Spacy Babelnet

A Spacy pipeline component that annotates tokens with their corresponding Babelnet Synsets (and Lemmas). The synsets are searched only in the specified language, but other languages can be retrieved through Babelnet. If the token has a POS annotation, the synsets are searched only with that POS. Notice that Babelnet uses only the 4 main POS tags: NAME,ADJ,VERB,ADV. Only tokens that map to these 4 types are searched in BabelNet.

Build and install the babelnet module

The babelnet module is a python wrapper to the Babelnet API jars

• install the jcc and openjdk packages
  conda install jcc openjdk
• or else
  pip install jcc
  and install openjdk for your Linux distribution
• edit Makefile and set JNI_DIR to the directory containing include/jni.h
• download and unzip the Babelnet-API archive version 5.0
  make get_api
• build and install the babelnet module
  make babelnet

Build and install the spacy_babelnet module

• make spacy-babelnet
  or else
• python setup.py install

Copy the config directory containing your key

The babelnet module must find the config directory in the current directory. Copy the config directory and edit the config/babelnet.var.properties to add your BabelNet API key.

Optional: local install of the BabelNet indices (29G compressed, 49G on disk)

• download the BabelNet indices from BabelNet.org
• unzip them
• edit the config/babelnet.vars.properties file to set the indices directory

Usage Example

The wrapper adds the 'babelnet' property to tokens, containing a Babelnet object that can be used to retrieve its synsets or lemmas

import spacy, babelnet
from spacy_babelnet import BabelnetAnnotator

nlp = spacy.load('it_core_news_lg')	# spacy 3

# with spacy 3
nlp.add_pipe('babelnet', config={
	#'domain': babelnet.BabelDomain.BUSINESS_INDUSTRY_AND_FINANCE.value,
	'source': babelnet.BabelSenseSource.OMWN_IT.toString()
	})

doc = nlp('Mi piace la pizza')    # I like pizza
for token in doc:
    print(token, token.pos_, token._.babelnet.synsets(), token._.babelnet.lemmas(), sep='\n\t')

That produces the output

Mi
    PRON
    []
    []
piace
    VERB
    [<BabelSynset: bn:00090362v__like#v#2>, <BabelSynset: bn:00086519v__delight#v#1>, <BabelSynset: bn:00090363v__like#v#3>, <BabelSynset: bn:00087646v__enjoy#v#3>, <BabelSynset: bn:00084526v__wish#v#2>]
    [<BabelLemma: piacere>, <BabelLemma: contentare>, <BabelLemma: provare_gioia>, <BabelLemma: piacere>, <BabelLemma: assecondare>, <BabelLemma: dilettare>, <BabelLemma: preferire>, <BabelLemma: compiacere>, <BabelLemma: piacere>, <BabelLemma: godere>, <BabelLemma: accontentare>, <BabelLemma: appagare>, <BabelLemma: piace>, <BabelLemma: deliziare>, <BabelLemma: gratificare>, <BabelLemma: piacere>, <BabelLemma: prediligere>, <BabelLemma: piacere>, <BabelLemma: amare>, <BabelLemma: soddisfare>]
la
    DET
    []
    []
pizza
    NOUN
    [<BabelSynset: bn:00062694n__pizza#n#1>, <BabelSynset: bn:00066766n__reel#n#1>, <BabelSynset: bn:00012225n__bore#n#1>]
    [<BabelLemma: pizza_al_taglio>, <BabelLemma: mattone>, <BabelLemma: pizzetta>, <BabelLemma: pizza_surgelata>, <BabelLemma: impiastro>, <BabelLemma: pizza_a_taglio>, <BabelLemma: pizza_congelata>, <BabelLemma: pizza_pie>, <BabelLemma: pizze>, <BabelLemma: palla>, <BabelLemma: pasta_della_pizza>, <BabelLemma: bobina>, <BabelLemma: pizza>, <BabelLemma: noioso>, <BabelLemma: pizza>, <BabelLemma: pizza>, <BabelLemma: pizza_classica>, <BabelLemma: noia>, <BabelLemma: forno_da_pizza>, <BabelLemma: lagna>, <BabelLemma: pasta_per_pizza>, <BabelLemma: cataplasma>, <BabelLemma: pizza_tonda>, <BabelLemma: pizzette>, <BabelLemma: pizza_in_teglia>, <BabelLemma: pittima>, <BabelLemma: forno_per_la_pizza>, <BabelLemma: pizza_alla_pala>]

TODO

• add tests
• find a better location for the babelnet configuration (~/babelnet_data ?)
• speed-up synsets retrieval
  • convert BN to a simpler/faster database?
  • build a lemma-synset external index?

DONE

• added domain and source parameters
• added (word,pos) caching of synsets shared at class level, with json save/load
• mapped other spacy POS types to Babelnet POS types
• filter out tokens with unmapped POS
• moved to faster LKB queries
• added source selection