Multi Modal Discussion Transformer: Integrating Text, Images and Graph Transformers to Detect Hate Speech on Social Media (AAAI 2024)

This repository contains the source code for the mDT Architecture, code to evaluate text-only baselines and how to create the HatefulDiscussions dataset. Our code base is based on the Graphormer and FairSeq repositories, with many modifications. This repository is also intended to be a living repository for ongoing follow-up work based on this research (mDT-Experimental). The dataset can be found here: https://vault.cs.uwaterloo.ca/s/5F9CsKMA2kmL3Wz

Abstract: We present the Multi-Modal Discussion Transformer (mDT), a novel method for detecting hate speech on online social networks such as Reddit discussions. In contrast to traditional comment-only methods, our approach to labelling a comment as hate speech involves a holistic analysis of text and images grounded in the discussion context. This is done by leveraging graph transformers to capture the contextual relationships in the discussion surrounding a comment and grounding the interwoven fusion layers that combine text and image embeddings instead of processing modalities separately. To evaluate our work, we present a new dataset, HatefulDiscussions, comprising complete multi-modal discussions from multiple online communities on Reddit. We compare the performance of our model to baselines that only process individual comments and conduct extensive ablation studies. ([Paper])

Key advantages of mDT:

• Contextual Discussion Embeddings: Our architecture integrates multi-modal models with graph transformers, allowing multi-modal comment representations to be actively and contextually grounded in the discussion context.
• Easily extendable to other modalities: Since we include signals from other modalities through bottleneck tokens, including more modalities through additional bottleneck tokens is trivial. Furthermore, bottleneck tokens mean that modalities are optional, where missing modalities are not given a bottleneck token.
• Updatable Comment Predictions: As the discussion unfolds and more comments are added, mDT can use that context to re-evaluate previous predictions based on earlier comments. This allows for more accurate predictions by leveraging the community's reaction as evidence.
• Extendable Implementation: This code base is built on FairSeq, which allows for native highly distributed training, and Huggingface Transformers, which make it easy to swap out text and image models for other domain-specific variants and bring in other modalities.

Organization

The code base is organized into three sections:

• Pre-Processing contains the code to create the HatefulDiscussions dataset from scratch.
• Comment Only Experiments contains the code to process the hateful discussions dataset and evaluate individual comment-only baselines
• mDT contains the source code for the Multi-Modal Discussion Transformer

mDT is organized as follows:

• Experiments: Contains folders for each dataset using this architecture, including launch scripts and data loading. Currently, this includes HatefulDiscussions, as well as WIP future work in a private nested submodule. Future work on other datasets should extend this folder
• src/criterions: Contains and registers each loss function which can be used
• src/data: Contains data loading utilities to collate and batch data into training examples.
• src/models: Contains model configurations and hyper-parameters. Model configurations registered are used in the launch script
• src/modules: Contains multi-purpose modules that make up models. These modules contain the actual model logic.
• src/tasks: Contains logic for building data loaders in src/data and loading pre-trained checkpoints. Tasks requiring advanced logic for loading pre-trained checkpoints should extend these files.

Citation

If you use our work or our dataset, we would greatly appreciated it if you could cite our paper here:

Hebert, Liam, et al. "Multi-modal discussion transformer: Integrating text, images and graph transformers to detect hate speech on social media." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 20. 2024.

@inproceedings{hebert2024multi,
  title={Multi-modal discussion transformer: Integrating text, images and graph transformers to detect hate speech on social media},
  author={Hebert, Liam and Sahu, Gaurav and Guo, Yuxuan and Sreenivas, Nanda Kishore and Golab, Lukasz and Cohen, Robin},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={20},
  pages={22096--22104},
  year={2024}
}