Integrating data coming from different knowledge bases has been one of the most important tasks in the Semantic Web the last years. Keys have been considered to be very useful in the data linking task. A set of properties is considered a key if it uniquely identifies every resource in the data. To cope with the incompleteness of the data, three different key semantics have been proposed so far. We propose BECKEY, a semantic agnostic approach that discovers keys for all three semantics, succeeding to scale on large datasets. Our approach is able to discover keys under the presence of erroneous data or duplicates (i.e., almost keys). A formalisation of the three semantics along with the relations among them is provided. An extended experimental comparison of the three key semantics has taken place. The results allow a better understanding of the three semantics, providing insights on when each semantic is more appropriate for the task of data linking.