Model Card for identrics/wasper_propaganda_classifier_bg

Model Description

• Developed by: Identrics
• Language: Bulgarian
• License: apache-2.0
• Finetuned from model: INSAIT-Institute/BgGPT-7B-Instruct-v0.2
• Context window : 8192 tokens

Model Description

This model consists of a fine-tuned version of BgGPT-7B-Instruct-v0.2 for a propaganda detection task. It is effectively a multilabel classifier, determining wether a given propaganda text in Bulgarian contains or not 5 predefined propaganda types. This model was created by Identrics, in the scope of the WASPer project. The detailed taxonomy could be found here.

Propaganda taxonomy

The propaganda techniques we want to identify are classified in 5 categories:

1. Self-Identification Techniques: These techniques exploit the audience's feelings of association (or desire to be associated) with a larger group. They suggest that the audience should feel united, motivated, or threatened by the same factors that unite, motivate, or threaten that group.

2. Defamation Techniques: These techniques represent direct or indirect attacks against an entity's reputation and worth.

3. Legitimisation Techniques: These techniques attempt to prove and legitimise the propagandist's statements by using arguments that cannot be falsified because they are based on moral values or personal experiences.

4. Logical Fallacies: These techniques appeal to the audience's reason and masquerade as objective and factual arguments, but in reality, they exploit distractions and flawed logic.

5. Rhetorical Devices: These techniques seek to influence the audience and control the conversation by using linguistic methods.

Uses

To be used as a multilabel classifier to identify if the Bulgarian sample text contains one or more of the five propaganda techniques mentioned above.

Example

First install direct dependencies:

pip install transformers torch accelerate

Then the model can be downloaded and used for inference:

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

labels = [
    "Legitimisation Techniques",
    "Rhetorical Devices",
    "Logical Fallacies",
    "Self-Identification Techniques",
    "Defamation Techniques",
]

model = AutoModelForSequenceClassification.from_pretrained(
    "identrics/wasper_propaganda_classifier_bg", num_labels=5
)
tokenizer = AutoTokenizer.from_pretrained("identrics/wasper_propaganda_classifier_bg")

text = "Газа евтин, американското ядрено гориво евтино, пълно с фотоволтаици а пък тока с 30% нагоре. Защо ?"

inputs = tokenizer(text, padding=True, truncation=True, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits

probabilities = torch.sigmoid(logits).cpu().numpy().flatten()

# Format predictions
predictions = {labels[i]: probabilities[i] for i in range(len(labels))}
print(predictions)

Training Details

During the training stage, the objective was to develop the multi-label classifier to identify different types of propaganda using a dataset containing both real and artificially generated samples.

The data has been carefully annotated by domain experts based on a predefined taxonomy, which covers five primary categories. Some examples are assigned to a single category, while others are classified into multiple categories, reflecting the nuanced nature of propaganda where multiple techniques can be found within a single text.

The model reached an F1-weighted score of 0.538 during training.

Compute Infrastructure

This model was fine-tuned using a GPU / 2xNVIDIA Tesla V100 32GB.

Citation [this section is to be updated soon]

If you find our work useful, please consider citing WASPer:

@article{...2024wasper,
  title={WASPer: Propaganda Detection in Bulgarian and English}, 
  author={....},
  journal={arXiv preprint arXiv:...},
  year={2024}
}