University of Engineering and Technology,

Taxila

Natural Language Processing (SE-345)

NLP-Project-Module-1+2+3

Instructor: Dr. Kanwal Yousaf

Project Group:
Muhammad Amad Ahmad (22-SE-10)
Muhammad Ammar Zahid (22-SE-12)
Zeeshan Ahmad (22-SE-15)
Sayab Arshad (22-SE-24)
Muhammad Zain (22-SE-32)

Project: FairWrite– AI-powered bias detection

Date: May 16, 2025

📑Table of Contents
🎯Project Goal....................................................................................................3
🧭Introduction...................................................................................................3
Broad Context and Motivation......................................................................3
Research Gap...............................................................................................4
Problem Statement.......................................................................................4
Objectives, Scope, and Contribution.............................................................4
📚Literature Review...........................................................................................5
Bias in AI and NLP.........................................................................................5
Bias Mitigation Techniques...........................................................................5
Pre-Processing Techniques.........................................................................5
In-Processing Techniques...........................................................................5
Post-Processing Techniques.......................................................................6
Existing Bias Detection and Mitigation Frameworks.....................................6
🧪Methodology..................................................................................................7
Vector Embedding Implementation..............................................................7
Data Collection..........................................................................................7
Data Preprocessing....................................................................................7
🧠Embeddings using BERT..........................................................................8
🧠Embeddings using GloVe.........................................................................9
⚙️Applying Neural Networks.......................................................................10
Bert Embeddings >> Simple NN (binary classification)..........................10
BERT Embeddings >> Transformer Based Model (multiclass
classification)...........................................................................................12
BERT Embeddings >> BiLSTM model......................................................14
🤖LLM Implementation....................................................................................17
LLM Evaluation Results (Gemini 2.0 Flash).................................................17
📈Results.........................................................................................................19
1. Binary Classification (Biased, Unbiased).................................................19
📊Binary Classification Report...................................................................20
2. Multiclass Classification (Biased, Unbiased, Opinion).............................21
 a) BiLSTM Model......................................................................................21
📊Classification Report (BiLSTM):..............................................................21
b) Transformer-Inspired Dense Model......................................................21
📊Classification Report (Transformer Model):............................................23
🧩Comparative Analysis..................................................................................23
📖References...................................................................................................26

🎯Project Goal
The goal of this project is to develop a specialized AI-driven bias detection and mitigation
system that enhances fairness in textual content. The system will identify biased words and
phrases, classify text based on bias presence, and suggest neutral alternatives. While existing
solutions like Dbias provide a research framework, our approach aims to fine-tune and extend
bias detection beyond news articles to general text, ensuring applicability in broader domains
such as blogs, opinion pieces, and everyday writing. This research will contribute to the field of
NLP by demonstrating how AI can be utilized for ethical and unbiased text processing.

🧭Introduction
Broad Context and Motivation
Bias in textual content, whether in news articles, blogs, or everyday writing, can reinforce
harmful stereotypes, misinform the public, and lead to unfair decision-making. With the
increasing reliance on AI-driven content generation and analysis, ensuring fairness in text has
become crucial. While large language models (LLMs) like GPT-4, GPT-3, llama3.4, phi4,
mistral, gemma, and deepseek-r1 aim for neutrality, they still reflect biases present in their
training data. There is a growing need for specialized AI models that not only detect but actively
mitigate biases in text, ensuring fairness and inclusiveness in various forms of written
communication.

Research Gap
While Dbias provides a promising framework for bias detection and mitigation, it remains
primarily a research tool rather than a widely accessible application for public use. Original
paper says, “We make this package (Dbias) as publicly available for the developers and
practitioners to mitigate biases in textual data (such as news articles), as well as to encourage
extension of this work.” Most existing llms are trained on vast datasets that may contain biases
and thus are susceptible to generating biased content. Furthermore, there is a lack of domain-
adaptive bias mitigation tools that cater to a broader audience beyond journalistic content, such
as bloggers, academic writers, and general users. A practical, web-based, or API-driven solution
that integrates bias detection and debiasing into daily writing tools is still missing.

