Sample usage for wordnet¶

WordNet Interface¶

WordNet is just another NLTK corpus reader, and can be imported like this:

>>> from nltk.corpus import wordnet

For more compact code, we recommend:

>>> from nltk.corpus import wordnet as wn

Words¶

Look up a word using synsets(); this function has an optional pos argument which lets you constrain the part of speech of the word:

>>> wn.synsets('dog')
[Synset('dog.n.01'), Synset('frump.n.01'), Synset('dog.n.03'), Synset('cad.n.01'),
Synset('frank.n.02'), Synset('pawl.n.01'), Synset('andiron.n.01'), Synset('chase.v.01')]
>>> wn.synsets('dog', pos=wn.VERB)
[Synset('chase.v.01')]

The other parts of speech are NOUN, ADJ and ADV. A synset is identified with a 3-part name of the form: word.pos.nn:

>>> wn.synset('dog.n.01')
Synset('dog.n.01')
>>> print(wn.synset('dog.n.01').definition())
a member of the genus Canis (probably descended from the common wolf) that has been domesticated by man since prehistoric times; occurs in many breeds
>>> len(wn.synset('dog.n.01').examples())
1
>>> print(wn.synset('dog.n.01').examples()[0])
the dog barked all night
>>> wn.synset('dog.n.01').lemmas()
[Lemma('dog.n.01.dog'), Lemma('dog.n.01.domestic_dog'), Lemma('dog.n.01.Canis_familiaris')]
>>> [str(lemma.name()) for lemma in wn.synset('dog.n.01').lemmas()]
['dog', 'domestic_dog', 'Canis_familiaris']
>>> wn.lemma('dog.n.01.dog').synset()
Synset('dog.n.01')

The WordNet corpus reader gives access to the Open Multilingual WordNet, using ISO-639 language codes. These languages are not loaded by default, but only lazily, when needed.

>>> wn.langs()
['eng']

>>> wn.synsets(b'\xe7\x8a\xac'.decode('utf-8'), lang='jpn')
[Synset('dog.n.01'), Synset('spy.n.01')]

>>> wn.synset('spy.n.01').lemma_names('jpn')
['いぬ', 'まわし者', 'スパイ', '回し者', '回者', '密偵',
'工作員', '廻し者', '廻者', '探', '探り', '犬', '秘密捜査員',
'諜報員', '諜者', '間者', '間諜', '隠密']

>>> sorted(wn.langs())
['als', 'arb', 'bul', 'cat', 'cmn', 'dan', 'ell', 'eng', 'eus',
'fin', 'fra', 'glg', 'heb', 'hrv', 'ind', 'isl', 'ita', 'ita_iwn',
'jpn', 'lit', 'nld', 'nno', 'nob', 'pol', 'por', 'ron', 'slk',
'slv', 'spa', 'swe', 'tha', 'zsm']

>>> wn.synset('dog.n.01').lemma_names('ita')
['Canis_familiaris', 'cane']
>>> wn.lemmas('cane', lang='ita')
[Lemma('dog.n.01.cane'), Lemma('cramp.n.02.cane'), Lemma('hammer.n.01.cane'), Lemma('bad_person.n.01.cane'),
Lemma('incompetent.n.01.cane')]
>>> sorted(wn.synset('dog.n.01').lemmas('dan'))
[Lemma('dog.n.01.hund'), Lemma('dog.n.01.k\xf8ter'),
Lemma('dog.n.01.vovhund'), Lemma('dog.n.01.vovse')]

>>> sorted(wn.synset('dog.n.01').lemmas('por'))
[Lemma('dog.n.01.cachorra'), Lemma('dog.n.01.cachorro'), Lemma('dog.n.01.cadela'), Lemma('dog.n.01.c\xe3o')]

>>> dog_lemma = wn.lemma(b'dog.n.01.c\xc3\xa3o'.decode('utf-8'), lang='por')
>>> dog_lemma
Lemma('dog.n.01.c\xe3o')
>>> dog_lemma.lang()
'por'
>>> len(list(wordnet.all_lemma_names(pos='n', lang='jpn')))
66031

The synonyms of a word are returned as a nested list of synonyms of the different senses of the input word in the given language, since these different senses are not mutual synonyms:

>>> wn.synonyms('car')
[['auto', 'automobile', 'machine', 'motorcar'], ['railcar', 'railroad_car', 'railway_car'], ['gondola'], ['elevator_car'], ['cable_car']]
>>> wn.synonyms('coche', lang='spa')
[['auto', 'automóvil', 'carro', 'máquina', 'turismo', 'vehículo'], ['automotor', 'vagón'], ['vagón', 'vagón_de_pasajeros']]

Synsets¶

Synset: a set of synonyms that share a common meaning.

