies on LLM safety highlight the need for new learning paradigms to handle code-switched inputs effectively.
  • Models for predicting preferences in generated code-switched text aim to improve the naturalness and acceptability of AI-generated multilingual content.

Limitations and Future Work

  • Data scarcity: A persistent challenge in code-switching research is the limited availability of large, diverse, and high-quality datasets across many language pairs. Future work will likely focus on developing more comprehensive and accessible corpora.
  • Generalizability: Models trained on specific language pairs or contexts may not generalize well to others. Further research is needed to develop more universally applicable code-switching models.
  • LLM robustness: While progress is being made, ensuring LLMs are consistently robust and unbiased across all forms of code-switching remains an open area for future exploration.
  • Integration of speech and text: The workshop emphasizes collaboration between speech and text researchers. Future work will likely involve deeper integration of these modalities for more holistic code-switching understanding.
  • Ethical considerations: As NLP systems become more involved in multilingual communication, addressing the ethical implications of code-switching modeling, such as potential biases and cultural representation, will be crucial.

Summary updated

BibTeX
@inproceedings{winata2025proceedings,
  author = {Genta Indra Winata and Sudipta Kar and Marina Zhukova and Thamar Solorio and Xi Ai and Injy Hamed and Mahardika Krisna Krisna Ihsani and Derry Tanti Wijaya and Garry Kuwanto},
  title = {{Proceedings of the 7th Workshop on Computational Approaches to Linguistic Code-Switching}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure