GermaNet API for Python.

Copyright (c) 23 March, 2014 Will Roberts <[email protected]>.

GermaNet is the German WordNet, a machine-readable lexical semantic resource which lists nouns, verbs and adjectives in German, along with lexical relations which connect these words together into a semantic network. This library gives a Python interface to this resource.

Start GermaNet by connecting to the MongoDB database which contains the lexical information (for setting up the MongoDB database, see the section Setup, below). On the local machine using the default port, this is as simple as:

>>> from pygermanet import load_germanet
>>> gn = load_germanet()

You can search GermaNet for synsets containing a particular lemmatised word form using the synsets function:

>>> gn.synsets('gehen')
[Synset(auseinandergehen.v.3),
 Synset(funktionieren.v.1),
 Synset(funktionieren.v.2),
 Synset(gehen.v.1),
 Synset(gehen.v.4),
 Synset(gehen.v.5),
 Synset(gehen.v.6),
 Synset(gehen.v.7),
 Synset(gehen.v.9),
 Synset(gehen.v.10),
 Synset(gehen.v.11),
 Synset(gehen.v.12),
 Synset(gehen.v.13),
 Synset(gehen.v.14),
 Synset(handeln.v.1)]

To look up synsets, you must have the canonical form of the word as it would appear in a dictionary (head word); this module calls this word form the lemma. The GermaNet instance can perform lemmatisation of words using data derived from the Projekt deutscher Wortschatz:

>>> gn.lemmatise(u'ginge')
[u'gehen']

Each Synset is represented by the orthographic form, part of speech, and sense number of its first Lemma; this acts as a unique identifier for synsets. If you know this identifier, you can also look up a synset in GermaNet:

>>> funktionieren = gn.synset(u'funktionieren.v.2')
>>> funktionieren
Synset(funktionieren.v.2)

Synset objects have data members and methods:

>>> funktionieren.hyponyms
[Synset(vorgehen.v.1), Synset(leerlaufen.v.2)]
>>> gn.synset('Husky.n.1').hypernym_paths
[[Synset(GNROOT.n.1),
  Synset(Entität.n.2),
  Synset(Objekt.n.4),
  Synset(Ding.n.2),
  Synset(Teil.n.2),
  Synset(Teilmenge.n.2),
  Synset(Gruppe.n.1),
  Synset(biologische Gruppe.n.1),
  Synset(Spezies.n.1),
  Synset(Rasse.n.1),
  Synset(Tierrasse.n.1),
  Synset(Hunderasse.n.1),
  Synset(Husky.n.1)],
 [Synset(GNROOT.n.1),
  Synset(Entität.n.2),
  Synset(kognitives Objekt.n.1),
  Synset(Kategorie.n.1),
  Synset(Art.n.1),
  Synset(Spezies.n.1),
  Synset(Rasse.n.1),
  Synset(Tierrasse.n.1),
  Synset(Hunderasse.n.1),
  Synset(Husky.n.1)],
 [Synset(GNROOT.n.1),
  Synset(Entität.n.2),
  Synset(Objekt.n.4),
  Synset(natürliches Objekt.n.1),
  Synset(Wesenheit.n.1),
  Synset(Organismus.n.1),
  Synset(höheres Lebewesen.n.1),
  Synset(Tier.n.1),
  Synset(Gewebetier.n.1),
  Synset(Chordatier.n.1),
  Synset(Wirbeltier.n.1),
  Synset(Säugetier.n.1),
  Synset(Plazentatier.n.1),
  Synset(Beutegreifer.n.1),
  Synset(Landraubtier.n.1),
  Synset(hundeartiges Landraubtier.n.1),
  Synset(Hund.n.2),
  Synset(Husky.n.1)],
 [Synset(GNROOT.n.1),
  Synset(Entität.n.2),
  Synset(Objekt.n.4),
  Synset(natürliches Objekt.n.1),
  Synset(Wesenheit.n.1),
  Synset(Organismus.n.1),
  Synset(höheres Lebewesen.n.1),
  Synset(Tier.n.1),
  Synset(Haustier.n.1),
  Synset(Hund.n.2),
  Synset(Husky.n.1)]]

Each Synset contains one or more Lemma objects:

>>> funktionieren.lemmas
[Lemma(funktionieren.v.2.funktionieren),
 Lemma(funktionieren.v.2.funzen),
 Lemma(funktionieren.v.2.gehen),
 Lemma(funktionieren.v.2.laufen),
 Lemma(funktionieren.v.2.arbeiten)]

A given orthographic form may be represented by multiple Lemma objects belonging to different Synset objects:

>>> gn.lemmas('brennen')
[Lemma(brennen.v.1.brennen),
 Lemma(verbrennen.v.1.brennen),
 Lemma(brennen.v.3.brennen),
 Lemma(brennen.v.4.brennen),
 Lemma(brennen.v.5.brennen),
 Lemma(destillieren.v.1.brennen),
 Lemma(brennen.v.7.brennen),
 Lemma(brennen.v.8.brennen)]

pygermanet includes several functions for calculating semantic similarity and semantic distance, somewhat like WN::Similarity. These metrics use word frequency information estimated on the SdeWaC corpus and then automatically lemmatised using the TreeTagger.