Problem Statement
This research aims to develop an AI-driven bias detection and mitigation system that is
accessible beyond the research community. Unlike generic LLMs that generate text with
potential biases, our model will focus specifically on identifying biased words and phrases,
masking them, and suggesting neutral alternatives. By leveraging transformer-based models and
fine-tuning on diverse textual data beyond news articles, we seek to create an adaptive system
that ensures fairness in different forms of writing while maintaining semantic integrity.

Objectives, Scope, and Contribution

The objective of this research is to design and implement a web-based or API-integrated bias
mitigation tool that extends the capabilities of existing frameworks like Dbias. This system will:
1. Detect biases in a variety of text formats.
2. Recognize and highlight biased phrases.
3. Provide unbiased alternative wordings.
4. Integrate into commonly used writing platforms. (not intended for semester project)
The scope of this work extends beyond news articles to general writing, ensuring that fairness
principles apply to blogs, opinion pieces, and everyday user-generated content. Our contribution
lies in making bias mitigation tools more practical, accessible, and adaptable to real-world
applications.

📚Literature Review
Bias in AI and NLP
With the increasing reliance on AI in sensitive domains such as journalism, social media, hiring,
and legal decision-making, concerns over bias and fairness in NLP models have gained
significant attention. Studies have shown that machine learning models trained on large text
corpora often inherit societal biases, leading to discriminatory outputs. For example, the
COMPAS algorithm exhibited racial biases by incorrectly predicting a higher likelihood of
recidivism for African American defendants. Similarly, biased AI systems have been observed in
hiring algorithms, targeted advertising, and medical recommendations. These issues highlight the
need for bias detection and mitigation techniques to ensure fairness in AI-driven decision-
making.
Bias in NLP is often categorized into individual fairness (ensuring similar individuals receive
similar predictions) and group fairness (ensuring unbiased outcomes across demographic
groups). However, achieving fairness in NLP remains challenging due to the complexity of
language and the subjective nature of bias. Therefore, researchers have proposed various
algorithmic techniques to detect, analyze, and mitigate bias at various stages of the NLP
pipeline.

Bias Mitigation Techniques

Bias mitigation in NLP models is typically approached using pre-processing, in-processing,
and post-processing methods.

Pre-Processing Techniques
Pre-processing techniques focus on modifying the training data to reduce bias before model
training. Reweighing adjusts the weights of training samples based on their group attributes to
ensure fairness across different demographics. The Learning Fair Representations (LFR)
method encodes data into a transformed representation that minimizes the effect of protected
attributes such as gender or race. Similarly, the Disparate Impact Remover modifies feature
values in a dataset while preserving rank ordering to improve group fairness. These methods help
reduce bias at the data level, but they do not address biases introduced during model training or
prediction.

In-Processing Techniques
In-processing techniques adjust the learning process of machine learning models to enforce
fairness constraints. Prejudice Remover applies a fairness-aware regularization term during
training to reduce discrimination in model outputs. Adversarial De-biasing employs an
adversarial framework where a secondary model attempts to predict a sensitive attribute (e.g.,
gender, race) while the primary model learns to minimize this prediction, thereby ensuring
fairness. Another notable approach is Exponentiated Gradient Reduction, which iteratively
adjusts classification decisions to optimize fairness metrics while minimizing performance loss.
These techniques improve fairness within the model, but they require modification of model
architectures, making them less accessible for pre-trained NLP models.
Post-Processing Techniques
Post-processing methods adjust model predictions after inference to improve fairness.
Equalized Odds Post-Processing modifies the output labels of a classifier using linear
programming to equalize false positive and false negative rates across demographic groups.
Calibrated Equalized Odds builds on this by optimizing score outputs to ensure unbiased
decision thresholds. Reject Option Classification selectively modifies predictions in uncertain
regions to favor disadvantaged groups. These methods are useful when retraining models is not
feasible, but they do not prevent biases from being learned during training.

