LikelihoodClassificationOfBiases

A likelihood-based statistical tool specifically designed to identify and classify various types of biased statements or comments related to different categories such as gender or religion within a given dataset.

Table of Contents

• Abstract
• Features
• Installation
• Usage
• Datasets
• Contributing

Abstract

Detecting biased or discriminatory language in so- cial media spaces towards protected attributes (such as gen- der, ethnicity or religion) is a laborious and subjective task, often necessitating the involvement of human moderators or computationally intensive neural networks. In this paper, we provide a likelihood-based statistical approach to identify and classify biased statements or comments in online communication platforms. Our model offers a fast solution that can aid in efficient content moderation, text analysis and annotation. Using word embeddings and statistical analysis, we can figure out comments or sentences that may contain certain biases that we may wish to set apart. For instance, identifying whether a sentence revolves around male, female, or perhaps none of them. To test the performance of our model and validate the classification pro- cess, experiments were performed on various publicly available datasets, including the Multi-Dimensional Gender Bias dataset

Features

• Identify Biases: Likelihood-based statistical approach specifically designed to identify and classify various types of biased statements or comments related to different categories such as gender or religion in language.
• Annotation: Can be used for content moderation, text analysis and annotation.
• Controversial Biases: Identify biased or discriminatory language in so- cial media spaces towards protected attributes (such as gender, ethnicity or religion.

Installation

Follow these steps to install and set up and use our model to anaylze text on your local machine:

1. Step 1: Install all the prerequisites using the following command

$ pip install requirements.txt

2. Step 2: Choose and install a pre-trained word embedding model from popular options like Bert, Sbert, spaCy, or gensim. (You can also train your own word-embeddings)
3. You can use the example notebooks, such as example1.ipynb to analyze your own data with the sbert_model, Please feel free to interact with the two experimentation notebooks located in the main directory. You have the flexibility to choose any word embedding model that you prefer for your experiments. By exploring and comparing the results obtained from different word embeddings.

Usage

By utilizing the model accessible through the src folder, you have the capability to detect textual bias within any dataset, while also analyzing the disparities between any two input target parameters. This functionality empowers you to identify and assess potential biases in the textual data, leading to a deeper understanding of how certain factors can influence the information present in the dataset.

# Load the dataset
import pandas as pd
data = pd.read_csv("path/to/data")

You can use the following piece of code to annotate your initial dataset to classify comments in thre different categories, namely $c_1$ biased, $c_2$ biased or neutral.

from src.wordbiases import CalculateWordBias
from src.model import LikelihoodModelForNormalDist

# Set two different target sets consitsting of words you wish to target
target_set_1 = ["abuse", "bad", "hate", "worst"]
target_set_2   = ["love", "good", "peace", "calm"]

# Select a list of Focus Words (i.e words that may contribute to the biases in any given comment)
F = ["ADJ", "NOUN", "PRON"]

wbcalc = CalculateWordBias(target_set_1, target_set_2, F, computing_device="cuda")
wbcalc.process_documents(data, "text")
c1, c2 = wbcalc.calculate_target_embeddings()

wbiases, _ , biased_words = wbcalc.calculate_biases()
total_pop = [b for _, b in biased_words]

mu = np.mean(total_pop)
sigma = np.std(total_pop)

likelihood_clf = LikelihoodModelForNormalDist(0.05, 0.95, threshold_limit=0)
likelihood_clf.fit_total_pop(total_pop)

# predictions
preds = likelihood_clf.predict(wbiases)[0]
data["prclass"] = preds

Datasets

Information regarding the datasets used in the experiments

Dataset	Targeted Bias	Annotated	Total comments	$c_1$-related	$c_2$-related	Neutral comments
Twitter (social_bias_frames)	Racism/Antisemitism	Yes	30,430	3124	1143	26,163
Reddit (one-million-reddit-jokes)	Racism	No	400,000	16,753	19,119	364,128
McDonald's Store Reviews (US)	Sentiment	No	33,396	6,705	2263	24428
ImageChat (md_gender_bias)	Gender-Identity	Yes	329,908	21,251	62,383	225,168
News Classification	News Categories	No	100,000	12,871	12,871	74,270

Contributing

We welcome contributions from the community! If you'd like to contribute to the project, please follow these guidelines:

1. Fork the repository. Click on the "Fork" button on the top right corner of this repository to create your own copy.

2. Create a new branch. In your forked repository, create a new branch to work on your changes. You can name the branch according to the feature or bug fix you are implementing.

3. Make your changes and commit them. Make the necessary changes and improvements to the code in your branch. Commit your changes with clear and descriptive commit messages.

4. Push the changes to your fork. Once you are satisfied with your changes, push the changes to the branch in your forked repository.

5. Submit a pull request to the original repository. Finally, navigate to the original repository and click on the "New Pull Request" button. Select your branch and describe the changes you have made. Submit the pull request, and we will review your contribution.

Thank you for your interest in contributing to our project! We appreciate your support in making this project better. If you have any questions or need assistance, don't hesitate to reach out to us through the Issues section or the pull request itself.