Unsupervised Dialogue Topic Segmentation with Topic-aware Utterance Representation