>>> dog = wn.synset('dog.n.01')
>>> sorted(dog.hypernyms())
[Synset('canine.n.02'), Synset('domestic_animal.n.01')]
>>> sorted(dog.hyponyms())
[Synset('basenji.n.01'), Synset('corgi.n.01'), Synset('cur.n.01'), Synset('dalmatian.n.02'), ...]
>>> sorted(dog.member_holonyms())
[Synset('canis.n.01'), Synset('pack.n.06')]
>>> sorted(dog.root_hypernyms())
[Synset('entity.n.01')]
>>> wn.synset('dog.n.01').lowest_common_hypernyms(wn.synset('cat.n.01'))
[Synset('carnivore.n.01')]

Each synset contains one or more lemmas, which represent a specific sense of a specific word.

Note that some relations are defined by WordNet only over Lemmas:

>>> good = wn.synset('good.a.01')
>>> good.antonyms()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'Synset' object has no attribute 'antonyms'
>>> good.lemmas()[0].antonyms()
[Lemma('bad.a.01.bad')]

The relations that are currently defined in this way are antonyms, derivationally_related_forms and pertainyms.

If you know the byte offset used to identify a synset in the original Princeton WordNet data file, you can use that to instantiate the synset in NLTK:

>>> wn.synset_from_pos_and_offset('n', 4543158)
Synset('wagon.n.01')

Likewise, instantiate a synset from a known sense key:

>>> wn.synset_from_sense_key("driving%1:04:03::")
Synset('drive.n.06')

Lemmas¶

>>> eat = wn.lemma('eat.v.03.eat')
>>> eat
Lemma('feed.v.06.eat')
>>> print(eat.key())
eat%2:34:02::
>>> eat.count()
4
>>> wn.lemma_from_key(eat.key())
Lemma('feed.v.06.eat')
>>> wn.lemma_from_key(eat.key()).synset()
Synset('feed.v.06')
>>> wn.lemma_from_key('feebleminded%5:00:00:retarded:00')
Lemma('backward.s.03.feebleminded')
>>> for lemma in wn.synset('eat.v.03').lemmas():
...     print(lemma, lemma.count())
...
Lemma('feed.v.06.feed') 3
Lemma('feed.v.06.eat') 4
>>> for lemma in wn.lemmas('eat', 'v'):
...     print(lemma, lemma.count())
...
Lemma('eat.v.01.eat') 61
Lemma('eat.v.02.eat') 13
Lemma('feed.v.06.eat') 4
Lemma('eat.v.04.eat') 0
Lemma('consume.v.05.eat') 0
Lemma('corrode.v.01.eat') 0
>>> wn.lemma('jump.v.11.jump')
Lemma('jump.v.11.jump')

Lemmas can also have relations between them:

>>> vocal = wn.lemma('vocal.a.01.vocal')
>>> vocal.derivationally_related_forms()
[Lemma('vocalize.v.02.vocalize')]
>>> vocal.pertainyms()
[Lemma('voice.n.02.voice')]
>>> vocal.antonyms()
[Lemma('instrumental.a.01.instrumental')]

The three relations above exist only on lemmas, not on synsets.

