The general approaches for polarity analysis in dialogue, e.g. Multiple Instance Learning (MIL), have achieved significant progress.
However, one significant drawback of current approaches is that the contribution of an utterance towards the polarity being a \emph{black-box}.
For existing methods, the polarity contained in each utterance, which we call meta-polarity, is not explicitly utilized.
In this paper, we study the problem of adding interpretability to the overall polarity by predicting the meta-polarity at the same time.
First, we reformulate a large advising dataset \cite{dstc20task2}, where the meta-polarity of each utterance is given.
Second, we propose an utterance classification model (UCM) and a two-stage progressive training method that strengthens the connection between the meta-polarity and the overall polarity.
Experimental results show that our overall approach outperforms all competitive baselines by substantial margins, achieving a new state-of-the-art performance on this dataset.