First, you should download and unzip the file http://aurelieherbelot.net/resources/data/ideallanguage.zip (38M zipped) to your data/ directory. This file contains a representation of the Visual Genome entities, attributes and relationships which we will use to produce a basic entity matrix. Should you want to re-generate this data, you can do so using the script in utils/parse.py (you will need the original Visual Genome json files -- see code for detail).
Gather all the stats we'll need for building the matrices:
cd utils; python3 stats.py
The following matrices will be produced: entity matrix, predicate matrix, probabilistic version of the predicate matrix. Run:
python3 extract.py [--att] [--rel] [--sit]
The flags are optional and process relations, attributes and situations in the Visual Genome, in addition to objects. Spaces are produced for the raw cooccurrence matrix, as well as its probabilistic version. In addition, PPMI and PCA-reduced versions are generated. If you're getting started, the combination of --att and --rel should give you a fairly rounded model.
All space files will be stored in the spaces/ directory. Formats are as follows:
The entity matrix
Each relation / attribute is shown, followed by the IDs of the entities that instantiate it. E.g.
has(-,jersey.n.03) 145052 145052 145052 ...
on(ice.n.01,-) 1564869 1564872 1073295 ...
in(-,mouth.n.01) 1066577 1571546 1580146 ...
on(visor.n.01,-) 1574593 1589180 307104 ...
on|top|of(-,food.n.01) 148294 337935 1113534 ...
on|top|of(sunglasses.n.01,-) 1580964 358974 1618779 ...
of(-,bathroom.n.01) 1543891 1546177 1075545 ...
has(ring.n.02,-) 156638 156638 156638 ...
Inverse entity matrix
For each entity, the predicates applicable to that entity are shown. E.g.
2258167 land.n.04
519773 chest_of_drawers.n.01 next|to(bag.n.01,-)
3265447 sweatshirt.n.01 white with(-,band.n.04)
175702 person.n.01
3506270 window.n.01 on|a(-,building.n.01)
943446 people.n.01 on(-,land.n.04) in(-,plaza.n.02)
2931124 man.n.01
3132144 hand.n.01 of(-,guy.n.01)
We'll also compute and store standard similarity and nearest neighbours information on the created space. Using the flags you picked for extract.py, run similarity.py. For instance:
python3 similarity.py --att --rel
All similarity files will be stored in the data/ directory.
The ext2vec model is an 'extensional' version of the famous Word2Vec model. It produces vectors at reduced dimensionality from a raw (extensional) co-occurrence matrix. In a nutshell, ext2vec re-uses the 'context prediction' task of the original skip-gram model, but porting it to a formal semantics setting. Specifically, the task consists in predicting whether a 'target' predicate (from the rows of the predicate matrix) and a 'context' predicate (from the columns of the matrix) have been seen together in the description of a unique, grounded entity.
Run ext2vec with:
python3 embed_model.py [--att] [--rel] [--sit]
(where the flags should be the same as those you used earlier with the extract.py script).
This will run on the predicate_matrix.dm file contained in the relevant folder of your data/ directory.
Spaces can be evaluated on the standard relatedness and similarity test sets MEN and SimLex-999. To do so, go to tests/MEN or tests/SimLex. Evaluation can be performed on count spaces (with or without PPMI and PCA), on the output space of ext2vec, or on external vectors for comparison. For instance:
python3 spearman.py ext2vec --att --rel
It is possible to choose the space that is evaluated by using the flags --att and --rel (to include attribute and relations dimensions). Adding the flag --ppmi will run on the PPMI version of the predicate matrix, similarly with the --pca flag. So for example:
python3 spearman.py count --rel --ppmi --pca
would run on a count space with relation dimensions, with PPMI weighting and PCA.
To run a comparison with FastText VG:
python3 spearman.py compare --file=data/MEN_fasttext_vg_desc_vecs.txt
(The data folder also contains the pretrained FastText and BERT vectors, issued from large corpora.)
The spaces can be inspected from the point of view of various aspects of semantic competence: a) the ability to refer; b) mastery of lexical relations; c) the ability to make graded semantic acceptability judgements with respect to 'normal use'. All code for this is to be found in the tests directory, under the relevant subfolder.
Run from the truth directory. To retrieve the extension of a particular constituent in the Visual Genome, run the following (with the appropriate flag):
python3 composition.py --att
and input a phrase, e.g. cute teddy bear.
The system will return a sorted list of predicates, indicating which proportion of instances of the phrase are associated with each one. For instance:
COUNTS FOR THIS READING:
bear.n 1.0
cute 1.0
brown 1.0
teddy 1.0
sitting 0.8
stuffed 0.8
nice 0.4
has(-,eye.n) 0.4
of(nose.n,-) 0.4
little 0.4
behind(curtain.n,-) 0.4
...
NB: the code involves a toy grammar linked to an interpretation function, and performs semantic space expansion / retraction as explained in the paper. For now, the grammar only implements bare NPs with adjectives and nouns (of any length). So in practice, you should run with the --att flag.
Run from the lexrel and compatibility directories.
We provide pretrained EVA models to reproduce the results reported in the paper. Models are contained in the pretrained directory and are available for the best scoring settings. As an example, to evaluate on the incompatibility dataset with EVA's ext2vec model, trained on the Visual Genome's relations, you might run:
python3 nn_compatibility_test.py --model=pretrained/eva/synrel/e106 --ext=../../spaces/synrel/ext2vec.dm
Note that in each pretrained directory, ten checkpoints are available, corresponding to the ten training runs averaged in the paper. The README in each directory provides the hyperparameters used for training.
First, prepare the data in the relevant data/ directory by running preprocess.py. Then, to train, do for instance:
python3 nn_lexrel.py --batch=700 --epochs=400 --hidden=300 --lr=0.001 --wdecay=0.001 --ext=data/models/lexrel_fasttext_vecs.txt --checkpoint=checkpoints/fasttext/check1
Run from the acceptability directory.
First, prepare the data in the data/ directory by running preprocess.py. Then, to train:
python3 nn_acceptability.py