Neural models trained for next utterance generation in dialogue task learn to\nmimic the n-gram sequences in the training set with training objectives like\nnegative log-likelihood (NLL) or cross-entropy. Such commonly used training\nobjectives do not foster generating alternate responses to a context. But, the\neffects of minimizing an alternate training objective that fosters a model to\ngenerate alternate response and score it on semantic similarity has not been\nwell studied. We hypothesize that a language generation model can improve on\nits diversity by learning to generate alternate text during training and\nminimizing a semantic loss as an auxiliary objective. We explore this idea on\ntwo different sized data sets on the task of next utterance generation in goal\noriented dialogues. We make two observations (1) minimizing a semantic\nobjective improved diversity in responses in the smaller data set (Frames) but\nonly as-good-as minimizing the NLL in the larger data set (MultiWoZ) (2) large\nlanguage model embeddings can be more useful as a semantic loss objective than\nas initialization for token embeddings.\n