A Comprehensive Multimodal Argument Mining Toolkit.
MAMKit is an open-source, publicly available PyTorch toolkit designed to access and develop datasets, models, and benchmarks for Multimodal Argument Mining (MAM). It provides a flexible interface for accessing and integrating datasets, models, and preprocessing strategies through composition or custom definition. MAMKit is designed to be extendible, ensure replicability, and provide a shared interface as a common foundation for experimentation in the field.
At the time of writing, MAMKit offers 4 datasets, 4 tasks and 6 distinct model architectures, along with audio and text processing capabilities, organized in 5 main components.
| Datasets | Tasks |
|---|---|
| UkDebates | Argumentative Sentence Detection (ASD) |
| MArgγ | Argumentative Relation Classification (ARC) |
| MM-USED | Argumentative Sentence Detection (ASD) Argumentative Component Classification (ACC) |
| MM-USED-fallacy | Argumentative Fallacy Classification (AFC) |
| Model | Text Encoding | Audio Encoding | Fusion |
|---|---|---|---|
| BiLSTM | GloVe + BiLSTM | (Wav2Vec2 ∨ MFCCs) + BiLSTM | Conat-Late |
| MM-BERT | BERT | (Wav2Vec2 ∨ HuBERT ∨ WavLM) + BiLSTM | Concat-Late |
| MM-RoBERTa | RoBERTa | (Wav2Vec2 ∨ HuBERT ∨ WavLM) + BiLSTM | Concat-Late |
| CSA | BERT | (Wav2Vec2 ∨ HuBERT ∨ WavLM) + Transformer | Concat-Early |
| Ensemble | BERT | (Wav2Vec2 ∨ HuBERT ∨ WavLM) + Transformer | Avg-Late |
| Mul_TA | BERT | (Wav2Vec2 ∨ HuBERT ∨ WavLM) + Transformer | Cross |
Clone the repository and install the requirements:
git clone [email protected]:TBA_AFTER_ACCEPTANCE/mamkit.git
cd MAMKit
pip install -r requirements.txtMAMKit provides a modular interface for defining datasets or allowing users to load datasets from the literature.
In the example that follows, illustrates how to load a dataset.
In this case, a dataset is loaded using the MMUSED class from mamkit.data.datasets, which extends the Loader interface and implements specific functionalities for data loading and retrieval.
Users can specify task and input mode (text-only, audio-only, or text-audio) when loading the data, with options to use default splits or load splits from previous works. The example uses splits from Mancini et al. (2022).
The get_splits method of the loader returns data splits in the form of a data.datasets.SplitInfo. The latter wraps split-specific data, each implementing PyTorch's Dataset interface and compliant to the specified input modality (i.e., text-only).
from mamkit.data.datasets import UKDebates, InputMode
loader = UKDebates(
task_name='asd',
input_mode=InputMode.TEXT_ONLY,
base_data_path=base_data_path)
split_info = loader.get_splits('mancini-et-al-2022')The Loader interface also allows users to integrate methods defining custom splits as follows:
from mamkit.data.datasets import SplitInfo
def custom_splits(self) -> List[SplitInfo]:
train_df = self.data.iloc[:50]
val_df = self.data.iloc[50:100]
test_df = self.data.iloc[100:]
fold_info = self.build_info_from_splits(train_df=...)
return [fold_info]
loader.add_splits(method=custom_splits,
key='custom')
split_info = loader.get_splits('custom')To add a new dataset, users need to create a new class that extends the Loader interface and implements the required functionalities for data loading and retrieval.
The new class should be placed in the mamkit.data.datasets module.
The toolkit provides a modular interface for defining models, allowing users to compose models from pre-defined components or define custom models. In particular, MAMkit offers a simple method for both defining custom models and leveraging models from the literature.
The following example demonstrates how to instantiate a model with a configuration found in the literature.
This configuration is identified by a key, ConfigKey, containing all the defining information.
The key is used to fetch the precise configuration of the model from the configs package.
Subsequently, the model is retrieved from the models package and configured with the specific parameters outlined in the configuration.
from mamkit.configs.base import ConfigKey
from mamkit.configs.text import TransformerConfig
from mamkit.data.datasets import InputMode
config_key = ConfigKey(
dataset='mmused',
task_name='asd',
input_mode=InputMode.TEXT_ONLY,
tags={'mancini-et-al-2022'})
config = TransformerConfig.from_config(
key=config_key)
model = Transformer(
model_card=config.model_card,
dropout_rate=config.dropout_rate
...)The example below illustrates that defining a custom model is straightforward. It entails creating the model within the models package, specifically by extending either the AudioOnlyModel, TextOnlyModel, or TextAudioModel classes in the models.audio, models.text, or models.text_audio modules, respectively, depending on the input modality handled by the model.
class Transformer(TextOnlyModel):
def __init__(
self,
model_card,
head,
dropout_rate=0.0,
is_transformer_trainable: bool = False,
): ...from mamkit.models.text import Transformer
model = Transformer(
model_card='bert-base-uncased',
dropout_rate=0.1, ...)Our models are designed to be encapsulated into a PyTorch LightningModule, which can be trained using PyTorch Lightning's Trainer class.
The following example demonstrates how to wrap and train a model using PyTorch Lightning.
from mamkit.utility.model import to_lighting_model
import lightning
model = to_lighting_model(model=model,
num_classes=config.num_classes,
loss_function=...,
optimizer_class=...)
trainer = lightning.Trainer(max_epochs=100,
accelerator='gpu',
...)
trainer.fit(model,
train_dataloaders=train_dataloader,
val_dataloaders=val_dataloader)The mamkit.configs package simplifies reproducing literature results in a structured manner.
Upon loading the dataset, experiment-specific configurations can be easily retrieved via a configuration key.
This enables instantiating a processor using the same features processor employed in the experiment.
In the example below, we adopt a configuration akin to Mancini et al. (2022), employing a BiLSTM model with audio encoded with MFCCs features. Hence, we define a MFCCExtractor processor using configuration parameters.
from mamkit.configs.audio import BiLSTMMFCCsConfig
from mamkit.configs.base import ConfigKey
from mamkit.data.datasets import UKDebates, InputMode
from mamkit.data.processing import MFCCExtractor, UnimodalProcessor
from mamkit.models.audio import BiLSTM
loader = UKDebates(task_name='asd',
input_mode=InputMode.AUDIO_ONLY)
config = BiLSTMMFCCsConfig.from_config(
key=ConfigKey(dataset='ukdebates',
input_mode=InputMode.AUDIO_ONLY,
task_name='asd',
tags='mancini-et-al-2022'))
for split_info in loader.get_splits(
key='mancini-et-al-2022'):
processor =
UnimodalProcessor(
features_processor=MFCCExtractor(
mfccs=config.mfccs, ...))
split_info.train = processor(split_info.train)
...
model = BiLSTM(embedding_dim=
config.embedding_dim, ...)The toolkit is organized into five main components: configs, data, models, modules and utility.
In addition to that, the toolkit provides a demos directory for running all the experiments presented in the paper.
The figure below illustrates the toolkit's structure.
The documentation is available at TBA.
We welcome contributions to MAMKit. Please refer to the contributing guidelines for more information. (TBA)
For any questions or suggestions, please contact us at TBA.
If you use MAMKit in your research, please cite the following paper:
@inproceedings{TBAmamkit,
title={MAMKit: A Comprehensive Multimodal Argument Mining Toolkit},
author={TBA},
booktitle={TBA},
year={TBA}
}