Create README.md

Model Details

Model Name: distilbert_fearspeech_classifier

Model Version: 1.0

Paper: "You are doomed!" Crisis-specific and Dynamic Use of Fear Speech in Protest and Extremist Radical Social Movements

Authors: Simon Greipl, Julian Hohner, Heidi Schulze, Patrick Schwabl, Diana Rieger

Institution: Ludwig-Maximilians-University of Munich, Germany

Contact: simon.greipl@ifkw.lmu.de

Date: 2024-04-26
Overview

This model is a DistilBERT-based classifier trained to detect fear speech (FS) in German language Telegram posts. The classifier was developed to study the prevalence and dynamics of FS in the communication of radical and extremist actors on Telegram.
Intended Use

Primary Use Case: Detection and analysis of fear speech in social media posts, particularly those from radical and extremist groups.

Primary Users: Researchers in social media analysis, hate speech detection, and radicalization studies.

Languages: German

Domains: Social media, Telegram, radical and extremist communication
Model Description

Model Architecture: DistilBERT

Training Data:

The model was fine-tuned on a dataset of manually annotated Telegram posts from radical and extremist groups.
The dataset consists of posts related to six crisis-specific topics: COVID-19, Conspiracy Narratives, Russian Invasion of Ukraine (RioU), Energy Crisis, Inflation, and Migration.
The training dataset includes 2,460 posts, equally split between fear speech (FS) and non-fear speech (no FS).

Performance:

Validation Set Macro F1 Score: 0.82
Test Set Macro F1 Score: 0.79
Detailed performance metrics for the classifier are provided in the paper.

Evaluation

Metrics:

Precision, Recall, F1-Score for both FS and no FS classes.

Validation and Test Set Performance:

Validation Set Precision: 0.82
Validation Set Recall: 0.82
Validation Set F1-Score: 0.82
Test Set Precision: 0.79
Test Set Recall: 0.79
Test Set F1-Score: 0.79

Ethical Considerations

Bias and Fairness: The model was trained on data from specific Telegram channels and groups known for their extremist content. While efforts were made to ensure balanced representation, inherent biases in the data may affect the model's predictions.
Misuse Potential: The classifier is intended for research purposes. Misuse of the model to label and penalize individuals or groups without context or understanding of the nuances in their communication could lead to unjust outcomes.

Training Data

Source:

Data collected from German language Telegram channels and groups associated with far-right, COVID-19 protest, and conspiracy-focused actors.

Annotation Process:

A nominal six-scale annotation scheme was used to capture and distinguish the varying degrees of FS manifestations.
The scheme was compressed into a dichotomous scale for training, where 0 indicates no FS and 1 indicates FS.

How to Use

The trained classifier can be accessed and used via the Hugging Face model hub:

Model URL: distilbert_fearspeech_classifier

Example Usage:

```
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("PatrickSchwabl/distilbert_fearspeech_classifier")
model = AutoModelForSequenceClassification.from_pretrained("PatrickSchwabl/distilbert_fearspeech_classifier")

# Tokenize input text
inputs = tokenizer("Your text here", return_tensors="pt")

# Get model predictions
outputs = model(**inputs)
predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)

# Print predictions
print(predictions)


```

README.md

@@ -0,0 +1,15 @@
+---
+language:
+- de
+metrics:
+- f1
+- precision
+- recall
+pipeline_tag: text-classification
+tags:
+- fearspeech
+- classification
+- social science
+- communication
+- hatespeech
+---