lexicons-builder 0.2.0

lexicons_builder, a tool to create lexicons

The lexicons_builder package aims to provide a basic API to create lexicons related to specific words.

Key principle: Given the input words, it will look for synonyms or neighboors in the dictionnaries or in the NLP model. For each of the new retreiven terms, it will look again for its neighboors or synonyms and so on..

The general method is implemented on 3 different supports:

1. Synonyms dictionnaries (See list of the dictionnaries : ref:here <list_dictionnaries.rst>)

2. NLP language models

3. WordNet (or WOLF)

Output can be text file, xlsx file, turtle file or a Graph object. See <Quickstart> section for examples.

Full documentation available on readthedocs

Note

Feel free to raise an issue on GitHub if something isn’t working for you.

Installation

With pip

$ pip install lexicons-builder

From source

To install the module from source:

$ pip install git+git://github.com/GuillaumeLNB/lexicons_builder

Download NLP models (optionnal)

Here’s a non exhaustive list of websites where you can download NLP models manually. The models should be in word2vec or fasttext format.

Link	Language(s)
https://fauconnier.github.io/#data	French
https://wikipedia2vec.github.io/wikipedia2vec/pretrained/	Multilingual
http://vectors.nlpl.eu/repository/	Multilingual
https://github.com/alexandres/lexvec#pre-trained-vectors	Multilingual
https://fasttext.cc/docs/en/english-vectors.html	English / Multilingual
https://github.com/mmihaltz/word2vec-GoogleNews-vectors	English

Download WOLF (French WordNet) (optionnal)

$ # download WOLF (French wordnet if needed)
$ wget https://gforge.inria.fr/frs/download.php/file/33496/wolf-1.0b4.xml.bz2
$ # (and extract it)
$ bzip2 -d wolf-1.0b4.xml.bz2

QuickStart

Command line interface (CLI)

To get words from input words through CLI, run

$ python -m lexicons_builder <words>  \
      --lang <LANG>                 \
      --out-file <OUTFILE>          \
      --format <FORMAT>             \
      --depth <DEPTH>               \
      --nlp-model <NLP_MODEL_PATHS> \
      --web                         \
      --wordnet                     \
      --wolf-path <WOLF_PATH>       \
      --strict

With:

• <words> The word(s) we want to get synonyms from

• <LANG> The word language (eg: fr, en, nl, …)

• <DEPTH> The depth we want to dig in the models, websites, …

• <OUTFILE> The file where the results will be stored

• <FORMAT> The wanted output format (txt with indentation, ttl or xlsx)

At least ONE of the following options is needed:

• --nlp-model <NLP_MODEL_PATHS> The path to the nlp model(s)

• --web Search online for synonyms

• --wordnet Search on WordNet using nltk

• --wolf-path <WOLF_PATH> The path to WOLF (French wordnet)

Optional

• --strict remove non relevant words

Eg: if we want to look for related terms linked to ‘eat’ and ‘drink’ on wordnet at a depth of 2, excecute:

$ python -m lexicons_builder eat drink  \
      --lang        en                  \
      --out-file    test_en.txt         \
      --format      txt                 \
      --depth       1                   \
      --wordnet
$ Note the indentation is linked to the depth a which the word was found
$ head test_en.txt
  drink
  eat
    absorb
    ade
    aerophagia
    alcohol
    alcoholic_beverage
    alcoholic_drink
    banquet
    bar_hop
    belt_down
    beverage
    bi
  ...

Python

To get related terms interactively through Python, run

>>> from lexicons_builder import build_lexicon
>>> # search for related terms of 'book' and 'read' in English at depth 1 online
>>> output = build_lexicon(["book", "read"], 'en', 1, web=True)
...
>>> # we then get a graph object
>>> # output as a list
>>> output.to_list()
['PS', 'accept', 'accommodate', 'according to the rules', 'account book', 'accountability', 'accountancy', 'accountant', 'accounting', 'accounts', 'accuse', 'acquire', 'act', 'adjudge', 'admit', 'adopt', 'afl', 'agree', 'aim', "al-qur'an", 'album', 'allege', 'allocate', 'allow', 'analyse', 'analyze', 'annuaire', 'anthology', 'appear in reading', 'apply', 'appropriate', 'arrange', 'arrange for', 'arrest', 'articulate', 'ascertain' ...
>>> # output as rdf/turtle
>>> print(output)
@prefix ns1: <http://taxref.mnhn.fr/lod/property/> .
@prefix ns2: <urn:default:baseUri:#> .
@prefix ns3: <http://www.w3.org/2004/02/skos/core#> .
@prefix xsd: <http://www.w3.org/2001/XMLSchema#> .

ns2:PS ns1:isSynonymOf ns2:root_word_uri ;
    ns3:prefLabel "PS" ;
    ns2:comesFrom <synonyms.com> ;
    ns2:depth 1 .

ns2:accept ns1:isSynonymOf ns2:root_word_uri ;
    ns3:prefLabel "accept" ;
    ns2:comesFrom <synonyms.com> ;
    ns2:depth 1 .
...

>>> # output to an indented file
>>> output.to_text_file("filename.txt")
>>> with open("filename.txt") as f:
...     print(f.read(1000))
...
read
book
  PS
  accept
  accommodate
  according to the rules
  account book
  accountability
...
>>> # output to xslx file
>>> output.to_xlsx_file("results.xlsx")

>>> # full search with 2 nlp models, wordnet and on the web
>>> # download and extract google word2vec model
>>> # from https://github.com/mmihaltz/word2vec-GoogleNews-vectors
>>>
>>> # download and extract FastText models
>>> # from https://fasttext.cc/docs/en/english-vectors.html
>>>
>>> nlp_models = ["GoogleNews-vectors-negative300.bin", "wiki-news-300d-1M.vec"]
>>> output = build_lexicon(["book", "letter"], "en", 1, web=True, wordnet=True, nlp_model_paths=nlp_models)
>>> # can take a while
>>> len(output.to_list())
614
>>> # deleting non relevant words
>>> output.pop_non_relevant_words()
>>> len(output.to_list())
57

Note

If the depth parameter is too high (higher than 3), the words found could seem unrelated to the root words. It can take also a long time to compute too.

Note

The word senses are taken equally, which means that you might get terms you would think are not related to the input word. Eg: looking for the word ‘test’ might give you words linked to Sea urchins, as a ‘test’ is also a type of shell https://en.wikipedia.org/wiki/Test_(biology).