Low-Shot Cross-Lingual Knowledge Transfer

The goal of the project is to analyze and compare methodologies for transferring supervision across languages, without (or with minimal) parallel data. The primary focus is around the tasks of Sentiment Analysis, Natural Language Inference, and Bilingual Lexicon Induction, and the main languages under study are English, Hindi, Dutch, and Russian. As part of the project, we also seek to answer important research questions like, which languages are easier to transfer to/from, are certain tasks more dependent on the language than others, can two completely unrelated tasks be used as a source of cross-lingual supervision, etc.