/zero_shot_classification

Implementation of the paper "Train Once, Test Anywhere: Zero-Shot Learning for Text Classification"

Primary LanguageJupyter Notebook

W2V-based Zero-shot Text Classification

zero-shot-architecture

Introduction

Zero-shot learning aims to create models capable of predicting unseen classes (i.e. classes not observed during the training phase).

This prototypic module is based on the notion of word embeddings as well as following the paper:

"Train Once, Test Anywhere: Zero-Shot Learning for Text Classification".

Throughout the documentation, I will refer to each piece of headline news in text column as a document and each keyword as a concept

A Jyputer notebook (W2V-based Zero-shot Text Classification.ipynb) was created to better illusrate the following steps:

  • Exploratory Data Analysis (EDA)
  • Computing Doc2vec
  • Cluster Analysis
  • Generating Document-Concept Vectors
  • Training DNN

In particular, the following are the primary modules in this demo:

  • data_pipeline.py:

    • EDA
    • Generating document vectors for clustering
    • Generating document-concept vectors for classificaiton

    The main entry point is at run_data_pipeline(), which generates all the required training data Set test_mode to True for a quick run with a small sample size

  • model.py:

    • DNN
    • Test functions for validating the preditability of the document-concept vectors using other classical ML methods
    • Test functions for the proposed DNN with other data sets

    The main entry point is at run_modeling_pipeline()

  • cluster.py:

    • clustering algorithms (e.g. HDBSCN, k-means, spectral clustering)
    • cluster evaluations (e.g. silhouette scores)

    See demo_clustering() for an example cluster analysis

Other supportive modules:

  • evaluate: Evaluate the output of a classifier
  • cluster_utils: utility functions for clustering
  • learn_manifold: non-linear embeddings for high-dimensional data e.g. T-SNE
  • sampling_utils: utility functions for sampling
  • utils: misc

For more details, please refer to the notebook W2V-based Zero-shot Text Classification.ipynb

Installation

  1. Unzip zero_shot.zip and cd to zero_shot containing all source codes
  2. Make a copy of the source data (news_data.csv) and keep it under zero_shot/data
  • the source data is not included in this zip file to save space

  1. Also download Google news W2V here and keep it directly under the project directory i.e. zero_shot/

  2. Run pip install -r requirements.txt

You may be seeing a longer list of dependent packages in requirement.txt ...

All modules were tested on a new Macbook Pro with the Apple M1 chip. Since the standard version of Tensorflow has not yet supported the M1 chip, the installation for Tensorflow is a bit more involved:

  1. Miniforge was installed to support Python on arm64 architecture.

  2. An environment.yml file (included) was used to create a specialized enviornment (zero_shot) for running Tensorflow

    E.g. conda env create --file=environment.yml --name=zero_shot

  3. Then an appropriate version of Tensorflow was pip-installed along with its dependent packages via the following command:

pip install --upgrade --force --no-dependencies https://github.com/apple/tensorflow_macos/releases/download/v0.1alpha3/tensorflow_macos-0.1a3-cp38-cp38-macosx_11_0_arm64.whl https://github.com/apple/tensorflow_macos/releases/download/v0.1alpha3/tensorflow_addons_macos-0.1a3-cp38-cp38-macosx_11_0_arm64.whl

  1. Additionally, make sure the following supporting packages are installed (but they should already be included in requirements.txt):

    • NLTK v3.6+
    • Gensim v4.0+
      • Also run pip install python-Levenshtein to avoid a warning
    • HDBSCAN
      • pip install HDBSCAN