Mod_HATE

Official implementation for: Modularized Networks for Few-shot Hateful Meme Detection

This includes an original implementation of "Modularized Networks for Few-shot Hateful Meme Detection" by Rui Cao, Roy Ka-Wei Lee, Jing Jiang.

This code provides:

• Codes for training LoRA modules for hate speech detection, meme comprehension and hateful meme interpretation.
• Composing trained LoRA modules and adapt composed modules to few-shot hateful meme detection.

Please leave issues for any questions about the paper or the code.

If you find our code or paper useful, please cite the paper:

@inproceedings{DBLP:conf/www/0002L024,
  author       = {Rui Cao and
                  Roy Ka{-}Wei Lee and
                  Jing Jiang},
  title        = {Modularized Networks for Few-shot Hateful Meme Detection},
  booktitle    = {Proceedings of the {ACM} on Web Conference 2024, {WWW} 2024, Singapore,
                  May 13-17, 2024},
  pages        = {4575--4584},
  publisher    = {{ACM}},
  year         = {2024}
}

Announcements

01/25/2023: We released our implementation for our WWW submission: Modularized Networks for Few-shot Hateful Meme Detection.

Content

1. Installation
2. Prepare Datasets
   • Datasets for Hateful Meme Detection
   • Datasets for Module Training
3. Training LoRA modules (Section 4.2 of the paper)
4. Modularized Networks for Hateful Meme Detection (Section 4.3 and Section 5 of the paper)
   • Step 1: Module Composition
   • Step 2: Experiments

Installation

The code is tested with python 3.9. To run the code, you should install the package of transformers provided by Huggingface (version 4.33.0), PyTorch library, PEFT Library (version 0.5.0), Nevergrad. The code is implemented with the CUDA of 11.2 (you can also implement with other compatible versions) on NIVIDA A40 GPU, each with a dedicated memory of 48GB. For the implementation of LLaMA model, we leverage the HuggingFace Library with the yahma/llama-7b-hf checkpoint.

Prepare Datasets

Datasets for Hateful Meme Detection

We have tested on two hateful meme benchmarks: Facebook Hateful Meme dataset (FHM) and Multimedia Automatic Misogyny Identification dataset (MAMI). The datasets can be download online. If you download dataset, you need to pre-process the datasets follow the code of the paper: Pro-Cap: Leveraging a Frozen Vision-Language Model for Hateful Meme Detection. Alternatively, you can directly leverage the converted data shared by Pro-Cap. Noted, they denote MAMI dataset as mimc.

Datasets for Module Training

We trained module capable of relevant tasks for hateful meme detection. Specifically, we focus on three relevant tasks: hate speech detection, meme comprehension and hateful meme interpretation. To train these modules, you need to first prepare data about these tasks. Below are datasets we considered for each task:

• Hate speech detection: we merge three hate speech detection dataset for training the module. Specifically, we leverage DT, WZ and Gab and consider hate speech detection as a generation task.
• Meme comprehension: we consider the MemeCap dataset. Given a meme, the task requires generation of its meaning. Beyond image captioning, the task also calls for recognizing and interpreting visual metaphors with respect to the text inside or around the meme.
• Hateful meme interpretation: we consider the HatReD dataset. Given a hateful meme, it requires generating the underlying hateful contextual reasons.

Alternatively, you can directly use our shared dataset in the data folder. hate-speech is the data for hate speech detection, meme-interp is for meme comprehension, and hateful-meme-explain is for explaining hateful memes.

Training LoRA modules

After preparing all data for relevant tasks, we train individual modules for each task. We leverage the parameter-efficient technique, low-rank adaptation (LoRA), to tune the large language model, LLaMA and regard the LoRA module as the module capable of each task. To obtain train the modules, please run the script in src/individual_mode.sh:

bash individual_mode.sh

Besides using the scripts for training LoRA modules yourself, you can also directly use our provided trained LoRA modules in the LoRA_modules folder. Do make sure the path of datasets is set properly (the path on your own matchine)!!!

Modularized Networks for Hateful Meme Detection

Based on the trained LoRA modules (LoRA_modules/hate-speech for hate speech detection, LoRA_modules/meme-captions for meme comprehension and LoRA_modules/hate-exp for hateful meme interpretation), we learn composed networks. Do make sure the path of datasets is set properly (the path on your own matchine)!!!

Step 1: Module Composition

The learning of the composition of modules is largely based on LoRAHub. We greatly appreciate the work of LoRAHub. By running src/lora_learning.py, we learn a module composer, assigning importance scores over each trained module.

Step 2: Experiments

Based on the learned importance scores, the modularized networks is learned by weighted averaging learnt modules. Then, we can test the few-shot capabilities regarding the hateful meme detection task. You can test with the script src/new_lora.sh by:

bash new_lora.sh

You can also see the logging file of our reported performance in the paper in the folder src/shot_4_LoRA and src/shot_8_LoRA.