Approaches for estimating the similarity between individual publications are an area of long -standing interest in the scientometrics and informetrics communities. Traditional techniques have generally relied on references and other metadata, while text mining approaches based on title and abstract text have appeared more frequently in recent years. In principle, topic models have great potential in this domain. But, in practice, they are often difficult to employ successfully, and are notoriously inconsistent as latent space dimension grows. In this manuscript we identify the three properties all usable topic models should have: robustness, descriptive power and reflection of reality. We develop a novel method for evaluating the robustness of topic models and suggest a metric to assess and benchmark descriptive power as number of topics scale. Employing that procedure, we find that the neural-network-based paragraph embedding approach seems capable of providing statistically robust estimates of the document-document similarities, even for topic spaces far larger than what is usually considered prudent for the most common topic model approaches.