Existing Bias Detection and Mitigation Frameworks

A number of toolkits have been developed to evaluate and mitigate bias in AI models. AI
Fairness 360 (AIF360) is an open-source library that provides fairness metrics and bias
mitigation algorithms for different ML pipelines. FairML uses influence functions to quantify
the contribution of input features to model predictions, helping researchers detect biased
decision-making patterns. FairTest applies statistical testing to discover biases in machine
learning models based on protected attributes. These toolkits offer valuable resources for
researchers but are often designed for technical users rather than the public.
Among domain-specific solutions, Dbias is a recent framework developed to mitigate biases in
news articles using a pipeline of three Transformer-based models:
(1) a DistilBERT classifier for bias detection.
(2) a RoBERTa-based Named Entity Recognition (NER) model for bias recognition.
(3) a Masked Language Model (MLM) for bias mitigation.
While Dbias provides an effective research-oriented solution, it remains a framework rather
than a widely accessible application for everyday writers. This highlights the need for a more
user-friendly system that integrates bias detection and mitigation into general writing
workflows, including blog posts, academic writing, and personal communication.

🧪Methodology
Vector Embedding Implementation
Data Collection
We used MBIC – A Media Bias Annotation Dataset. MBIC is the first available dataset about
media bias reporting detailed information on annotator characteristics and their individual
background. The first sample of statements represents various media bias instances. We are
following the research paper on Dbias for comparative analysis, so we used the single dataset.
Data Example:
Sentence: “The increasingly bitter dispute between American women’s national soccer team and
the U.S. Soccer Federation spilled onto the field Wednesday night when players wore their
warm-up jerseys inside outing a protest before their 3-1 victory over Japan.”
Outlet: MSNBC
Topic: sport
Type: left
Label_Bias: Non-biased
Label_Opinion: Entirely factual
Biased_Words: ['bitter']

Data Preprocessing
Took the important columns for embeddings and model training ['sentence', 'topic', 'type',
'biased_words4', 'Label_bias', 'Label_opinion'], Dropped the under representative label “No
Agreement”, and then worked the problem as both:
a) Binary Classification
b) Multiclass Classification
Although the original paper mentioned the work with two class Biased and Unbiased. But we
also did create a third label “Opinion” based on the following case:
If Label_bias == ‘Biased’ and Label_opinion == ‘Writer’s Opinion’ we label the final label as
“Opinion”, this balances the three classes perfectly.
For both the sentence data was lowercased and ridden of any punctuation. We One Hot Encoded
the 2 labels in case (a) and 3 labels in case (b). Also, we encoded the 'topic' and 'type'
columns.
The data was split into train and validation sets based on class equalization i.e. Using stratified
split.

🧠Embeddings using BERT

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

model_bert = BertModel.from_pretrained('bert-base-uncased')

def get_bert_embeddings(text):

try:

inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True,

max_length=512)

with torch.no_grad():

outputs = model_bert(**inputs)

return outputs.last_hidden_state.mean(dim=1).squeeze().numpy()

except Exception as e:

print(f"Error processing: {text} → {e}")

return np.zeros(768)

Tokenization
1. Input text is first normalized (lowercased for 'uncased' models)
2. The tokenizer splits text into WordPiece tokens (subword units)
3. Special tokens are added: [CLS] at beginning and [SEP] at end
4. Tokens are truncated/padded to max_length (512 for BERT)
Token to ID Conversion:
1. Each token is mapped to its corresponding ID from BERT's vocabulary (30,522 words)
2. Attention masks are created (1 for real tokens, 0 for padding)
Embedding Layers:
1. Token Embeddings: Each token ID is mapped to a 768-dim vector (for base model)
2. Position Embeddings: Learnable vectors encoding absolute position (0-511)
3. Segment Embeddings: For sentence-pair tasks (single sentence uses 0)
Transformer Encoder:
1. l2 layers (for base model) of multi-head self-attention
 2. Each layer applies:
a) Multi-head attention (12 heads for base model)
b) Layer normalization
c) Feed-forward network
d) Residual connections
Output Processing:
1. The last hidden state contains contextual embeddings for each token
2. Mean pooling across sequence dimension creates sentence embedding
3. Final output is a 768-dim vector representing the input text

