Multi-Author Writing Style Analysis - PAN 2025

This project is developed for the PAN 2025 competition, focusing on the task of multi-author writing style analysis. The goal is to detect changes in authorship within a given text by analyzing stylistic features and using machine learning models to predict where these changes occur.

Overview

The system processes input texts, extracts a variety of linguistic and stylistic features, and applies pre-trained models to infer authorship changes. It supports different difficulty levels (easy, medium, hard) with tailored models and configurations for each.

Project Structure

• src/: Contains the core source code for the project.
  • preprocessing.py: Handles text preprocessing tasks such as tokenization, lemmatization, and stopword removal.
  • features.py: Extracts various features from text, including lexical, syntactic, and semantic features.
  • pipeline_pan25.py: Defines the main pipeline for processing texts and extracting features for the PAN 2025 task.
  • inference.py: Runs inference using pre-trained models to predict authorship changes.
  • test.py: Manages the testing process across different problem types.
• models/: Stores pre-trained machine learning models and scalers for different difficulty levels.
• main.py: Entry point for running the analysis on input data.
• requirements.txt: Lists the dependencies required for the project.
• Dockerfile: Configuration for containerizing the application.
• Makefile: Contains build and run commands for the project.
• eval.sh and run_test.sh: Scripts for evaluation and testing.

Features

• Text Preprocessing: Customizable preprocessing with options for handling punctuation, stopwords, lemmatization, and POS tagging.
• Feature Extraction: Comprehensive feature set including:
  • Lexical and syntactic features
  • Semantic features using SBERT embeddings
  • Contextual and deep style features
• Difficulty Levels: Separate models and configurations for easy, medium, and hard problem sets.
• Inference: Predicts authorship changes using gradient boosting models with tailored thresholds.

Installation

1. Clone this repository to your local machine.

2. Ensure you have Python 3.8+ installed.

3. Install the required dependencies:

   pip install -r requirements.txt

Usage

To run the analysis on a dataset:

python main.py -i <input_directory> -o <output_directory>

• <input_directory>: Path to the directory containing the input text files, organized by difficulty level (easy, medium, hard).
• <output_directory>: Path where the output predictions will be saved as JSON files.

The input directory should have subdirectories for each difficulty level (e.g., input/easy, input/medium, input/hard), containing text files named problem-<id>.txt and corresponding truth files truth-<id>.json if available.

Output Format

For each processed problem, a JSON file named solution-<id>.json will be created in the output directory under the respective difficulty folder. The JSON file contains a list of binary values indicating predicted authorship changes between consecutive sentences.

Example:

{
    "changes": [0, 1, 0, ...]
}

Models

The project uses pre-trained gradient boosting models for each difficulty level:

• Easy: grad_boost_C3_easy.pkl with configuration C3
• Medium: grad_boost_C14_medium.pkl with configuration C14
• Hard: grad_boost_C9_hard.pkl with configuration C9

Additionally, MinMax scalers are applied for easy and medium levels to normalize features before inference.