Digital data plays a central role in sciences, journalism, environment, digital humanities, etc. Open Data sharing initiatives lead to many large, interesting datasets being shared online. Some of these are RDF graphs, but other formats like CSV, relational, property graphs, JSON or XML documents are also frequent. Practitioners need to understand a dataset to decide whether it is suited to their needs. Datasets may come with a schema and/or may be summarized, however the first is not always provided and the latter is often too technical for non-IT users. To overcome these limitations, we present an end-to-end dataset abstraction approach, which (i) applies on any (semi)structured data model; (ii) computes a description meant for human users, in the form of an Entity- Relationship diagram; (ii) integrates Information Extraction and data profiling to classify dataset entities among a large set of intelligible categories. We implemented our approach in a system called Abstra, and detail its performance on various datasets.