🧠Embeddings using GloVe

# Load GloVe embeddings (assumes glove.6B.100d.txt is already
downloaded)
def load_glove_embeddings(glove_file_path):
embeddings_index = {}
with open(glove_file_path, encoding='utf8') as f:
for line in f:
values = line.split()
word = values[0]
coeffs = np.asarray(values[1:], dtype='float32')
embeddings_index[word] = coeffs
return embeddings_index
# Generate GloVe embedding for a sentence
def get_glove_embedding(text, glove_embeddings, dim=100):
words = clean_text(text)
valid_vectors = [glove_embeddings[word] for word in words if word in
glove_embeddings]
if valid_vectors:
return np.mean(valid_vectors, axis=0)
else:
return np.zeros(dim)
Tokenization:
1. Input text is first lowercased.

2. Punctuation is removed using regex.

3. Text is tokenized into individual words using word_tokenize() from

NLTK.

Embedding Lookup:
1. Pretrained GloVe vectors (e.g., glove.6B.100d) are loaded into a
dictionary.

2. Each word is matched to its corresponding 100-dimensional vector.

3. If a word is not found in GloVe vocabulary, it is skipped.

Sentence Embedding Construction:

1. All found word vectors in the sentence are averaged.

2. This produces a fixed-size 100-dimensional vector for the input

sentence.

3. If no words are found in the GloVe vocabulary, a zero vector is

returned.

Key Notes:
 GloVe is a static word embedding method: it doesn't consider context.

 Faster and lighter than BERT but lacks dynamic contextual

understanding.

 Best for models where interpretability or efficiency is prioritized.

⚙️Applying Neural Networks

Bert Embeddings >> Simple NN (binary classification)

model = Sequential([

Dense(256, activation='relu', input_dim=768, kernel_regularizer='l2'),

BatchNormalization(),

Dropout(0.6), # Increased from 0.5

 Dense(128, activation='relu', kernel_regularizer='l2'),

BatchNormalization(),

Dropout(0.5),

Dense(1, activation='sigmoid')

])

Batch size: 16, Epochs: 20, loss=BinaryCrossentropy(label_smoothing=0.1) with adjusting

learning rate lr_scheduler = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=3)

Results
Performance When Saved:
Validation Accuracy: 0.8006
Training Accuracy: ~0.7382
The original Dbias has about 78% G-AUC score. Threshold used here for binary classification is
0.5.
Unbiased Class
The economy of the United Kingdom grew by 2% last year.
Prediction Score: 0.0077 → Unbiased
Scientists discovered a new exoplanet orbiting a nearby star.
Prediction Score: 0.1834 → Unbiased
Biased Class
Those people don't belong here.

Prediction Score: 0.6101 → Biased

Only the elite benefit from the current system.

Prediction Score: 0.6692 → Biased

BERT Embeddings >> Transformer Based Model (multiclass classification)

model = Sequential([
Dense(512, kernel_regularizer=l2(0.001), input_shape=(X_train_final.shape[1],)),
BatchNormalization(),
 tf.keras.layers.Activation('relu'),
Dropout(0.5),

Dense(256, kernel_regularizer=l2(0.001)),
BatchNormalization(),
tf.keras.layers.Activation('relu'),
Dropout(0.4),

Dense(128, kernel_regularizer=l2(0.001)),
BatchNormalization(),
tf.keras.layers.Activation('relu'),
Dropout(0.3),

Dense(y_train_final.shape[1], activation='softmax')
])
Batch size: 16, Epochs: 50, loss='categorical_crossentropy’ with adjusting learning rate
lr_scheduler = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=3) we achieved
similar results but the highest accuracy one didn’t had this as tuning parameter.
Results
Confusion Matrix

G-AUC Score
(macro): 0.8327
We got better G-AUC score than the original Dbias paper mentions of 78%.
BERT Embeddings >> BiLSTM model
model = Sequential([
# Reshape layer to convert 2D input to 3D for LSTM
Reshape((input_dim, 1), input_shape=(input_dim,)),

# Bidirectional LSTM layer

Bidirectional(LSTM(lstm_units, return_sequences=True)),
Dropout(dropout_rate_lstm),
BatchNormalization(),

