Towards Unsupervised Recognition of Dialogue Acts