Verb Frames¶

>>> wn.synset('think.v.01').frame_ids()
[5, 9]
>>> for lemma in wn.synset('think.v.01').lemmas():
...     print(lemma, lemma.frame_ids())
...     print(" | ".join(lemma.frame_strings()))
...
Lemma('think.v.01.think') [5, 9]
Something think something Adjective/Noun | Somebody think somebody
Lemma('think.v.01.believe') [5, 9]
Something believe something Adjective/Noun | Somebody believe somebody
Lemma('think.v.01.consider') [5, 9]
Something consider something Adjective/Noun | Somebody consider somebody
Lemma('think.v.01.conceive') [5, 9]
Something conceive something Adjective/Noun | Somebody conceive somebody
>>> wn.synset('stretch.v.02').frame_ids()
[8]
>>> for lemma in wn.synset('stretch.v.02').lemmas():
...     print(lemma, lemma.frame_ids())
...     print(" | ".join(lemma.frame_strings()))
...
Lemma('stretch.v.02.stretch') [8, 2]
Somebody stretch something | Somebody stretch
Lemma('stretch.v.02.extend') [8]
Somebody extend something

Similarity¶

>>> dog = wn.synset('dog.n.01')
>>> cat = wn.synset('cat.n.01')

>>> hit = wn.synset('hit.v.01')
>>> slap = wn.synset('slap.v.01')

synset1.path_similarity(synset2): Return a score denoting how similar two word senses are, based on the shortest path that connects the senses in the is-a (hypernym/hypnoym) taxonomy. The score is in the range 0 to 1. By default, there is now a fake root node added to verbs so for cases where previously a path could not be found—and None was returned—it should return a value. The old behavior can be achieved by setting simulate_root to be False. A score of 1 represents identity i.e. comparing a sense with itself will return 1.

>>> dog.path_similarity(cat)
0.2...

>>> hit.path_similarity(slap)
0.142...

>>> wn.path_similarity(hit, slap)
0.142...

>>> print(hit.path_similarity(slap, simulate_root=False))
None

>>> print(wn.path_similarity(hit, slap, simulate_root=False))
None

synset1.lch_similarity(synset2): Leacock-Chodorow Similarity: Return a score denoting how similar two word senses are, based on the shortest path that connects the senses (as above) and the maximum depth of the taxonomy in which the senses occur. The relationship is given as -log(p/2d) where p is the shortest path length and d the taxonomy depth.

>>> dog.lch_similarity(cat)
2.028...

>>> hit.lch_similarity(slap)
1.312...

>>> wn.lch_similarity(hit, slap)
1.312...

>>> print(hit.lch_similarity(slap, simulate_root=False))
None

>>> print(wn.lch_similarity(hit, slap, simulate_root=False))
None

synset1.wup_similarity(synset2): Wu-Palmer Similarity: Return a score denoting how similar two word senses are, based on the depth of the two senses in the taxonomy and that of their Least Common Subsumer (most specific ancestor node). Note that at this time the scores given do not always agree with those given by Pedersen’s Perl implementation of Wordnet Similarity.

The LCS does not necessarily feature in the shortest path connecting the two senses, as it is by definition the common ancestor deepest in the taxonomy, not closest to the two senses. Typically, however, it will so feature. Where multiple candidates for the LCS exist, that whose shortest path to the root node is the longest will be selected. Where the LCS has multiple paths to the root, the longer path is used for the purposes of the calculation.

>>> dog.wup_similarity(cat)
0.857...

>>> hit.wup_similarity(slap)
0.25

>>> wn.wup_similarity(hit, slap)
0.25

>>> print(hit.wup_similarity(slap, simulate_root=False))
None

>>> print(wn.wup_similarity(hit, slap, simulate_root=False))
None

wordnet_ic Information Content: Load an information content file from the wordnet_ic corpus.

>>> from nltk.corpus import wordnet_ic
>>> brown_ic = wordnet_ic.ic('ic-brown.dat')
>>> semcor_ic = wordnet_ic.ic('ic-semcor.dat')

Or you can create an information content dictionary from a corpus (or anything that has a words() method).

>>> from nltk.corpus import genesis
>>> genesis_ic = wn.ic(genesis, False, 0.0)

synset1.res_similarity(synset2, ic): Resnik Similarity: Return a score denoting how similar two word senses are, based on the Information Content (IC) of the Least Common Subsumer (most specific ancestor node). Note that for any similarity measure that uses information content, the result is dependent on the corpus used to generate the information content and the specifics of how the information content was created.