# Second Bidirectional LSTM layer for deeper feature extraction

Bidirectional(LSTM(64, return_sequences=False)),
Dropout(dropout_rate_lstm),

# Dense hidden layer with ReLU activation

Dense(dense_units, activation='relu'),
Dropout(dropout_rate_dense),
BatchNormalization(),

# Output layer with softmax activation for multi-class classification

Dense(num_classes, activation='softmax')
 ])

# Model configuration parameters

input_dim = 785 # Feature dimension from BERT embeddings
num_classes = 3 # Number of bias classes
lstm_units = 128 # Size of LSTM units (increased for better feature learning)
dense_units = 96 # Size of dense layer
learning_rate = 5e-5 # Lower learning rate for more stable training
dropout_rate_lstm = 0.4 # Dropout after LSTM
dropout_rate_dense = 0.2 # Dropout after dense layer
batch_size = 64 # Larger batch size for more stable gradients
epochs = 50 # Maximum number of epochs
patience = 10 # Increased patience for early stopping

Results
For some reason this model performed the worst with 38% accuracy.
It may have been trained on potentially altered data. But we may further work on improving it.
G-AUC SCORE
(macro): 0.5578
🤖LLM Implementation
To evaluate the capabilities of large language models (LLMs) in detecting bias in text, we
integrated Google’s Gemini 2.0 Flash model via the google-genai API. This model was
prompted with a system instruction to classify input sentences into one of three categories:
Biased, Unbiased, or Opinion.
We used the validation set from our multiclass classification pipeline and constructed prompts
that included the sentence, along with its topic and source type (left, right, center) as
contextual metadata. The model was instructed to return only the final label. Example prompt
structure:

system_prompt = (

"You are a helpful assistant that classifies text as one of the following: "

"1. Biased, 2. Unbiased, or 3. Opinion. "

"Return only the label (e.g., Biased, Unbiased, or Opinion)."

full_prompt = f"{system_prompt}\n\nText: {sentence}\nTopic: {topic}\nType: {ttype}"

LLM Evaluation Results (Gemini 2.0 Flash)

To benchmark LLM performance, we evaluated 15 validation samples using Google’s Gemini
Flash model. The LLM was prompted with contextual metadata (sentence, topic, type) to classify
inputs as Biased, Unbiased, or Opinion.
Performance Summary:
Metric Score

Accuracy 53.33%

Precision 65.28%

Recall 51.85%

F1-Score 51.09%
 1. The model showed high precision for Opinion (1.00) but low recall (0.33), indicating it
was selective but confident.
2. Unbiased texts had the lowest precision (0.33), but a recall of 0.67, suggesting it
overpredicted this class in some borderline cases.
3. Biased class performance was balanced across all metrics.
Limitations:
1. The model could only respond to ~15/50 samples, likely due to API rate limits, input
formatting, or silent response truncation.
2. LLMs like Gemini require more structured input or fine-tuning to handle nuanced
classification reliably.
📈Results
We implemented and evaluated multiple models for bias classification: a binary classifier and
two multiclass models (BiLSTM and Transformer-based). Preprocessing included sentence-level
text cleaning, tokenization using BERT, and metadata encoding (topic, type, biased word count).
For multiclass classification, additional encoded metadata features were used as inputs.

1. Binary Classification (Biased, Unbiased)

Component Details

Input Features Sentence only (BERT embeddings)

Model Architecture Dense (MLP) with dropout & batch norm

Train/Val Split 80/20 random (balanced)

Accuracy 80.06%
📊Binary Classification Report
Class Precision Recall F1-Score Support

Biased 0.84 0.87 0.85 208

Non-biased 0.71 0.67 0.69 103

Accuracy 0.80 311

Macro Avg 0.78 0.77 0.77 311

Weighted Avg 0.80 0.80 0.80 311

G-AUC Score (binary): 0.8440

The binary model effectively distinguished between "Biased" and "Unbiased" classes using only
sentence embeddings. Minimal overfitting was observed, and the model generalized well on
validation data.

