MARRO is a novel architecture for rhetorical role classification for legal documents from the Indian and the UK Supreme Court. Rhetrocial roles are labels such as facts, arguments, precedence, etc. assigned to each sentence in a court document.

We proposed a BiLSTM-CRF model that uses self-attention inspired by transformer models to achieve state-of-the-art results. We further train the said model(s) in a multi-task setting using label shift prediction as an auxiliary task and rhetorical role prediction as a primary task. Our proposed model(s) outperform any transformer model on the task.