TLA (Text-Label Alignment)

Implementation for the ECML-PKDD 2024 accepted paper "Modeling Text-Label Alignment for Hierarchical Text Classification" arxiv ecml

Requirements

• Python >= 3.6
• torch >= 1.6.0
• transformers >= 4.2.1
• fairseq == 0.10.0
• torch-geometric == 2.4.0
• torch-sparse == 0.6.17

Data

• All datasets are publically available and can be accessed at WOS, RCV1-V2 and NYT.
• We followed the specific details mentioned in the contrastive-htc repository to obtain and preprocess the original datasets (WOS, RCV1-V2, and NYT).
• After accessing the dataset, run the scripts in the folder preprocess for each dataset separately to obtain tokenized version of dataset and the related files. These will be added in the data/x folder where x is the name of dataset with possible choices as: wos, rcv and nyt.
• For reference we have added tokenized versions of the WOS dataset along with its related files in the data/wos folder. Similarly do for the other two datasets.

Train

The train.py can be used to train all the models by setting different arguments.

For BERT (does flat multi-label classification)

python train.py --name='ckp_bert' --batch 10 --data='wos' --graph 0

Some Important arguments:

• --name name of directory in which your model will be saved. For e.g. the above model will be saved in ./TLA/data/wos/ckp_bert
• --data name of dataset directory which contains your data and related files
• --graph whether to use graph encoder

For HTLA (BERT+GPTrans+TLA; does Hierarchical Text Classification)

python train.py --name='ckp_htla' --batch 10 --data='wos' --graph 1 --graph_type='GPTrans' --edge_dim 30 --tla 1 --tl_temp 0.07

Some Important arguments:

• --graph_type type of graph encoder. Possible choices are 'GCN, 'GAT', 'graphormer' and 'GPTrans'. HTLA uses GPTrans as the graph encoder
• --edge_dim edge feature size for GPTrans (We use 30 as edge feature size for each dataset )
• --tla whether Text-Label Alignment (TLA) Loss required or not. If set to 0, the model will be optimized only on BCE loss, which we refer to as BERT-GPTrans in the paper.
• --tl_temp Temperature value for the TLA loss (We use 0.07 as the temp. value for all datasets)
• The node feature is fixed as 768 to match the text feature size and is not included as run time argument

For multiple random runs

In train.py set the --seed=None for multiple random runs

Other arguments for TLA in train.py:

Arguments of train.py, namely --norm, --proj, and --hsize, are part of TLA but have not been used in this work and can be ignored.

Test

To run the trained model on test set run the script test.py
python test.py --name ckp_htla --data wos --extra _macro

Some Important arguments

• --name name of the directory which contains the saved checkpoint. The checkpoint is saved in ../TLA/data/wos/ when working with WOS dataset
• --data name of dataset directory which contains your data and related files
• --extra two checkpoints are kept based on macro-F1 and micro-F1 respectively. The possible choices are _macro and _micro to choose from the two checkpoints

Citation

@InProceedings{10.1007/978-3-031-70365-2_10,
author="Kumar, Ashish
and Toshniwal, Durga",
editor="Bifet, Albert
and Davis, Jesse
and Krilavi{\v{c}}ius, Tomas
and Kull, Meelis
and Ntoutsi, Eirini
and {\v{Z}}liobait{\.{e}}, Indr{\.{e}}",
title="Modeling Text-Label Alignment for Hierarchical Text Classification",
booktitle="Machine Learning and Knowledge Discovery in Databases. Research Track",
year="2024",
publisher="Springer Nature Switzerland",
address="Cham",
pages="163--179",
isbn="978-3-031-70365-2"
}