>>> dog.res_similarity(cat, brown_ic)
7.911...
>>> dog.res_similarity(cat, genesis_ic)
7.204...

synset1.jcn_similarity(synset2, ic): Jiang-Conrath Similarity Return a score denoting how similar two word senses are, based on the Information Content (IC) of the Least Common Subsumer (most specific ancestor node) and that of the two input Synsets. The relationship is given by the equation 1 / (IC(s1) + IC(s2) - 2 * IC(lcs)).

>>> dog.jcn_similarity(cat, brown_ic)
0.449...
>>> dog.jcn_similarity(cat, genesis_ic)
0.285...

synset1.lin_similarity(synset2, ic): Lin Similarity: Return a score denoting how similar two word senses are, based on the Information Content (IC) of the Least Common Subsumer (most specific ancestor node) and that of the two input Synsets. The relationship is given by the equation 2 * IC(lcs) / (IC(s1) + IC(s2)).

>>> dog.lin_similarity(cat, semcor_ic)
0.886...

Access to all Synsets¶

Iterate over all the noun synsets:

>>> for synset in list(wn.all_synsets('n'))[:10]:
...     print(synset)
...
Synset('entity.n.01')
Synset('physical_entity.n.01')
Synset('abstraction.n.06')
Synset('thing.n.12')
Synset('object.n.01')
Synset('whole.n.02')
Synset('congener.n.03')
Synset('living_thing.n.01')
Synset('organism.n.01')
Synset('benthos.n.02')

Get all synsets for this word, possibly restricted by POS:

>>> wn.synsets('dog')
[Synset('dog.n.01'), Synset('frump.n.01'), Synset('dog.n.03'), Synset('cad.n.01'), ...]
>>> wn.synsets('dog', pos='v')
[Synset('chase.v.01')]

Walk through the noun synsets looking at their hypernyms:

>>> from itertools import islice
>>> for synset in islice(wn.all_synsets('n'), 5):
...     print(synset, synset.hypernyms())
...
Synset('entity.n.01') []
Synset('physical_entity.n.01') [Synset('entity.n.01')]
Synset('abstraction.n.06') [Synset('entity.n.01')]
Synset('thing.n.12') [Synset('physical_entity.n.01')]
Synset('object.n.01') [Synset('physical_entity.n.01')]

Morphy¶

Look up forms not in WordNet, with the help of Morphy:

>>> wn.morphy('denied', wn.NOUN)
>>> print(wn.morphy('denied', wn.VERB))
deny
>>> wn.synsets('denied', wn.NOUN)
[]
>>> wn.synsets('denied', wn.VERB)
[Synset('deny.v.01'), Synset('deny.v.02'), Synset('deny.v.03'), Synset('deny.v.04'),
Synset('deny.v.05'), Synset('traverse.v.03'), Synset('deny.v.07')]

Morphy uses a combination of inflectional ending rules and exception lists to handle a variety of different possibilities:

>>> print(wn.morphy('dogs'))
dog
>>> print(wn.morphy('churches'))
church
>>> print(wn.morphy('aardwolves'))
aardwolf
>>> print(wn.morphy('abaci'))
abacus
>>> print(wn.morphy('book', wn.NOUN))
book
>>> wn.morphy('hardrock', wn.ADV)
>>> wn.morphy('book', wn.ADJ)
>>> wn.morphy('his', wn.NOUN)
>>>

Synset Closures¶

Compute transitive closures of synsets

>>> dog = wn.synset('dog.n.01')
>>> hypo = lambda s: s.hyponyms()
>>> hyper = lambda s: s.hypernyms()
>>> sorted(dog.closure(hypo, depth=1)) == sorted(dog.hyponyms())
True
>>> sorted(dog.closure(hyper, depth=1)) == sorted(dog.hypernyms())
True
>>> sorted(dog.closure(hypo))
[Synset('affenpinscher.n.01'), Synset('afghan_hound.n.01'), Synset('airedale.n.01'), Synset('american_foxhound.n.01'), ...]
>>> sorted(dog.closure(hyper))
[Synset('animal.n.01'), Synset('canine.n.02'), Synset('carnivore.n.01'), Synset('chordate.n.01'),
Synset('domestic_animal.n.01'), Synset('entity.n.01'), Synset('living_thing.n.01'),
Synset('mammal.n.01'), Synset('object.n.01'), Synset('organism.n.01'), Synset('physical_entity.n.01'),
Synset('placental.n.01'), Synset('vertebrate.n.01'), Synset('whole.n.02')]

