Emotion classification is a challenging task in NLP due to the inherent idiosyncratic and subjective nature of linguistic expression,especially with code-mixed data. Pre-trained language models (PLMs) have achieved high performance for many tasks and languages, but it remains to be seen whether these models learn and are robust to the differences in emotional expression across languages. Sociolinguistic studies have shown that Hinglish speakers switch to Hindi when expressing negative emotions and to English when expressing positive emotions. To understand if language models can learn these associations, we study the effect of language on emotion prediction across 3 PLMs on a Hinglish emotion classification dataset. Using LIME and token level language ID, we find that models do learn these associations between language choice and emotional expression. Moreover, having code-mixed data present in the pre-training can augment that learning when task-specific data is scarce. We also conclude from the misclassifications that the models may overgeneralise this heuristic to other infrequent examples where this sociolinguistic phenomenon does not apply.

Monolinguals make up a minority of the world’s speakers, and yet most language technologies lag behind in handling linguistic behaviours produced by bilingual and multilingual speakers. A commonly observed phenomenon in such communities is code-mixing, which is prevalent on social media, and thus requires attention in NLP research. In this work, we look into the ability of pretrained language models to handle code-mixed data, with a focus on the impact of languages present in pretraining on the downstream performance of the model as measured on the task of sentiment analysis. Ultimately, we find that the pretraining language has little effect on performance when the model sees code-mixed data during downstream finetuning. We also evaluate the models on code-mixed data in a zero-shot setting, after task-specific finetuning on a monolingual dataset. We find that this brings out differences in model performance that can be attributed to the pretraining languages. We present a thorough analysis of these findings that also looks at model performance based on the composition of participating languages in the code-mixed datasets.