We introduce a realistic expert recommendation problem called advocate recommendation. To facilitate investigation of the problem, we develop a rich dataset of \(\sim\) 25k documents called the Automatic Advocate Recommendation Dataset in the Indian Legal System (ARDI), which also contains additional attributes. Extra information about areas is generated through an expert annotation process that we incorporate into our experimentation. Treating the problem as a multi-label classification task and carrying out extensive experimentation with various strategies, including using area-based representations, summarization, ensembling methods and multi-task learning, we find the advocate recommendation task quite challenging. Our results suggest that using longer contexts and combining information from different models are beneficial in reaching a viable solution.
