We propose an unsupervised representation learning method for a job title similarity model using noisy skill labels. We show that it is highly effective for tasks such as text ranking and job normalization.