Neural conversational models require substantial amounts of dialogue data for\ntheir parameter estimation and are therefore usually learned on large corpora\nsuch as chat forums or movie subtitles. These corpora are, however, often\nchallenging to work with, notably due to their frequent lack of turn\nsegmentation and the presence of multiple references external to the dialogue\nitself. This paper shows that these challenges can be mitigated by adding a\nweighting model into the architecture. The weighting model, which is itself\nestimated from dialogue data, associates each training example to a numerical\nweight that reflects its intrinsic quality for dialogue modelling. At training\ntime, these sample weights are included into the empirical loss to be\nminimised. Evaluation results on retrieval-based models trained on movie and TV\nsubtitles demonstrate that the inclusion of such a weighting model improves the\nmodel performance on unsupervised metrics.\n