TOXIC COMMENT ANALYSER

submitted to
Indian Institute of Information Technology, Kalyani
for partial fulfillment of degree of
Bachelor of
Technology In
Computer Science and Engineering

By
Ashish Kumar (Reg No. 678)
Amarjit Hore (Reg No. 669)
Roshan Kumar (Reg No. 729)

Under the supervision of

Dr. Anirban Lakshman

(Assistant Professor)

Department of Computer Science and Engineering

Indian Institute of Information Technology,
Kalyani Autumn Semester 2024-2025
 CERTIFICATE

This is to certify that project report entitled “TOXIC COMMENT ANALYSER”

being submitted by Ashish Kumar (Reg No. 524), Amarjit Hore (Reg No.
669), Roshan Kumar (Reg No.729), undergraduate students of the Indian
Institute of Information Technology Kalyani, West Bengal, 741235, India,
for the award of Bachelor of Technology in Computer Science and
Engineering, is an original research work carried by them under my
supervision and guidance.

The project has fulfilled all the requirements as per the regulations of the
Indian Institute of Information Technology Kalyani and in my opinion, has
reached the standards needed for submission. The work, techniques and
the results presented have not been submitted to any other university or
institute for the award of any other degree or diploma.

………………………..

(Dr. Anirban Lakshman)

Assistant Professor

Department of Computer Science and Engineering

Indian Institute of Information Technology Kalyani

Kalyani, W.B.-741235, India.

(i)
 DECLARATION

We hereby declare that the work being presented in this project entitled
TOXIC COMMENT ANALYSER, submitted to Indian Institute of Information
Technology Kalyani in partial fulfilment for the award of the degree of
Bachelor of Technology in Computer Science and Engineering during the
period from Aug 2023 to Nov 2023 under the supervision of Dr. Anirban
Lakshman, Indian Institute of Information Technology Kalyani, West Bengal
- 741235, India, does not contain any classified information.

Name of the Candidates: Ashish Kumar (Reg No. 678)

Roshan Kumar (Reg No. 729)

Amarjit Hore (Reg No. 669)

Name of the Department: Computer Science and Engineering

Institute Name: Indian Institute of Information Technology Kalyani

Date: 14/11/2023

(ii)
 ACKNOWLEDGEMENT

First of all, we would like to thank our guide, Dr. Anirban Lakshman, for his
encouragement, guidance and cooperation to carry out this project, and for
giving us an opportunity to work on this project and providing us with a
great environment to carry our work in ease. We also thank other
resources which we have mentioned in our references.

We would also like to express our genuine gratitude to the Department of

Computer science and Engineering, IIIT Kalyani for providing us the
platform and opportunity to work on this project and for helping us in
every possible way.

IIIT Kalyani Ashish Kumar (Reg No. 678)

Date: 14/11/2023 Amarjit Hore (Reg No. 537)

Roshan Kumar (Reg No. 540)

(iii)
 ABSTRACT
In the ever-expanding digital landscape, online platforms provide spaces
for diverse interactions, yet the prevalence of toxic comments poses
significant challenges to maintaining a healthy online environment. This
project addresses the critical issue of identifying and classifying toxic
comments using advanced machine learning techniques. The Toxic Comment
Classifier employs a state-of-the-art deep learning model, incorporating a
bidirectional LSTM network and embedding layers for effective feature
extraction from textual data.

The project adheres to the rigorous standards set by the Indian Institute of
Information Technology Kalyani, fulfilling all regulatory requirements. Under
the expert supervision and guidance of their mentor, the students
demonstrate a comprehensive understanding of natural language processing
and deep learning techniques.

The results showcase the efficacy of the Toxic Comment Classifier in

distinguishing toxic comments, with precision, recall, and accuracy metrics
reflecting the model's robust performance. The work presented in this
project is an original contribution, and the findings have not been previously
submitted for any other degree or diploma.

This research not only contributes to the field of natural language

processing but also addresses the pressing need for effective tools to
maintain a positive and inclusive online environment. The Toxic Comment
Classifier stands as a testament to the commitment of the students and the
academic community at the Indian Institute of Information Technology
Kalyani towards fostering responsible and respectful online communication.