2. Multiclass Classification (Biased, Unbiased, Opinion)

a) BiLSTM Model
Component Details

Input Shape 785 (768 BERT + 17 metadata features)

Layers 2 BiLSTM layers, 1 Dense layer

Dropout/B-Norm Dropout (0.4/0.2), Batch Normalization

Optimizer/Loss Adam (5e-5), categorical crossentropy

Accuracy 38%

G-AUC (macro) 0.5578

📊Classification Report (BiLSTM):

Class Precision Recall F1-Score Support

Unbiased 0.00 0.00 0.00 106

Biased 0.41 0.37 0.39 98

Opinion 0.38 0.79 0.51 107

Macro Avg 0.26 0.38 0.30 311

Weighted Avg 0.26 0.39 0.30 311

b) Transformer-Inspired Dense Model

Component Details

Input Shape 785

Layers Dense (512→256→128) with L2 regularization

Dropout/B-Norm Dropout (0.5/0.4/0.3), Batch Normalization

Component Details

Optimizer/Loss Adam (1e-4), categorical crossentropy

Accuracy 67%

G-AUC (macro) 0.8327

📊Classification Report (Transformer Model):

Class Precision Recall F1-Score Support

Unbiased 0.62 0.60 0.61 106

Biased 0.69 0.73 0.71 98

Class Precision Recall F1-Score Support

Opinion 0.71 0.68 0.70 107

Macro Avg 0.67 0.67 0.67 311

Weighted Avg 0.67 0.67 0.67 311

🧩Comparative Analysis

1. Binary vs Multiclass: Binary classification was simpler and highly effective using only
BERT embeddings. In contrast, the multiclass setup required richer input and more
complex learning.
2. BiLSTM Limitations: The model failed to capture generalizable features, especially for
the “Unbiased” class (recall: 0%). Despite deeper sequential layers, it underperformed
across most metrics.
3. Transformer Performance: The transformer-inspired dense network significantly
outperformed BiLSTM in all metrics. Its architecture better leveraged metadata and
embeddings, maintaining generalization with regularization.