The probability of encountering an instance of a given synset s is estimated over SdeWaC using the procedure described by Resnik (1995). Briefly, for each instance of a noun in the corpus, we find the set of synsets S containing a sense of that noun; each of these synsets is then credited with a count of 1/|S|. A count added to a synset is also added to all of its hypernyms, so that count observations are propagated up the taxonomy. By dividing by the total number of noun instances in the corpus, each synset is assigned a probability value; these probabilities increase monotonically up the taxonomy, and the root node has p = 1.

Using this interface, we can replicate the results of (Gurevych, 2005) and (Gurevych and Niederlich, 2005), who collected human semantic similarity judgements on 65 word pairs and then measured the correlation of these judgements against similarity scores reported by various automatic similarity metrics. These two papers reported Pearson's r of 0.715 for (Resnik, 1995), 0.738 for a normalised distance version of (Jiang and Conrath, 1997), and 0.734 for (Lin, 1998), with inter-annotator agreement of 0.810.

Replication of the two studies, using the gur65 data set:

from pygermanet import load_germanet, Synset
from scipy.stats.stats import pearsonr
import codecs
import numpy as np

GUR65_FILENAME = 'gur65.csv'

def load_gurevych():
    gur65 = []
    with codecs.open(GUR65_FILENAME, 'r', 'latin-1') as input_file:
        for idx, line in enumerate(input_file):
            fields = line.strip().split(';')
            if idx == 0:
                header = fields
            else:
                # fix typo in gur65
                fields[1] = {'Reis': 'Reise'}.get(fields[1], fields[1])
                fields[2] = float(fields[2])
                fields[3] = float(fields[3])
                gur65.append(fields)
    gur65 = np.core.records.array(
        gur65,
        dtype=np.dtype({'formats': ['U30', 'U30', '<f8', '<f8'],
                        'names': header}))
    return gur65

gur65 = load_gurevych()
gn    = load_germanet()

# select those words which are found in GermaNet; exclude the
# adjective "jung"
pred = lambda w1, w2: bool(gn.synsets(w1) and gn.synsets(w2) and
                           w1 != 'jung' and w2 != 'jung')

print 'Semantic similarity computed on {0} of {1} word pairs'.format(
    sum([1 for word1, word2 in zip(gur65['Word1'], gur65['Word2'])
         if pred(word1, word2)]),
    len(gur65))

sim_funcs = [('lch', Synset.sim_lch,  np.max),
             ('res', Synset.sim_res,  np.max),
             ('jcn', Synset.dist_jcn, np.min),
             ('lin', Synset.sim_lin,  np.max)]

print
print 'metric   r'
print '---------------'
for sim_name, sim_func, comb_func in sim_funcs:
    scores = []
    for word1, word2, human, _hstd in gur65:
        if not pred(word1, word2):
            continue
        score = comb_func(np.array([sim_func(ss1, ss2)
                                    for ss1 in gn.synsets(word1)
                                    for ss2 in gn.synsets(word2)]))
        scores.append([score, human])
    scores = np.array(scores)
    r, _p = pearsonr(scores[:,0],scores[:,1])
    print '{0}      {1:.3f}'.format(sim_name, r)

This script outputs:

Semantic similarity computed on 60 of 65 word pairs

metric   r
---------------
lch      0.742
res      0.715
jcn      -0.770
lin      0.737

• The GermaNet resource (available from the University of Tübingen)
• Python 2.7 or Python 3 (tested with Python 3.4)
• MongoDB
• pymongo
• future (for Python 2)
• repoze.lru (optional)

Example setup:

sudo apt-get install mongodb
sudo pip install repoze.lru pygermanet

GermaNet is distributed as a set of XML files, or as a PostgreSQL database dump, neither of which is a convenient format to handle from inside Python. This library delegates responsibility for handling the data to a MongoDB database. As such, setup happens in two steps.

1. Start a MongoDB instance running. For example, the start_mongo.sh script contains:

   mkdir -p ./mongodb
   mongod --dbpath ./mongodb

2. Import GermaNet into the MongoDB instance. The mongo_import.py script needs the path to the directory that contains the GermaNet XML files:

   python -m pygermanet.mongo_import ~/corpora/germanet/GN_V80/GN_V80_XML/

   This step only needs to be performed once, before you use pygermanet for the first time.

3. pygermanet can now be used by connecting to the running MongoDB instance. Using default settings and connecting to a database on the local machine, this is accomplished with:

   >>> from pygermanet import load_germanet
   >>> gn = load_germanet()

This README file and the source code in this library are licensed under the MIT License (see source file LICENSE.txt for details).

The file baseforms_by_projekt_deutscher_wortschatz.txt.gz contains data derived from the Projekt deutscher Wortschatz; this database is free for educational and researching purposes but not for commercial use. For more information visit: http://wortschatz.uni-leipzig.de/.

The file sdewac-gn-words.tsv.gz contains data derived from the SdeWaC corpus, whose license and copyright status are slightly unclear to me. I'm releasing the file under the Creative Commons Attribution 4.0 International License.

The NLTK project had an API once upon a time for interacting with GermaNet, but this has now been removed from the current NLTK distribution. This API was called GermaNLTK and was described in some detail in NLTK Issue 604. pygermanet shamelessly imitates the interface of this older NLTK code, which was, in turn, based on the standard NLTK interface to WordNet.

The GermaNLTK project had a script to push the content of the XML files into a sqlite database; I haven't tested this code myself, and the GermaNet database has changed over the years since GermaNLTK was written. This mongo_import.py script included in this library does much the same thing, and could easily be adapted to use sqlite as a back end instead of MongoDB.

Thanks to stefanpernes for his work on making a Python 3 port of pygermanet.