Exemplar-based generative models for open-domain conversation produce\nresponses based on the exemplars provided by the retriever, taking advantage of\ngenerative models and retrieval models. However, they often ignore the\nretrieved exemplars while generating responses or produce responses over-fitted\nto the retrieved exemplars. In this paper, we argue that these drawbacks are\nderived from the one-to-many problem of the open-domain conversation. When the\nretrieved exemplar is relevant to the given context yet significantly different\nfrom the gold response, the exemplar-based generative models are trained to\nignore the exemplar since the exemplar is not helpful for generating the gold\nresponse. On the other hand, when the retrieved exemplar is lexically similar\nto the gold response, the generative models are trained to rely on the exemplar\nhighly. Therefore, we propose a training method selecting exemplars that are\nsemantically relevant to the gold response but lexically distanced from the\ngold response to mitigate the above disadvantages. In the training phase, our\nproposed training method first uses the gold response instead of dialogue\ncontext as a query to select exemplars that are semantically relevant to the\ngold response. And then, it eliminates the exemplars that lexically resemble\nthe gold responses to alleviate the dependency of the generative models on that\nexemplars. The remaining exemplars could be irrelevant to the given context\nsince they are searched depending on the gold response. Thus, our proposed\ntraining method further utilizes the relevance scores between the given context\nand the exemplars to penalize the irrelevant exemplars. Extensive experiments\ndemonstrate that our proposed training method alleviates the drawbacks of the\nexisting exemplar-based generative models and significantly improves the\nperformance in terms of appropriateness and informativeness.\n