VISPool

Codebase for the paper VISPool: Enhancing Transformer Encoders with Vector Visibility Graph Neural Networks published in Findings of the Association for Computational Linguistics ACL 2024.

TL;DR

We propose a vector visibility-based dynamic graph construction method for text documents and a corresponding VISPool architecture that allows seamless integration of transformers and graph neural networks.

Abstract

The emergence of transformers has revolutionized natural language processing (NLP), as evidenced in various NLP tasks. While graph neural networks (GNNs) show recent promise in NLP, they are not standalone replacements for transformers. Rather, recent research explores combining transformers and GNNs. Existing GNN-based approaches rely on static graph construction methods requiring excessive text processing, and most of them are not scalable with the increasing document and word counts. We address these limitations by proposing a novel dynamic graph construction method for text documents based on vector visibility graphs (VVGs) generated from transformer output. Then, we introduce visibility pooler (VISPool), a scalable model architecture that seamlessly integrates VVG convolutional networks into transformer pipelines. We evaluate the proposed model on the General Language Understanding Evaluation (GLUE) benchmark datasets. VISPool outperforms the baselines with less trainable parameters, demonstrating the viability of the visibility-based graph construction method for enhancing transformers with GNNs.

Vector Visibility Graph Dependency

Vector visibility graph (VVG) is a novel approach that maps mutlivariate time series into a graph structure. For this project @tunakasif implemented vector-vis-graph for fast and parallel VVG generation. Due to its promising applications, it is implemented as a standalone library, which is used as an external dependency in this project. For details, please refer to its own GitHub page. It is also available on PyPI:

pip install vector-vis-graph

and Conda:

conda install -c conda-forge vector-vis-graph

Installation

As a Dependency

It is possible to install this project as a dependency from the GitHub repository, either with pip:

pip install git+https://github.com/koc-lab/vispool

or with Poetry:

poetry add git+https://github.com/koc-lab/vispool

Development or Altering Behavior

After cloning the repository, install the required dependencies either from pinned versions in the requirements.txt file:

pip install -r requirements.txt

or with Conda through environment.yml:

conda env create --name vispool --file=environment.yml

or with Poetry by running the following command in the root directory:

poetry update

Project Structure

The vispool/glue/ directory contains the baseline code for data loading, training and evaluating transformers on the GLUE benchmark datasets. Datasets and their corresponding metrics are obtained from the Hugging Face's datasets and evaluate modules. Base transformers are obtained from the Hugging Face's transformers library. To eliminate boilerplate code and to have reproducible experiments with auto accelerators, the lightning library is used for data loading and training. Finally, for logging and tracking, the wandb library is used.

The vispool/model/ directory contains the proposed VVG-based GCN architecture and the overall VISPool model that integrates the transformer and the VVG-based GCN. Similar structure to the baseline is preserved for easy integration and comparison.

Finally, vispool/baseline.py and vispool/our.py are used to define the hyperparameter tuning sweeps for the baseline and the proposed model, respectively. It allowed easy integration and agent attachment to the WandB sweeps. With click library, a CLI is implemented for easy usage of the sweeps and the agents in vispool/__main__.py.

Usage

Usage: python -m vispool [OPTIONS] COMMAND [ARGS]...

Options:
  --help  Show this message and exit.

Commands:
  baseline-sweep        Initialize a WandB sweep for fine-tuning a baseline transformer model
                        generated from MODEL_CHECKPOINT on a GLUE task with name TASK_NAME.
  baseline-agent        Attach an agent to the created baseline sweep with the given SWEEP_ID.
  vispool-sweep         Initialize a WandB sweep for fine-tuning a vispool model generated
                        from MODEL_CHECKPOINT on a GLUE task with name TASK_NAME.
  vispool-agent         Attach an agent to the created vispool sweep with the given SWEEP_ID.
  vispool-single-sweep  Initialize a WandB grid sweep for different seeds with the
                        hyperparameter values obtained in the given run with the specified
                        RUN_ID.
  vispool-single-agent  Attach an agent to the created vispool single sweep with the given SWEEP_ID

Example: Creating a Sweep for VISPool and Attaching an Agent

First create a sweep for the VISPool model for a GLUE benchmark dataset with a given TASK_NAME and an initial MODEL_CHECKPOINT which can be any path or name accepted by the Hugging Face's AutoModel class. For example, to create a sweep for the MRPC dataset with the distilbert-base-uncased checkpoint:

python -m vispool vispool-sweep "distilbert-base-uncased" "mrpc"

This will output the created SWEEP_ID:

Created sweep with id: a1b2c3d4

Then attach an agent to the created sweep with the given SWEEP_ID:

python -m vispool vispool-agent a1b2c3d4

Reference

If you are using ACL anthology .bib file, you can cite the paper with alikasifoglu-etal-2024-vispool entry key, where the full entry is provided below.

@inproceedings{alikasifoglu-etal-2024-vispool,
  title     = {{VISP}ool: Enhancing Transformer Encoders with Vector Visibility Graph Neural Networks},
  author    = {Alika{\c{s}}ifo{\u{g}}lu, Tuna  and Aras, Arda  and Koc, Aykut},
  editor    = {Ku, Lun-Wei  and Martins, Andre  and Srikumar, Vivek},
  booktitle = {Findings of the Association for Computational Linguistics ACL 2024},
  month     = aug,
  year      = {2024},
  address   = {Bangkok, Thailand and virtual meeting},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2024.findings-acl.149},
  pages     = {2547--2556}
}