Movatterモバイル変換

Style Variation as a Vantage Point for Code-Switching

Khyathi Raghavi Chandu, Alan W. Black

Code-Switching (CS) is a prevalent phenomenon observed in bilingualand multilingual communities, especially in digital and social mediaplatforms. A major problem in this domain is the dearth of substantialcorpora to train large scale neural models. Generating vast amountsof quality synthetic text assists several downstream tasks that heavilyrely on language modeling such as speech recognition, text-to-speechsynthesis etc,. We present a novel vantage point of CS to be stylevariations between both the participating languages. Our approach doesnot need any external dense annotations such as lexical language ids.It relies on easily obtainable monolingual corpora without any parallelalignment and a limited set of naturally CS sentences. We propose atwo-stage generative adversarial training approach where the firststage generates competitive negative examples for CS and the secondstage generates more realistic CS sentences. We present our experimentson the following pairs of languages: Spanish-English, Mandarin-English,Hindi-English and Arabic-French. We show that the trends in metricsfor generated CS move closer to real CS data in the above languagepairs through the dual stage training process. We believe this viewpointof CS as style variations opens new perspectives for modeling varioustasks in CS text.