(iv)
 CONTENTS
Contents

CHAPTER – 1.............................................................................................................................................................................................. 1
PROBLEM STATEMENT.............................................................................................................................................................. 1
OBJECTIVE........................................................................................................................................................................................ 2
CHAPTER – 2.............................................................................................................................................................................................. 1
LITERATURE SURVEY.................................................................................................................................................................. 1
CHAPTER – 3.............................................................................................................................................................................................. 3
CHAPTER – 4.............................................................................................................................................................................................. 6
ARCHITECTURE & USER INTERACTION FLOW...............................................................................................................6
CHAPTER – 5........................................................................................................................................................................................... 11
WORKING....................................................................................................................................................................................... 11
16.................................................................................................................................................................................................................. 12
CHAPTER – 6........................................................................................................................................................................................... 13
EVALUATION AND RESULTS................................................................................................................................................. 13
MODEL SAVING AND LOADING............................................................................................................................................ 13
INTEGRATION WITH GRADIO............................................................................................................................................... 14
CHAPTER – 7........................................................................................................................................................................................... 15
RESULTS.......................................................................................................................................................................................... 15
Non-Toxic Comment as input, Output shows toxic is false.................................................................................................16
CHAPTER – 8........................................................................................................................................................................................... 17
CONCLUSION................................................................................................................................................................................. 17
FUTURE SCOPE OF WORK.......................................................................................................................................................18
REFERENCES................................................................................................................................................................................. 21
FIGURES
CHAPTER – 1
PROBLEM STATEMENT
To build a prototype of online hate and abuse comment classifier which can
be used to classify hate and offensive comments so that it can be controlled
and restricted from spreading hatred and cyberbullying.
The aim is to develop a prototype for an online hate and abuse comment
classifier. This classifier will play a crucial role in identifying and
categorizing hate speech and offensive comments. The primary goal is to
enable effective control and restriction of the dissemination of such content,
thereby mitigating the spread of hatred and preventing instances of
cyberbullying. The development of this prototype underscores a
commitment to fostering a safer and more responsible online environment.
OBJECTIVE
❖ Automated Detection of Toxic Comments:

▪ Develop a machine learning model capable of automatically

detecting toxic comments within digital content.
▪ Implement natural language processing techniques to
analyze and understand the linguistic features associated
with toxicity.

❖ Multi-Class Categorization

▪ Enable the classifier to categorize toxic comments into

specific classes, such as "toxic," "severe toxic," "obscene,"
"threat," "insult," and "identity hate."
▪ Provide a nuanced classification system to better
understand and address different forms of toxic language.

❖ Model Robustness and Accuracy

▪ Train the model to achieve high accuracy in identifying and

categorizing toxic comments.
▪ Ensure the robustness of the classifier to effectively handle
variations in language and context.

❖ Scalability and Efficiency

▪ Design the Toxic Comment Classifier to scale efficiently,

allowing it to process a large volume of comments in real-
time.
▪ Optimize the model for computational efficiency without
 compromising accuracy.

❖ User-Friendly Integration

▪ Develop an interface or integration method that allows

users to easily incorporate the Toxic Comment Classifier
into various online platforms.
CHAPTER – 2

LITERATURE SURVEY

❖ Early Approaches

Early efforts in toxic comment classification primarily relied on

rule-based methods and keyword filtering. These methods, though
straightforward, often lacked the ability to capture the nuances and
context-dependent nature of toxic language. As a result, researchers
began to shift towards machine learning-based approaches.

❖ Machine Learning Approaches

Supervised Learning:
The majority of research in toxic comment classification has
adopted supervised learning techniques. Various algorithms,
including Support Vector Machines (SVM), Naive Bayes, and
decision trees, have been employed to train models on labelled
datasets. These models leverage features such as bag-of-words, TF-
IDF, and word embeddings to identify patterns associated with toxic
language.
Deep Learning:
The advent of deep learning has significantly impacted the field,
with recurrent neural networks (RNNs), long short-term memory
networks (LSTMs), and more recently, transformer-based models
such as BERT and GPT, achieving state-of-the-art performance.
These models excel in capturing contextual information and
semantic relationships, enabling them to effectively identify subtle
instances of toxicity.
❖ Challenges and Open Problems