4. LLM Baseline (Gemini Flash):

1. Applied to 15 real validation sentences.
2. Accuracy: 53.3%, F1 Score: 0.51
3. While LLMs can generate plausible labels, they suffer from inconsistent outputs
and fail to match specialized model accuracy.
5. Comparison with Literature: Our transformer model exceeded the G-AUC score (0.84
vs. 0.78) reported in the Dbias paper (Raza et al., 2022), validating the effectiveness of
combining BERT with light transformer layers and metadata.
📖References
1. Nielsen, A.: Practical fairness. O’Reilly Media, Sebastopol (2020).
2. Bellamy, R.K.E., et al.: AI Fairness 360: an extensible toolkit for detecting and mitigating
algorithmic bias. IBM J. Res. Dev. 63(4–5), 401–415 (2019).
3. Orphanou, K., et al.: "Mitigating Bias in Algorithmic Systems—A Fish-Eye View." ACM
Comput. Surv., 2021.
4. Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., Galstyan, A.: "A survey on bias and
fairness in machine learning." ACM Comput. Surv. 54(6), 1–35 (2021).
5. Narayanan, A.: "Fairness Definitions and Their Politics." In: Tutorial presented at the
Conf. on Fairness, Accountability, and Transparency, 2018.
6. Kamiran, F., Calders, T.: "Data preprocessing techniques for classification without
discrimination." Knowl. Inf. Syst. 33(1), 1–33 (2012).
7. Feldman, M., Friedler, S. A., Moeller, J., Scheidegger, C., Venkatasubramanian, S.:
"Certifying and removing disparate impact." In: Proceedings of the 21st ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, 2015, pp. 259–268.
8. Zemel, R., Wu, Y., Swersky, K., Pitassi, T., Dwork, C.: "Learning fair representations."
In: International Conference on Machine Learning, 2013, pp. 325–333.
9. Calmon, F. P., Wei, D., Vinzamuri, B., Ramamurthy, K. N., Varshney, K. R.: "Optimized
pre-processing for discrimination prevention." Adv. Neural Inf. Process. Syst., vol. 2017-
Decem, no. Nips, pp. 3993–4002, 2017.
10. Kamishima, T., Akaho, S., Asoh, H., Sakuma, J.: "Fairness-aware classifier with
prejudice remover regularizer." In: Lecture Notes in Computer Science (including
subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics),
pp. 35–50. Springer, New Jersey (2012).
11. Celis, L. E., Huang, L., Keswani, V., Vishnoi, N. K.: "Classification with fairness
constraints: A meta-algorithm with provable guarantees." In: Proceedings of the
Conference on Fairness, Accountability, and Transparency, 2019, pp. 319–328.
12. Zhang, B. H., Lemoine, B., Mitchell, M.: "Mitigating unwanted biases with adversarial
learning." In: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society,
2018, pp. 335–340.
13. Agarwal, A., Beygelzimer, A., Dudík, M., Langford, J., Wallach, H.: "A reductions
approach to fair classification." In: International Conference on Machine Learning, 2018,
pp. 60–69.
14. Kamiran, F., Karim, A., Zhang, X.: "Decision theory for discrimination-aware
classification." In: 2012 IEEE 12th International Conference on Data Mining, 2012, pp.
924–929.
15. Hardt, M., Price, E., Srebro, N.: "Equality of opportunity in supervised learning." Adv.
Neural Inf. Process. Syst. 29, 3315–3323 (2016).
16. Pleiss, G., Raghavan, M., Wu, F., Kleinberg, J., Weinberger, K. Q.: "On fairness and
calibration." arXiv Prepr. arXiv1709.02012, 2017.
17. Udeshi, S., Arora, P., Chattopadhyay, S.: "Automated directed fairness testing." In:
Proceedings of the 33rd ACM/IEEE International Conference on Automated Software
Engineering, 2018, pp. 98–108.
18. Adebayo, J. A., et al.: "FairML: Toolbox for diagnosing bias in predictive modeling."
Massachusetts Institute of Technology, 2016.
19. Tramèr, F., et al.: "FairTest: Discovering Unwarranted Associations in Data-Driven
Applications." Proc. 2nd IEEE Eur. Symp. Secur. Privacy, EuroS&P 2017, pp. 401–416,
2017.
20. Bantilan, N.: "Themis-ml: a fairness-aware machine learning interface for end-to-end
discrimination discovery and mitigation." J. Technol. Hum. Serv. 36(1), 15–30 (2018).
21. Mehrabi, N., Gowda, T., Morstatter, F., Peng, N., Galstyan, A.: "Man is to person as
woman is to location: Measuring gender bias in named entity recognition." Proc. 31st
ACM Conf. Hypertext Soc. Media, HT 2020, pp. 231–232, 2020.
22. Tan, C., Sun, F., Kong, T., Zhang, W., Yang, C., Liu, C.: "A survey on deep transfer
learning." In: International Conference on Artificial Neural Networks, 2018, pp. 270–
279.
23. Devlin, J., Chang, M. W., Lee, K., Toutanova, K.: "BERT: Pre-training of deep
bidirectional transformers for language understanding." arXiv Prepr. arXiv1810.04805,
2018.
24. Li, B., et al.: "Detecting Gender Bias in Transformer-based Models: A Case Study on
BERT." 1, 2021.
25. Sinha, M., Dasgupta, T.: "Determining Subjective Bias in Text through Linguistically
Informed Transformer-based Multi-Task Network." In: Proceedings of the 30th ACM
International Conference on Information & Knowledge Management, 2021, pp. 3418–
3422.
26. Nadeau, D., Sekine, S.: "A survey of named entity recognition and classification."
Lingvisticae Investig. 30(1), 3–26 (2007).
27. Excell, E., Al Moubayed, N.: "Towards Equal Gender Representation in the Annotations
of Toxic Language Detection," 2021.
28. Mishra, S., He, S., Belli, L.: "Assessing Demographic Bias in Named Entity
Recognition," 2020.
29. Kaneko, M., Bollegala, D.: "Unmasking the Mask—Evaluating Social Biases in Masked
Language Models." arXiv Prepr. arXiv2104.07496, 2021.