From Machine Translation to Code-Switching: Generating High-Quality Code-Switched Text

Generating code-switched text is a problem of growing interest, especially\ngiven the scarcity of corpora containing large volumes of real code-switched\ntext. In this work, we adapt a state-of-the-art neural machine translation\nmodel to generate Hindi-English code-switched sentences starting from\nmonolingual Hindi sentences. We outline a carefully designed curriculum of\npretraining steps, including the use of synthetic code-switched text, that\nenable the model to generate high-quality code-switched text. Using text\ngenerated from our model as data augmentation, we show significant reductions\nin perplexity on a language modeling task, compared to using text from other\ngenerative models of CS text. We also show improvements using our text for a\ndownstream code-switched natural language inference task. Our generated text is\nfurther subjected to a rigorous evaluation using a human evaluation study and a\nrange of objective metrics, where we show performance comparable (and sometimes\neven superior) to code-switched text obtained via crowd workers who are native\nHindi speakers.\n