Despite the progress made in toxic comment classification, several

challenges persist. The dynamic nature of online language, evolving
patterns of toxicity, and the existence of cultural and contextual
variations pose difficulties for model generalization. Additionally,
issues related to bias and fairness in models, especially those
trained on biased datasets, need careful consideration.
CHAPTER – 3

PROPOSED SYSTEM
In the dynamic realm of online communication, the unrestricted exchange
of ideas on digital platforms has empowered diverse voices. However, this
openness has also given rise to the persistent challenge of toxic comments,
which can undermine the constructive nature of online discussions.
Recognizing the gravity of this issue, our project, the Toxic Comment
Classifier, spearheaded by undergraduate students Ashish Kumar, Amarjit
Hore, and Roshan Kumar from the Department of Computer Science and
Engineering at the Indian Institute of Information Technology Kalyani,
delves into the intricate domain of natural language processing and machine
learning.

As we navigate through the intricacies of the Toxic Comment Classifier, we

not only showcase the technical prowess of our students but also emphasize
the importance of responsible and respectful online communication. The
outcomes of this project stand not only as a testament to academic
excellence but also as a proactive step toward fostering a digital landscape
where constructive dialogue can thrive, free from the shadows of toxicity

METHODOLOGY

❖ Dataset
- Source of Dataset: The dataset for the Toxic Comment Classification
project is obtained from Kaggle, a popular platform for machine
 learning datasets and competitions.

- Size of Dataset: The dataset comprises a total of 15,971 comments,

providing a substantial corpus for training and evaluating the Toxic
Comment Classifier.

- Data Format: The dataset is structured in CSV format, facilitating easy

handling and integration into machine learning pipelines.

- Origin of Comments: These comments are collected from various

online platforms, reflecting a diverse range of user-generated content.

- Manual Labeling: To create a robust training set, each comment in the

dataset has undergone manual classification into specific categories,
including:
▪ Toxic

▪ Severe toxic

▪ Obscene

▪ Threat

▪ Insult

▪ Identity hate

- Labeling Criteria: The categorization process involves human

annotators identifying and assigning labels based on the nature and
intensity of toxicity exhibited in each comment.

- Purpose of Labels: The labeling of comments into distinct categories

enables the machine learning model to learn and distinguish between
different forms of toxic language, enhancing its ability to classify
comments accurately.

- Dataset Structure: With comments and their corresponding labels

clearly outlined, the dataset serves as a valuable resource for training
and evaluating the Toxic Comment Classifier.

- Versatility: Given the diverse nature of comments and the range of

toxicity categories, the dataset is designed to foster the development of
a classifier capable of addressing various forms of harmful content in
online platforms.

❖ Hardware and Software requirements:

Software
- VS Code Editor
- Kaggle
- Python and Modules
- Jupyter
Library Used
- Pandas
- Numpy
- MatplotLib
- TensorFlow
- skLearn
- Re
- NLTK
- WordCloud
- Gradio

11
CHAPTER – 4
ARCHITECTURE & USER INTERACTION
FLOW

In the ever-evolving landscape of online communication, the prevalence of toxic

comments poses significant challenges to maintaining a positive and inclusive
digital environment. Recognizing the need for proactive measures against online
toxicity, the Toxic Comment Analyzer emerges as a pivotal solution. This
innovative tool harnesses the power of advanced machine learning and natural
language processing techniques to automatically identify and categorize toxic
language within digital content.

With a dataset sourced from Kaggle, encompassing 15,971 comments manually

labeled for various forms of toxicity, the Toxic Comment Analyzer is designed to
offer a nuanced understanding of harmful online expressions. This introduction
provides a glimpse into the architecture and user flow of the analyzer, outlining
its robust features in preprocessing, feature extraction, and model prediction.

As we delve into the intricacies of this solution, it becomes evident that the Toxic
Comment Analyzer not only aims to enhance online safety but also embodies a
commitment to fostering responsible and respectful digital communication. Let's
explore how this tool, equipped with a sophisticated architecture, navigates the
challenges posed by toxic comments, ultimately contributing to a healthier
online discourse.

1.Data Preprocessing
- Dataset Loading: Load the toxic comment dataset
(e.g., 'train.csv') containing comments and corresponding toxicity labels.
- Text Vectorization: Utilise the TextVectorization layer to convert raw
text into numerical vectors, allowing for efficient processing by the model.
- Label Preparation: Extract the target labels (toxicity categories) and
format them for model training.

