The similarity model is not constructed as a direct classifier of binary labels that indicate similar vs not-similar.

Instead, it is learnt as part of a the model that has booked vs not-booked as label, and the similarity model helps refine the prediction for booked vs not-booked based on the similarity to the antecedent listing.

Please see https://arxiv.org/pdf/2210.07774.pdf Section 5 for a more detailed explanation.