Abstract
We develop a generative attention-based approach to modeling structuredentities comprising different property types, such as numerical, categorical,string, and composite. This approach handles such heterogeneous data through amixed continuous-discrete diffusion process over the properties. Our flexibleframework can model entities with arbitrary hierarchical properties, enablingapplications to structured Knowledge Base (KB) entities and tabular data. Ourapproach obtains state-of-the-art performance on a majority of cases across 15datasets. In addition, experiments with a device KB and a nuclear physicsdataset demonstrate the model's ability to learn representations useful forentity completion in diverse settings. This has many downstream use cases,including modeling numerical properties with high accuracy - critical forscience applications, which also benefit from the model's inherentprobabilistic nature.