For my work, the most important feature would be "same-as" (like
Laura), with the same justification.
Afterwards, I would use reasoning that is achievable with simple
construct queries and then store the result.
An actual example:
construct {?tok a ?pattern .}
FROM <http://gerastree.at/test1> # contains ?tok nlp:lemma3 ?lem
from <http://gerastree.at/wc12>
# contains ?lexentry lemon:writtenRep ?lem.
where {
?lexentry ?kind ?pattern .
?lexentry lemon:writtenRep ?lem .
?tok nlp:lemma3 ?lem .
?tok nlp:pos ?cats.
filter (not exists {wn:isNotClass ?kind ?lem} ) .
}