Accurately learning from user data while ensuring quantifiable privacy\nguarantees provides an opportunity to build better Machine Learning (ML) models\nwhile maintaining user trust. Recent literature has demonstrated the\napplicability of a generalized form of Differential Privacy to provide\nguarantees over text queries. Such mechanisms add privacy preserving noise to\nvectorial representations of text in high dimension and return a text based\nprojection of the noisy vectors. However, these mechanisms are sub-optimal in\ntheir trade-off between privacy and utility. This is due to factors such as a\nfixed global sensitivity which leads to too much noise added in dense spaces\nwhile simultaneously guaranteeing protection for sensitive outliers. In this\nproposal paper, we describe some challenges in balancing the tradeoff between\nprivacy and utility for these differentially private text mechanisms. At a high\nlevel, we provide two proposals: (1) a framework called LAC which defers some\nof the noise to a privacy amplification step and (2), an additional suite of\nthree different techniques for calibrating the noise based on the local region\naround a word. Our objective in this paper is not to evaluate a single solution\nbut to further the conversation on these challenges and chart pathways for\nbuilding better mechanisms.\n