2. Model Architecture

The model architecture for a Toxic Comment Classifier typically involves a

combination of natural language processing (NLP) techniques and deep
learning components. Here's a simplified representation of a common
architecture:

1. Input Layer:
- Accepts the preprocessed text input, which goes through initial
tokenization and embedding.

2. Embedding Layer:
- Converts words into dense vectors to capture semantic relationships.
- Pre-trained word embeddings (such as Word2Vec, GloVe, or FastText)
may be used to leverage contextual information.

3. Recurrent Neural Network (RNN) or Long Short-Term Memory (LSTM)

Layer:
- Sequentially processes the embedded tokens to understand the context
and dependencies within the text.
- Captures long-range dependencies, critical for analyzing the sequence
of words in a comment.

4. Attention Mechanism (Optional):

- Enhances the model's focus on specific parts of the input sequence,
allowing it to weigh the importance of different words.

5. Global Average/Max Pooling Layer:

- Aggregates information from the entire sequence to create a fixed-size
representation.
 - Reduces the dimensionality of the data while retaining essential
information.

6. Dense (Fully Connected) Layers:

- Multiple dense layers are employed for learning complex relationships
and representations.
- The number of nodes in the output layer corresponds to the number of
toxicity categories.

7. Output Layer:
- Utilizes a sigmoid activation function for binary classification or
softmax for multi-class classification.
- Generates probability scores for each toxicity category.

8. Loss Function:
- Cross-entropy loss is commonly used for multi-class classification
tasks.

9. Optimizer:
- Adam or RMSprop optimizers are often chosen for efficient gradient
descent during model training.

This architecture leverages the sequential nature of language, allowing the

model to understand the context and relationships between words in a
comment. The use of embeddings, recurrent layers, and attention
mechanisms enables the model to capture both local and global
 dependencies, making it effective for classifying toxic comments. The
specific hyperparameters and layer configurations may vary based on the
dataset and desired model complexity.

3. Linear Regression Implementation

Logistic Regression is a statistical method for analyzing a dataset in which there
are one or more independent variables that predict an outcome. It is commonly
used for binary classification problems.

❖ Advantages:
o Interpretability: Results are easily interpretable, providing
probabilities for class membership.
o Efficiency: Computationally efficient and does not require high
computational resources.
o Less Prone to Overfitting: Less susceptible to overfitting
compared to more complex models when the feature space is
small.
❖ Disadvantages:
o Linear Decision Boundary: Limited to linear decision
boundaries, which might be a drawback for complex datasets.
o Assumption of Linearity: Assumes a linear relationship between
independent variables and the log-odds of the dependent
variable.
o Sensitivity to Outliers: Sensitive to outliers, which can impact the
model's performance.
❖ Use Cases:
o Binary Classification: Well-suited for problems with two classes,
such as spam detection or disease diagnosis.
 o Probabilistic Predictions: Useful when probability estimates for
class membership are required.
❖ Implementation:
o Algorithm: Uses the logistic function to model the probability of
a particular outcome.
o Optimization: Typically optimized using techniques like gradient
descent.
❖ Scalability:
o Scalability: Scales well with the number of features but may not
be the best choice for large and highly complex datasets.

4. Bidirectional LSTM Implementation

❖ Embedding Layer: Create an embedding layer to convert
integer-encoded vocabulary into dense vectors.

Bidirectional LSTM: Implement a Bidirectional LSTM layer to

capture contextual information from both directions in the
sequence.

❖ Bidirectional LSTM Layer: A bidirectional LSTM layer is added

to capture contextual information from the comment text. The
LSTM layer has 32 units and uses the hyperbolic tangent (tanh)
activation function.

❖ Fully Connected Layers: Three fully connected layers with

ReLU activation functions are added as feature extractors. The
layer sizes are 128, 256, and 128, respectively.

❖ Output Layer: Include a dense output layer with sigmoid

activation for multi-label classification.
The model is compiled with the Binary Cross Entropy loss function
and the Adam optimizer.

