Surprisal Predicts Code-Switching in Chinese-English Bilingual Text

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4029–4039, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4029 Surprisal Predicts Code-Switching in Chinese-English Bilingual Text Jes´us Calvillo 1 jzc1104@psu.edu Le Fang 1 fredfang1203@gmail.com Jeremy Cole 2,1 jrcole@google.com David Reitter 2,1 reitter@google.com 1Pennsylvania State University 2Google Research Abstract Why do bilinguals switch languages within a sentence? The present observational study asks whether word surprisal and word en- tropy predict code-switching in bilingual writ- ten conversation. We describe and model a new dataset of Chinese-English text with 1476 clean code-switched sentences, translated back into Chinese. The model includes known control variables together with word surprisal and word entropy. We found that word sur- prisal, but not entropy, is a significant predic- tor that explains code-switching above and be- yond other well-known predictors. We also found sentence length to be a significant pre- dictor, which has been related to sentence com- plexity. We propose high cognitive effort as a reason for code-switching, as it leaves fewer resources for inhibition of the alternative lan- guage. We also corroborate previous findings, but this time using a computational model of surprisal, a new language pair, and doing so for written language. 1 Introduction Code-Switching (CS) occurs when a speaker alter- nates from one language to another during linguis- tic communication (e.g., Poplack, 1980). For ex- ample, in: “洗衣房在basement 。” (“The laundry room is in the basement.”), the speaker alternates from Chinese to English by introducing the word “basement”, replacing the Chinese word “地下室”. This behavior is very common among bilinguals. Many factors have been shown to affect the propensity of a bilingual to code-switch. Among others, there are variables related to the partici- pants in the conversation (e.g., Blom and Gumperz, 1972), the ease of production of the relevant words (e.g., Gollan and Ferreira, 2009), the linguistic con- text (e.g., Clyne, 1991), memory limitations of the speaker (Eppler, 2011), cognitive load and emo- tional state of the speaker (e.g., Grosjean, 1982; Dornic, 1978), and the type of information to be conveyed (e.g., Karrebæk, 2003; Mysl´ın and Levy, 2015). Among the latter, predictability, as mea- sured by word completion, has been correlated with code-switching (Mysl´ın and Levy, 2015). In this paper, we model predictability using word surprisal calculated with a language model. We ask whether word surprisal (Hale, 2001) and word entropy (Roark et al., 2009) affect the proba- bility of CS within a sentence (intra-sentential CS), while controlling for other known psycholinguis- tic factors. Word surprisal measures how unpre- dictable a word is in its context, typically opera- tionalized as the negative log-probability of a word wi conditioned on a window of t previous words: surp(wi) = −logP(wi|wi−1, ..., wi−t) (1) Word entropy before wi measures the uncertainty when wi is still unknown, operationalized as the expectation over the vocabulary of word surprisal: Hi−1 = � w∈vocab −logP(w|wi−1, ..., wi−t) ∗ P(w|wi−1, ..., wi−t) (2) Thus, given a context, word surprisal measures how unpredictable a specific word is, while word entropy …

Content truncated to 500 words