Code-switching in text and speech challenges information-theoretic speaker design

Technical Summary: Code-switching in text and speech challenges information-theoretic speaker design

Problem Statement
Code-switching, the alternation between language varieties (e.g., Chinese and English), is a prevalent phenomenon in multilingual communication. Previous research has established a correlation between code-switching and regions of low predictability (high surprisal) in the primary language. However, the underlying motivation for this correlation remains debated. The dominant "speaker-driven" hypothesis posits that speakers switch languages to reduce cognitive load when the primary language continuation is difficult to produce; essentially, the secondary language is assumed to be more accessible or easier to produce at these points. Conversely, an "audience-driven" or "speaker design" perspective suggests that speakers might switch to signal to listeners that upcoming material is complex, thereby prompting listeners to allocate more attention. This paper investigates whether the relationship between predictability and insertional code-switching is entirely driven by speaker production ease or if it involves other mechanisms, such as signaling to the audience.

Methodology
The authors employ language modeling to analyze both written and spoken Chinese–English code-switching.

• Corpora:
  • Written: An expanded version of the Calvillo et al. (2020) corpus containing 1,476 code-switched sentences and 1,476 control non-code-switched sentences from online forums. These were translated into fully Chinese and fully English versions to allow for monolingual modeling.
  • Spoken: A subset of the South East Asia Mandarin-English (SEAME) corpus, containing 6,171 code-switched utterances and 22,770 non-code-switched Chinese utterances.
• Models:
  • Written: Monolingual 5-gram Kneser–Ney smoothed language models trained on Chinese and English Wikipedia datasets. Large Language Models (LLMs) like GPT-2 were also used for validation.
  • Spoken: Monolingual 5-gram models trained on conversational speech transcripts (CallHome for English, HUB5 for Chinese).
• Metrics: The study utilizes surprisal (negative log probability) as the primary measure of predictability.
• Experimental Design:
  1. Replication: Verify that code-switched words (CS1) in the primary language (Chinese) have higher surprisal than non-code-switched words.
  2. Secondary Language Analysis: Compare the features (length, frequency, part-of-speech, and surprisal) of the code-switched English words against standard monolingual English.
  3. Cross-Linguistic Comparison: Normalize the surprisal of CS1 words in both languages relative to their respective monolingual baselines to determine if the secondary language continuation is actually easier (lower surprisal) or harder (higher surprisal) than the primary language alternative.
  4. Regression: Logistic regression models were used to control for confounding variables such as word length, part-of-speech, and position.

Key Results

1. Replication of Primary Language Effects: The study confirms that code-switches in both writing and speech occur in regions of high surprisal within the primary language (Chinese).
2. Secondary Language Complexity: Contrary to the speaker-driven hypothesis, the English words inserted into Chinese sentences are more complex than standard monolingual English. Specifically:
   • Code-switched English words are longer (in characters/phonemes) than average monolingual English words.
   • They are significantly less frequent and have higher surprisal than monolingual English words.
3. Relative Difficulty: Even after normalizing for the baseline difficulty of each language, the surprisal of the English code-switch (CS1) is significantly higher than the surprisal of the corresponding Chinese continuation would have been.
   • Writing: English CS1 surprisal is higher than Chinese CS1 surprisal relative to their monolingual baselines.
   • Speech: The same pattern holds for spoken data; switching to English results in a more surprising (information-theoretically heavier) utterance than remaining in Chinese.
4. Modality Consistency: These findings are consistent across both written and spoken modalities, despite the differences in production constraints (e.g., the ability to revise in writing).

Key Contributions

• Dual-Language Analysis: Unlike prior work that focused exclusively on the primary language, this study analyzes the properties of the secondary language (English) at the point of the switch.
• Rejection of Pure Speaker-Driven Theory: The findings demonstrate that code-switching does not always result in a "easier" production for the speaker. In fact, switching to English often increases the information load compared to staying in Chinese.
• Evidence for Audience Design: The results suggest that code-switching may serve a communicative function beyond production ease, potentially acting as a signal to listeners that the upcoming content is challenging, thereby prompting them to allocate more cognitive resources.

Significance and Claims
The authors claim that their work provides strong evidence that the "speaker-driven" account of code-switching (where switching is solely a mechanism to ease production) is insufficient to explain the phenomenon. By showing that code-switched regions are often more surprising in the secondary language than in the primary, the paper argues that speakers may be engaging in "audience design," using code-switching to manage listener attention and comprehension.

The paper modestly notes that while the results are consistent with audience-driven pressures, they do not definitively prove this mechanism, as other factors could be at play. The authors suggest that future research should explore how these patterns vary across different language pairs with varying typological distances. The study concludes that information-theoretic constraints on code-switching are not solely driven by production ease but likely involve a trade-off where speakers accept higher production complexity to aid listener comprehension.