Unsupervised Synthetic Code-Mixed Data Generation

Code-mixed text is common among non-English users. Code-mixed parallel training data fosters the research of code-mixed text. We propose a phrase-based translation approach to generate a synthetic code-mixed corpus. In this approach, we first parse the sentences of both source and target languages, the parsed output contains the sentences with POS tags given to the words, and sentences divided into phrases with head words identified. Next, we extract head words from the parsed output and determine their alignment to replace them in the target language, producing code-mixed sentences. We later use this synthetically generated code-mixed data to train an NMT to perform machine translation. The proposed methodology is independent of the languages used for the code-mixing; Hindi and English languages are used for the demonstration and evaluation. We compare the proposed approach with an alignment-based approach and a pre-trained model approach. Results have shown that our approach has yielded comparable BLEU scores to both the alignment-based approach and the pre-trained model approach, while effectively resolving the issues associated with both methods.