DWReCo at CLEF-2023 News Subjectivity

This repository contains our approach for the shared task: CLEF 2023 Task 2 subjectivity detection. We evaluate our method on the datasets in English and Turkish.

You can find more details about the task and the dataset at the link.

You can check the sample distibution via python -m app.analysis.stats

Installation

We use docker to build the environment and the libraries.

docker build -t news_subjectivity:latest .

Create .env file and add your OPENAI key in .env file:

OPENAI_API_KEY=HEREISYOURAPIKEY

How to run

To augment the dataset, modify the script scripts/augment.sh and then execute the following code:

docker run --rm --gpus 0 -v /reco/news_subjectivity:/app --name app news_subjectivity:latest bash scripts/augment.sh

Augmentation Styles

To change the paraphrase style, you need to change --style in scripts/augment.sh.

normal: it paraphrases the text without any style. It is a baseline augmentation method. When this style is on, it only paraphrases the minority class which is subjective in the datasets.

Reference

If you find this work useful, please cite the following information:

@InProceedings{clef-checkthat:2023:task2:dwreco, 
	author = "Schlicht, Ipek Baris and Khellaf, Lynn and Altiok, Defne", 
	title = "DWReCO at {CheckThat}! 2023: Enhancing Subjectivity Detection through Style-based Data Sampling", 
	crossref = "clef2023-workingnotes"
}