Welcome to the FusionSent repository. FusionSent is an efficient few-shot learning model designed for multi-label classification of scientific documents with many classes.
💾 Data: https://huggingface.co/datasets/TimSchopf/arxiv_categories
Figure 1: The training process of FusionSent comprises three steps:
- Fine-tune two different sentence embedding models from the same Pre-trained Language Model (PLM), with parameters θ₁, θ₂ respectively.
- θ₁ is fine-tuned on pairs of training sentences using cosine similarity loss, and θ₂ is fine-tuned on pairs of training sentences and their corresponding label texts, using contrastive loss.
- Label texts can consist of simple label/class names or more extensive texts that semantically describe the meaning of a label/class.
- Merge parameter sets θ₁, θ₂ into θ₃ using Spherical Linear Interpolation (SLERP).
- Freeze θ₃ to embed the training sentences, which are then used as input features to train a classification head.
By fine-tuning sentence embedding models using contrastive learning, FusionSent achieves high performance even with limited labeled data. The model initially leverages two distinct sub-models: one, using regular contrastive learning with item pairs ('setfit'), and another using label embeddings with class-description pairs ('label_embedding'). These two models are then fused, via (spherical) linear intterpolation, to create the robost FusionSent model that excels in diverse classification tasks. For detailed insights into the model and its performance, please refer to our published paper.
FusionSent is integrated with the Hugging Face Hub and provides two main classes:
- FusionSentModel: This class encapsulates the dual fine-tuning process of the two sentence embedding models ('setfit, and 'label_embedding') and their fusion into a single model ('fusionsent'). It is the core model class for embedding sentences and performing classification tasks.
- FusionTrainer: Responsible for loading, cleaning, and preparing datasets for training and evaluation.
To install the fusionSent package, use pip:
pip install fusionsentfrom fusionsent.training_args import TrainingArguments
from fusionsent.modeling import FusionSentModel
from fusionsent.trainer import Trainer
from datasets import Dataset
# Example dataset objects with sentences belonging to classes: ["Computer Science", "Physics", "Philosophy"]
train_dataset = Dataset.from_dict({
"text": [
"Algorithms and data structures form the foundation of computer science.",
"Quantum mechanics explores the behavior of particles at subatomic scales.",
"The study of ethics is central to philosophical inquiry."
],
"label": [
[1, 0, 0], # Computer Science
[0, 1, 0], # Physics
[0, 0, 1] # Philosophy
],
"label_description": [
["Computer Science"],
["Physics"],
["Philosophy"]
]
})
eval_dataset = Dataset.from_dict({
"text": [
"Artificial intelligence is transforming the landscape of technology.",
"General relativity revolutionized our understanding of gravity.",
"Epistemology questions the nature and limits of human knowledge."
],
"label": [
[1, 0, 0], # Computer Science
[0, 1, 0], # Physics
[0, 0, 1] # Philosophy
],
"label_description": [
["Computer Science"],
["Physics"],
["Philosophy"]
]
})
# Load the model.
model_id = "malteos/scincl"
model = FusionSentModel._from_pretrained(model_id=model_id)
# Set training arguments.
training_args = TrainingArguments(
batch_sizes=(16, 1),
num_epochs=(1, 3),
sampling_strategies="undersampling"
)
# Initialize trainer.
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset
)
# Train the model.
trainer.train()
# Evaluate the model.
eval_scores = trainer.evaluate(
x_eval=eval_dataset["text"],
y_eval=eval_dataset["label"]
)
# Perform inference.
texts = [
"Computational complexity helps us understand the efficiency of algorithms.",
"Thermodynamics studies the relationships between heat, work, and energy.",
"Existentialism explores the freedom and responsibility of individual existence."
]
predictions = model.predict(texts)
print(predictions)For a more elaborate example, please refer to the Jupyter notebook of a Description-Embedding Experiment.
Dataset is available at: https://huggingface.co/datasets/TimSchopf/arxiv_categories
For evaluation of FusionSent, we constructed a dataset of scientific documents derived from arXiv metadata. The arXiv metadata provides information about more than 2 million scholarly articles published in arXiv from various scientific fields. We use this metadata to create a dataset of 203,961 titles and abstracts categorized into 130 different classes. To this end, we first perform a stratified downsampling of the metadata to only 10% of all articles while retaining the original class distribution. Afterward, articles assigned to categories occurring less than 100 times in the downsampled dataset are removed. To obtain the final dataset, we then perform a stratified train/validation/test split of the processed dataset in an 80:10:10 ratio. The number of examples in each set is shown below.
| Split | Number of Samples |
|---|---|
| Train | 163,168 |
| Validation | 20,396 |
| Test | 20,397 |
Each article in the resulting arXiv dataset is categorized into one or more distinct categories. The figure below shows the distribution of papers across the 130 categories of the dataset.
When citing our work in academic papers and theses, please use this BibTeX entry:
@inproceedings{schopf-etal-2024-efficient,
title = "Efficient Few-shot Learning for Multi-label Classification of Scientific Documents with Many Classes",
author = "Schopf, Tim and
Blatzheim, Alexander and
Machner, Nektarios and
Matthes, Florian",
editor = "Abbas, Mourad and
Freihat, Abed Alhakim",
booktitle = "Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024)",
month = oct,
year = "2024",
address = "Trento",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.icnlsp-1.21",
pages = "186--198",
}For additional details and advanced configurations, please refer to the original paper linked at the beginning of this document.