Regression Tests¶

Bug 85: morphy returns the base form of a word, if it’s input is given as a base form for a POS for which that word is not defined:

>>> wn.synsets('book', wn.NOUN)
[Synset('book.n.01'), Synset('book.n.02'), Synset('record.n.05'), Synset('script.n.01'), Synset('ledger.n.01'), Synset('book.n.06'), Synset('book.n.07'), Synset('koran.n.01'), Synset('bible.n.01'), Synset('book.n.10'), Synset('book.n.11')]
>>> wn.synsets('book', wn.ADJ)
[]
>>> wn.morphy('book', wn.NOUN)
'book'
>>> wn.morphy('book', wn.ADJ)
>>>

Bug 160: wup_similarity breaks when the two synsets have no common hypernym

>>> t = wn.synsets('picasso')[0]
>>> m = wn.synsets('male')[1]
>>> t.wup_similarity(m)
0.631...

Issue #2278: wup_similarity not commutative when comparing a noun and a verb. Patch #2650 resolved this error. As a result, the output of the following use of wup_similarity no longer returns None.

>>> t = wn.synsets('titan')[1]
>>> s = wn.synsets('say', wn.VERB)[0]
>>> t.wup_similarity(s)
0.142...

Bug 21: “instance of” not included in LCS (very similar to bug 160)

>>> a = wn.synsets("writings")[0]
>>> b = wn.synsets("scripture")[0]
>>> brown_ic = wordnet_ic.ic('ic-brown.dat')
>>> a.jcn_similarity(b, brown_ic)
0.175...

Bug 221: Verb root IC is zero

>>> from nltk.corpus.reader.wordnet import information_content
>>> s = wn.synsets('say', wn.VERB)[0]
>>> information_content(s, brown_ic)
4.623...

Bug 161: Comparison between WN keys/lemmas should not be case sensitive

>>> k = wn.synsets("jefferson")[0].lemmas()[0].key()
>>> wn.lemma_from_key(k)
Lemma('jefferson.n.01.Jefferson')
>>> wn.lemma_from_key(k.upper())
Lemma('jefferson.n.01.Jefferson')

Bug 99: WordNet root_hypernyms gives incorrect results

>>> from nltk.corpus import wordnet as wn
>>> for s in wn.all_synsets(wn.NOUN):
...     if s.root_hypernyms()[0] != wn.synset('entity.n.01'):
...         print(s, s.root_hypernyms())
...
>>>

Bug 382: JCN Division by zero error

>>> tow = wn.synset('tow.v.01')
>>> shlep = wn.synset('shlep.v.02')
>>> from nltk.corpus import wordnet_ic
>>> brown_ic =  wordnet_ic.ic('ic-brown.dat')
>>> tow.jcn_similarity(shlep, brown_ic)
1...e+300

Bug 428: Depth is zero for instance nouns

>>> s = wn.synset("lincoln.n.01")
>>> s.max_depth() > 0
True

Bug 429: Information content smoothing used old reference to all_synsets

>>> genesis_ic = wn.ic(genesis, True, 1.0)

Bug 430: all_synsets used wrong pos lookup when synsets were cached

>>> for ii in wn.all_synsets(): pass
>>> for ii in wn.all_synsets(): pass

Bug 470: shortest_path_distance ignored instance hypernyms

>>> google = wordnet.synsets("google")[0]
>>> earth = wordnet.synsets("earth")[0]
>>> google.wup_similarity(earth)
0.1...

