GACaps-HTC: Graph Attention Capsule Network for Hierarchical Text Classification

This is the official implementation of [GACaps-HTC: Graph Attention Capsule Network for Hierarchical Text Classification].

1. Configuration

run.py takes the path to a configuration file as its argument. (e.g. python run.py configs/cfg.json)

2. Data

• config.path.data.train, config.path.data.val, and config.path.data.test

  • These files are .json files where each line is a JSON object composed as follows:

    {"text": "this is an example document", "label": ["<study_unit_label_name>", "<parent_topic_label_name>", ancester topics]}

  • The labels are required to be sorted from the leaf node to the node just below the root.

    e.g. ["Vector, Matrix, and Tensor", "Basics of Linear Algebra", "Linear Algebra", "AI Prerequisites"]

• config.path.labels

  • This is a json file containing a dictionary where label names and label indices are keys and values, respectively.

• config.path.prior

  • This is a json file containing a dictionary of the following format:

    {"<parent_label_name>": {"<child_1_label_name>": <prior 1>, "<child_2_label_name>": <prior 2> ...} ...}

    e.g. {"AI Prerequisites": {"Calculus": 0.4, "Linear Algebra": 0.3, "Probability": 0.2, "Statistics": 0.1} ...}

  • For each parent node - child node pair, prior is simply obtained as the proportion of the number of instances.

• config.path.hierarchy

  • This is a tsv file where each line is written as follows:

    <parent_label_name>\t<child_1_label_name>\t<child_2_label_name>...

3. Requirements

run.py requires the following libraries: numpy, torch, transformers.