The (Semi)Automatic Vossian Antonomasia generation method harnesses the power of Knowledge Graph Embeddings (KGE) trained on Wikidata to generate Vossian Antonomasias. Vossian Antonomasia (VA) is a figure of speech that refers to the substitution of a proper name with a common noun or vice versa. This technique can be used to create memorable and impactful expressions, making it popular in literature, rhetoric, and creative writing.
A VA is generally of the type
A is (or was) the B of C
where A is the target entity that is going to be substituted by a different, popular name. In our setting, A is an arbitrary Wikidata individual and B is a popular Wikidata individual. We classify individuals as popular if they have a wide coverage of Wikipedia articles among different languages.
C is the context that allows a VA to convey a similarity relation in a creative way. In our case, it is always equal to the profession of an entity.
It should hold that A and B can be classified as similar individual. However, with C the profession of A it should hold that C is not the profession of B.
For intance, in the VA Floyd Mayweather is the Michael Jordan of boxing. [1, 2], we have that A = Floyd Mayweather, B = Micheal Jordan and C = boxing.
Clone the repository locally with
git clone git@github.com:MordorISWS23/antonomasia.git
and install the python requirements with
pip install -r requirements.txt
In order to use KGE, the pretrained models on Wikidata needs to be downloaded. We rely on the model shared by GraphVite available via the following link. Currently, only the TransE model is supported. Download the model locally in order to use them. On Linux and MacOS, this can be done using
wget https://udemontreal-my.sharepoint.com/:u:/g/personal/zhaocheng_zhu_umontreal_ca/EX4c1Ud8M61KlDUn2U_yz_sBP_bXNuFnudfhRnYzWUFA2A?download=1 -O transe.pkl
The script antonomasia.py can be used to generate VA.
usage: antonomasia.py [-h] -b B_POOL -a A [--num NUM] [--confidence] [--funny-first] {kge,we,meta} ...
Generate a Vossian Antonomasia
positional arguments:
{kge,we,meta} Method specific parameters
kge Use Knowledge Graph Embeddings
we Use Word Embeddings
meta Use Meta Embeddings
options:
-h, --help show this help message and exit
-b B_POOL, --b_pool B_POOL
Path to the file containing the csv for the set of B entities.
-a A A entity expressed as Wikidata ID - e.g. Q76.
--num NUM Number of sentences to generate.
--confidence Add a confidence score to each generated sentence.
usage: antonomasia.py kge [-h] -i INPUT -p {translate,project}
options:
-h, --help show this help message and exit
-i INPUT, --input INPUT
Path to the KGE weigths.
-p {translate,project}, --projection {translate,project}
Projection method to use.
usage: antonomasia.py we [-h] -m {word2vec,glove} -p {translate,project}
options:
-h, --help show this help message and exit
-m {word2vec,glove}, --model {word2vec,glove}
Word embedding method to use.
-p {translate,project}, --projection {translate,project}
Projection method to use.
usage: antonomasia.py meta [-h] -kge KGE -we {word2vec,glove} -p {translate,project} -c {concatenate,average}
options:
-h, --help show this help message and exit
-kge KGE Path to the KGE weigths.
-we {word2vec,glove} Word embedding method to use.
-p {translate,project}, --projection {translate,project}
Projection method to use.
-c {concatenate,average}, --combination {concatenate,average}
Combination method to use.
TBD
[1] https://vossanto.weltliteratur.net/emnlp-ijcnlp2019/vossantos.html
[2] https://www.nytimes.com/2007/05/03/sports/othersports/03boxing.html