Bug 484: similarity metrics returned -1 instead of None for no LCS

>>> t = wn.synsets('fly', wn.VERB)[0]
>>> s = wn.synsets('say', wn.VERB)[0]
>>> print(s.shortest_path_distance(t))
None
>>> print(s.path_similarity(t, simulate_root=False))
None
>>> print(s.lch_similarity(t, simulate_root=False))
None
>>> print(s.wup_similarity(t, simulate_root=False))
None

Bug 427: “pants” does not return all the senses it should

>>> from nltk.corpus import wordnet
>>> wordnet.synsets("pants",'n')
[Synset('bloomers.n.01'), Synset('pant.n.01'), Synset('trouser.n.01'), Synset('gasp.n.01')]

Bug 482: Some nouns not being lemmatised by WordNetLemmatizer().lemmatize

>>> from nltk.stem.wordnet import WordNetLemmatizer
>>> WordNetLemmatizer().lemmatize("eggs", pos="n")
'egg'
>>> WordNetLemmatizer().lemmatize("legs", pos="n")
'leg'

Bug 284: instance hypernyms not used in similarity calculations

>>> wn.synset('john.n.02').lch_similarity(wn.synset('dog.n.01'))
1.335...
>>> wn.synset('john.n.02').wup_similarity(wn.synset('dog.n.01'))
0.571...
>>> wn.synset('john.n.02').res_similarity(wn.synset('dog.n.01'), brown_ic)
2.224...
>>> wn.synset('john.n.02').jcn_similarity(wn.synset('dog.n.01'), brown_ic)
0.075...
>>> wn.synset('john.n.02').lin_similarity(wn.synset('dog.n.01'), brown_ic)
0.252...
>>> wn.synset('john.n.02').hypernym_paths()
[[Synset('entity.n.01'), ..., Synset('john.n.02')]]

Issue 541: add domains to wordnet

>>> wn.synset('code.n.03').topic_domains()
[Synset('computer_science.n.01')]
>>> wn.synset('pukka.a.01').region_domains()
[Synset('india.n.01')]
>>> wn.synset('freaky.a.01').usage_domains()
[Synset('slang.n.02')]

Issue 629: wordnet failures when python run with -O optimizations

>>> # Run the test suite with python -O to check this
>>> wn.synsets("brunch")
[Synset('brunch.n.01'), Synset('brunch.v.01')]

Issue 395: wordnet returns incorrect result for lowest_common_hypernyms of chef and policeman

>>> wn.synset('policeman.n.01').lowest_common_hypernyms(wn.synset('chef.n.01'))
[Synset('person.n.01')]

Bug https://github.com/nltk/nltk/issues/1641: Non-English lemmas containing capital letters cannot be looked up using wordnet.lemmas() or wordnet.synsets()

>>> wn.lemmas('Londres', lang='fra')
[Lemma('united_kingdom.n.01.Londres'), Lemma('london.n.01.Londres'), Lemma('london.n.02.Londres')]
>>> wn.lemmas('londres', lang='fra')
[Lemma('united_kingdom.n.01.Londres'), Lemma('london.n.01.Londres'), Lemma('london.n.02.Londres')]

Patch-1 https://github.com/nltk/nltk/pull/2065 Adding 3 functions (relations) to WordNet class

>>> sorted(sorted(wn.synsets("computer_science"))[0].in_topic_domains())
[Synset('access.n.05'), Synset('access.v.01'), Synset('access_time.n.01'), Synset('accumulator.n.03'), ...]
>>> sorted(sorted(wn.synsets("France"))[0].in_region_domains())
[Synset('agincourt.n.01'), Synset('ancien_regime.n.01'), Synset('apache_dance.n.01'), Synset('bastille.n.01'), ...]
>>> sorted(sorted(wn.synsets("slang"))[2].in_usage_domains())
[Synset(''hood.n.01'), Synset('airhead.n.01'), Synset('arse.n.02'), Synset('baby.n.05'), Synset('bad_egg.n.01'), ...]

