The file size of the APT representations used in the paper is too large to host them on GitHub.
Please email the first author of the paper (either using the email address on this GitHub repository or the email address in the paper) if you would like to have them.
The code relies on several 3rd party libraries:
- numpy
- scipy
- scikit-learn
- dill
- joblib
- nltk
- sqlitedict
In addition the following code (which has its own dependencies) is necessary for performing distributional inference:
- DiscoUtils:
https://github.com/tttthomasssss/DiscoUtils
Apart from DiscoUtils which needs to be installed manually, all requirements as well as the codebase itself can be installed with:
cd path/to/apt-toolkit
pip install -e .
Vectors from the paper Improving Semantic Composition with Offset Inference will be posted on github - unfortunately they were too large for the ACL submission system.
from apt_toolkit.utils import vector_utils
vectors = vector_utils.load_vector_cache('path/to/vectors', filetype='dill') # Loads the higher-order dependency-typed vectors as a `dict` of `dicts`
from apt_toolkit.composition import mozart
from apt_toolkit.distributional_inference import distributional_inference
from apt_toolkit.utils import vector_utils
from discoutils.thesaurus_loader import Vectors
'''
Loads a standard set of vectorised APTs, creates some offsets and adds them back to the lexicon
This is one way - and perhaps the simplest way - of doing creating neighbours for offsets
'''
# Load Vectors
vectors = vector_utils.load_vector_cache('path/to/vectors', filetype='dill')
noun_vector = vectors['quantity']
adj_vector = vectors['large']
# Offset and add the resulting representation back to the lexicon
offset_vector_adj = vector_utils.create_offset_vector(adj_vector, 'amod')
vectors['__OFFSET_amod_large'] = offset_vector_adj
# Can also offset the noun vector or any other vectors and add them back to the lexicon
offset_vector_noun = vector_utils.create_offset_vector(noun_vector, '!amod')
vectors['__OFFSET_!amod_quantity'] = offset_vector_noun
# Store vectors back to file
vector_utils.save_vector_cache(vectors, 'path/to/offsets', filetype='dill')
from apt_toolkit.distributional_inference import distributional_inference
from apt_toolkit.utils import vector_utils
# Load Vectors with previously added offset representations
vectors = vector_utils.load_vector_cache('path/to/offsets', filetype='dill')
adj_vector = vectors['exciting']
# Use offset Inference to enrich the noun-view of the adjective "exciting"
exciting_offset = vector_utils.create_offset_vector(adj_vector, 'amod')
rich_exciting_offset = distributional_inference.static_top_n(vectors=vectors, words=['__OFFSET_amod_exciting'], num_neighbours=20)
# Use the standard DI algorithm to infer unobserved co-occurrence features for the noun "book"
rich_book = distributional_inference.static_top_n(vectors=vectors, words=['book'], num_neighbours=20)
# Now the two vectors can be composed
composed_vector = mozart.intersect_apts(rich_exciting_offset, rich_book)