No description or website provided.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
WN18
WN18RR
LICENSE
README.md

README.md

Transforming WN18 / WN18RR back to Text

I wanted to work with the datasets WN18 and WN18RR that contain 18/11 relations from wordnet data.

The original WN18RR dataset has the following form:

02174461  _hypernym 02176268
05074057  _derivationally_related_form  02310895
08390511  _synset_domain_topic_of 08199025
02045024  _member_meronym 02046321
01257145 _derivationally_related_form 07488875
...

I wanted to have the textual representation of the entities, but only the wordnet offsets are given as entites, transforming them back is problematic cause they are ambiguous within the 4 datafiles from wordnet.

For example 01257145 _derivationally_related_form 07488875 has two offsets: 01257145 and 07488875.

01257145 07488875
ADJ sensual.s.02
ADV
NOUN precession.n.02 sensuality.n.01
VERB

I transformed the dataset back to wordnet synsets by validating if the given relation holds between the ambiguous entities.

The transformed textual data then looks like this:

clangor.v.01  _hypernym sound.v.02
straightness.n.02 _derivationally_related_form  straight.a.02
militia.n.01  _synset_domain_topic_of military.n.01
alcidae.n.01  _member_meronym pinguinus.n.01
sensual.s.02  _derivationally_related_form  sensuality.n.01

You can load it into NLTK by executing

from nltk.corpus import wordnet as wn
wn.synset('sensual.s.02')

Working with WN18 (a warning)

As first stated by Toutanova in 2015 and confirmed by Dettmers in 2018, the dataset suffers from informative value, cause >80% of the test triples (e1, r1, e2) can be found in the training set with another relation: (e1, r2, e2) or (e2, r2, e1). Dettmers used a rule-based model which learned the inverse relation and achieved state-of-the-art results on that dataset. It should therefore not used for research evaluation anymore.

Source/Credit

I got the WN18RR dataset from TimDettmers/ConvE. As the original WN18 is down, I obtained a copy from Github.