Issue 2721: WordNetCorpusReader.ic() does not add smoothing to N

>>> class FakeCorpus:
...     def words(self): return ['word']
...
>>> fake_ic = wn.ic(FakeCorpus(), False, 1.0)
>>> word = wn.synset('word.n.01')
>>> information_content(word, fake_ic) > 0
True

Issue 3077: Incorrect part-of-speech filtering in all_synsets

>>> next(wn.all_synsets(pos="a"))
Synset('able.a.01')
>>> next(wn.all_synsets(pos="s"))
Synset('emergent.s.02')
>>> wn.add_omw()
>>> next(wn.all_synsets(lang="hrv"))
Synset('able.a.01')
>>> next(wn.all_synsets(lang="hrv", pos="n"))
Synset('entity.n.01')
>>> next(wn.all_synsets(lang="hrv", pos="v"))
Synset('breathe.v.01')
>>> next(wn.all_synsets(lang="hrv", pos="s"))
Synset('ideological.s.02')
>>> next(wn.all_synsets(lang="hrv", pos="a"))
Synset('able.a.01')

Endlessness vs. intractability in relation trees¶

1. Endlessness¶

Until NLTK v. 3.5, the tree() function looped forever on symmetric relations (verb_groups, attributes, and most also_sees). But in the current version, tree() now detects and discards these cycles:

>>> from pprint import pprint
>>> pprint(wn.synset('bound.a.01').tree(lambda s:sorted(s.also_sees())))
[Synset('bound.a.01'),
 [Synset('unfree.a.02'),
  [Synset('confined.a.02'),
   [Synset('restricted.a.01'), [Synset('classified.a.02')]]],
  [Synset('dependent.a.01')],
  [Synset('restricted.a.01'),
   [Synset('classified.a.02')],
   [Synset('confined.a.02')]]]]

Specifying the “cut_mark” parameter increases verbosity, so that the cycles are mentioned in the output, together with the level where they occur:

>>> pprint(wn.synset('bound.a.01').tree(lambda s:sorted(s.also_sees()),cut_mark='...'))
[Synset('bound.a.01'),
 [Synset('unfree.a.02'),
  "Cycle(Synset('bound.a.01'),-3,...)",
  [Synset('confined.a.02'),
   [Synset('restricted.a.01'),
    [Synset('classified.a.02')],
    "Cycle(Synset('confined.a.02'),-5,...)",
    "Cycle(Synset('unfree.a.02'),-5,...)"],
   "Cycle(Synset('unfree.a.02'),-4,...)"],
  [Synset('dependent.a.01'), "Cycle(Synset('unfree.a.02'),-4,...)"],
  [Synset('restricted.a.01'),
   [Synset('classified.a.02')],
   [Synset('confined.a.02'),
    "Cycle(Synset('restricted.a.01'),-5,...)",
    "Cycle(Synset('unfree.a.02'),-5,...)"],
   "Cycle(Synset('unfree.a.02'),-4,...)"]]]

2. Intractability¶

However, even after discarding the infinite cycles, some trees can remain intractable, due to combinatorial explosion in a relation. This happens in WordNet, because the also_sees() relation has a big Strongly Connected Component (_SCC_) consisting in 758 synsets, where any member node is transitively connected by the same relation, to all other members of the same SCC. This produces intractable relation trees for each of these 758 synsets, i. e. trees that are too big to compute or display on any computer.

For example, the synset ‘concrete.a.01’ is a member of the largest SCC, so its also_sees() tree is intractable, and can normally only be handled by limiting the depth parameter to display a small number of levels:

>>> from pprint import pprint
>>> pprint(wn.synset('concrete.a.01').tree(lambda s:sorted(s.also_sees()),cut_mark='...',depth=2))
[Synset('concrete.a.01'),
 [Synset('practical.a.01'),
  "Cycle(Synset('concrete.a.01'),0,...)",
  [Synset('possible.a.01'), '...'],
  [Synset('realistic.a.01'), '...'],
  [Synset('serviceable.a.01'), '...']],
 [Synset('real.a.01'),
  "Cycle(Synset('concrete.a.01'),0,...)",
  [Synset('genuine.a.01'), '...'],
  [Synset('realistic.a.01'), '...'],
  [Synset('sincere.a.01'), '...']],
 [Synset('tangible.a.01'), "Cycle(Synset('concrete.a.01'),0,...)"]]

2.1 First solution: `acyclic_tree()`¶

On the other hand, the new acyclic_tree() function is able to also handle the intractable cases. The also_sees() acyclic tree of ‘concrete.a.01’ is several hundred lines long, so here is a simpler example, concerning a much smaller SCC: counting only five members, the SCC that includes ‘bound.a.01’ is tractable with the normal tree() function, as seen above.

