Audio Transcription Classification

This Python project processes audio transcriptions and classifies them for aggression and hate speech categories using various language models.

Features

• Binary Classification: Determines if text is aggressive (0 = not aggressive, 1 = aggressive).
• Multiclass Classification: Categorizes text into one of five classes:
  • 0: Racism (attacks based on race, nationality, or religion)
  • 1: Sexism (attacks directed at women, gender roles, objectification)
  • 2: Hate Speech (general hate speech not fitting other categories)
  • 3: Vulgarism (vulgar language without targeted attacks)
  • 4: Neutral (non-aggressive language)
• Supports multiple model providers: Local (Ollama), OpenAI, and Google Gemini.
• Processes input files from the input/ directory and outputs results to output/.
• Handles errors gracefully (e.g., unreachable models, invalid responses).

Requirements

• Python 3.7+
• Required packages (install via pip install -r requirements.txt):
  • litellm
  • requests
  • openai (for OpenAI provider)
  • google-genai (for Gemini provider)
• For local models: Ollama server running with supported models (e.g., llama3.3, mistral-large).
• API keys for OpenAI and Gemini if using those providers.

Installation

1. Clone the repository:

   git clone <repository-url>
   cd Classification
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Set up API keys (if using OpenAI or Gemini):

   • Edit the script files to set OPENAI_API_KEY and GEMINI_API_KEY, or set environment variables.

4. For local models, ensure Ollama is installed and running:

   • Install Ollama from ollama.ai.
   • Pull required models: ollama pull llama3.3, etc.
   • Set SELF_HOSTED_MODELS_URL to your Ollama API base (e.g., http://localhost:11434).

Usage

1. Place input files in the input/ directory. Each file should contain lines in the format: filename;text (semicolon-separated).

2. Run the classification script:

   • For one-shot classification: python classification-oneshot.py
   • For zero-shot classification: python classification-zeroshot.py

3. Results will be saved in the output/ directory as CSV files, one per model and input file.

Output Format

Output files are CSV with semicolon-separated values:

• filename: Original filename
• binary_aggressive: 0 or 1 (or error codes)
• multiclass_label: 0-4 (or error codes)
• text: Sanitized input text

Supported Models

• Local: llama3.3, mistral-large, SpeakLeash/bielik-11b-v2.3-instruct:Q8_0
• OpenAI: gpt-5 (adjust as needed)
• Gemini: gemini-2.5-flash