Skip to content

Latest commit

 

History

History
78 lines (49 loc) · 3.07 KB

Readme_DI.md

File metadata and controls

78 lines (49 loc) · 3.07 KB

Dependencies

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 or from the original repository: https://github.com/mbatchkarov/DiscoUtils (both should be fine, if in doubt the forked version by tttthomasssss should work)

Installation

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 .

Resources

Vectors from the EMNLP 2016 paper Improving Sparse Word Representations with Distributional Inference for Semantic Composition are in the subfolder resources/vectors.

The folder contains 5 files:

  • APT_ML2010_AN.dill.xz: Composed AN vectors (using 1000 neighbours), keys are prefixed with __CV_.
  • APT_ML2010_NN.dill.xz: Composed NN vectors (using 10 neighbours), keys are prefixed with __CV_
  • APT_ML2010_VO.dill.xz: Composed VO vectors (using 50 neighbours), keys are prefixed with __CV_
  • APT_wikipedia_clean_shift-10.dill.xz: Raw APT vectors, with an SPPMI shift of log(10).
  • APT_wikipedia_clean_shift-40.dill.xz: Raw APT vectors, with an SPPMI shift of log(40).

All files have been compressed with xz. All files can be read with the supplied code by using vector_utils.load_vector_cache() (see below).

Usage

Loading vectors:

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`

Composing Vectors:

from apt_toolkit.composition import mozart
from apt_toolkit.utils import vector_utils

# Load Vectors
vectors = vector_utils.load_vector_cache('path/to/vectors', filetype='dill')

noun_vector = vectors['quantity']
adj_vector = vectors['large']

# Suppose we want to compose the AN pair "large quantity" with a an "amod" relation between the adjective and the noun, we first need to offset the adjective by "amod" to align the feature spaces (doing so makes the adjective more look like a noun; the offset needs to be "amod" as the adjective has an inverse "amod" relation to its head noun)
offset_vector = vector_utils.create_offset_vector(adj_vector, 'amod')

# Now the two vectors can be composed
composed_vector = mozart.intersect_apts(offset_vector, noun_vector)

Distributional Inference:

from apt_toolkit.distributional_inference import distributional_inference
from apt_toolkit.utils import vector_utils

# Load Vectors
vectors = vector_utils.load_vector_cache('path/to/vectors', filetype='dill')

# Use the top 20 neighbours (using the `static top n` approach) to infer unobserved co-occurrence features for the noun "book"
rich_book = distributional_inference.static_top_n(vectors=vectors, words=['book'], num_neighbours=20)