But while tree() only prunes redundancy within local branches, acyclic_tree() prunes the tree globally, thus discarding any additional redundancy, and produces a tree that includes all reachable nodes (i.e., a spanning tree). This tree is minimal because it includes the reachable nodes only once, but it is not necessarily a Minimum Spanning Tree (MST), because the Depth-first search strategy does not guarantee that nodes are reached through the lowest number of links (as Breadth-first search would).

>>> pprint(wn.synset('bound.a.01').acyclic_tree(lambda s:sorted(s.also_sees())))
[Synset('bound.a.01'),
 [Synset('unfree.a.02'),
  [Synset('confined.a.02'),
   [Synset('restricted.a.01'), [Synset('classified.a.02')]]],
  [Synset('dependent.a.01')]]]

Again, specifying the cut_mark parameter increases verbosity, so that the cycles are mentioned in the output, together with the level where they occur:

>>> pprint(wn.synset('bound.a.01').acyclic_tree(lambda s:sorted(s.also_sees()),cut_mark='...'))
[Synset('bound.a.01'),
 [Synset('unfree.a.02'),
  "Cycle(Synset('bound.a.01'),-3,...)",
  [Synset('confined.a.02'),
   [Synset('restricted.a.01'),
    [Synset('classified.a.02')],
    "Cycle(Synset('confined.a.02'),-5,...)",
    "Cycle(Synset('unfree.a.02'),-5,...)"],
   "Cycle(Synset('unfree.a.02'),-4,...)"],
  [Synset('dependent.a.01'), "Cycle(Synset('unfree.a.02'),-4,...)"],
  "Cycle(Synset('restricted.a.01'),-3,...)"]]

2.2 Better solution: mst()¶

A Minimum Spanning Tree (MST) spans all the nodes of a relation subgraph once, while guaranteeing that each node is reached through the shortest path possible. In unweighted relation graphs like WordNet, a MST can be computed very efficiently in linear time, using Breadth-First Search (BFS). Like acyclic_tree(), the new unweighted_minimum_spanning_tree() function (imported in the Wordnet module as mst) handles intractable trees, such as the example discussed above: wn.synset('concrete.a.01').mst(lambda s:sorted(s.also_sees())).

But, while the also_sees() acyclic_tree of ‘bound.a.01’ reaches ‘classified.a.02’ through four links, using depth-first search as seen above (bound.a.01 > unfree.a.02 > confined.a.02 > restricted.a.01 > classified.a.02), in the following MST, the path to ‘classified.a.02’ is the shortest possible, consisting only in three links (bound.a.01 > unfree.a.02 > restricted.a.01 > classified.a.02):

>>> pprint(wn.synset('bound.a.01').mst(lambda s:sorted(s.also_sees())))
[Synset('bound.a.01'),
 [Synset('unfree.a.02'),
  [Synset('confined.a.02')],
  [Synset('dependent.a.01')],
  [Synset('restricted.a.01'), [Synset('classified.a.02')]]]]

Loading alternative Wordnet versions¶

>>> print("Wordnet {}".format(wn.get_version()))
Wordnet 3.0

>>> from nltk.corpus import wordnet31 as wn31
>>> print("Wordnet {}".format(wn31.get_version()))
Wordnet 3.1

>>> print(sorted(wn.synset('restrain.v.01').hyponyms()))
[Synset('confine.v.03'), Synset('control.v.02'), Synset('hold.v.36'), Synset('inhibit.v.04')]

>>> print(sorted(wn31.synset('restrain.v.01').hyponyms()))
[Synset('enchain.v.01'), Synset('fetter.v.01'), Synset('ground.v.02'), Synset('impound.v.02'), Synset('pen_up.v.01'), Synset('pinion.v.01'), Synset('pound.v.06'), Synset('tie_down.v.01')]

>>> print(sorted(wn31.synset('restrain.v.04').hyponyms()))
[Synset('baffle.v.03'), Synset('confine.v.02'), Synset('control.v.02'), Synset('hold.v.36'), Synset('rule.v.07'), Synset('swallow.v.06'), Synset('wink.v.04')]

Teardown test¶

>>> from nltk.corpus import wordnet
>>> wordnet._unload()