Dialogue level quality estimation is vital for optimizing data driven\ndialogue management. Current automated methods to estimate turn and dialogue\nlevel user satisfaction employ hand-crafted features and rely on complex\nannotation schemes, which reduce the generalizability of the trained models. We\npropose a novel user satisfaction estimation approach which minimizes an\nadaptive multi-task loss function in order to jointly predict turn-level\nResponse Quality labels provided by experts and explicit dialogue-level ratings\nprovided by end users. The proposed BiLSTM based deep neural net model\nautomatically weighs each turn's contribution towards the estimated\ndialogue-level rating, implicitly encodes temporal dependencies, and removes\nthe need to hand-craft features.\n On dialogues sampled from 28 Alexa domains, two dialogue systems and three\nuser groups, the joint dialogue-level satisfaction estimation model achieved up\nto an absolute 27% (0.43->0.70) and 7% (0.63->0.70) improvement in linear\ncorrelation performance over baseline deep neural net and benchmark Gradient\nboosting regression models, respectively.\n