CHAPTER – 5
WORKING

Long Short-Term Memory networks, or LSTMs, are a type of recurrent

neural network (RNN) architecture designed to overcome the challenges of
learning long-term dependencies in sequential data. Unlike traditional RNNs,
LSTMs possess a more sophisticated memory cell that allows them to capture
and retain information over extended sequences, making them particularly
effective in natural language processing tasks, time series analysis, and other
applications involving sequential data.

Here's a breakdown of the key components and mechanisms that make

LSTMs work effectively:

❖ Memory Cell:
- The core of an LSTM is its memory cell, which serves as a storage unit
capable of retaining information over long periods. This memory cell is
responsible for keeping track of relevant information from earlier parts of
 the sequence.

❖ Three Gates:
- LSTMs employ three gates to regulate the flow of information: the input
gate, forget gate, and output gate

- The input gate determines which values from the input should be stored
in the memory cell.
- The forget gate decides what information to discard from the memory cell.
- The output gate regulates the information that should be output based on
the current input and the memory cell content.

❖ Cell State:
- The memory cell maintains a continuous 'cell state' that runs through the
entire sequence. This state is modified by the gates, allowing the LSTM to
selectively update, add, or remove information from the cell state.

❖ Hidden State:
- The hidden state is the LSTM's way of capturing and storing information
from previous time steps. It acts as a summary or representation of the
relevant information learned from the entire sequence.

❖ Training and Backpropagation:

- During training, LSTMs use a variant of backpropagation called
Backpropagation Through Time (BPTT) . This involves unfolding the
network across time steps and calculating gradients to update the model's
parameters.

❖ Advantages of LSTMs:
- Long-Term Dependencies: LSTMs excel at capturing and learning
dependencies over extended sequences, making them suitable for tasks
requiring an understanding of context over time.
- Gradient Flow: The gating mechanisms help in mitigating the vanishing
and exploding gradient problems that often hinder the training of
traditional RNNs.
- Versatility: LSTMs can be applied to a wide range of sequential data tasks,
including natural language processing, speech recognition, and time series
prediction.
 In summary, LSTMs address the limitations of traditional RNNs by
introducing memory cells and gating mechanisms, enabling them to effectively
capture long-term dependencies in sequential data. This makes them a
powerful tool for tasks that involve understanding context and relationships
across extended sequences.

Bidirectional LSTM
One shortcoming of conventional RNNs is that they are only able to make use of
previous context. … Bidirectional RNNs (BRNNs) do this by processing the data
in both directions with two separate hidden layers, which are then fed forwards
to the same output layer. … Combining BRNNs with LSTM gives Bidirectional
LSTM, which can access long-range context in both input directions
16

CHAPTER – 6
EVALUATION AND RESULTS
Precision, Recall, Accuracy: Utilize precision, recall, and categorical accuracy
metrics to evaluate the model's performance on the test set.

model.compile(loss='BinaryCrossentropy', optimizer='Adam')
history = model.fit(train, epochs=6, validation_data=val)

MODEL SAVING AND LOADING

from tensorflow.keras.metrics import Precision, Recall, CategoricalAccuracy

pre = Precision()
re = Recall()
acc = CategoricalAccuracy()

for batch in test.as_numpy_iterator():

X_true, y_true = batch
yhat = model.predict(X_true)

y_true = y_true.flatten()
yhat = yhat.flatten()

pre.update_state(y_true, yhat)
re.update_state(y_true, yhat)
acc.update_state(y_true, yhat)

print(f'Precision: {pre.result().numpy()}, Recall:{re.result().numpy()},

Accuracy:{acc.result().numpy()}')

model.save('toxicity.h5')

INTEGRATION WITH GRADIO

import gradio as gr

def score_comment(comment):
vectorized_comment = vectorizer([comment])
results = model.predict(vectorized_comment)

text = ''
for idx, col in enumerate(df.columns[2:]):
text += '{}: {}\n'.format(col, results[0][idx]>0.5)
 return text

interface = gr.Interface(fn=score_comment,
inputs=gr.inputs.Textbox(lines=2, placeholder='Comment to score'),
outputs='text')
interface.launch(share=True)

CHAPTER – 7
RESULTS
1. LOSS GRAPH

2. MODEL SCORES
 3. OUTPUT

Toxic Comment as input, Output shows toxic is True

 Non-Toxic Comment as input, Output shows toxic is false

CHAPTER – 8
CONCLUSION

In the culmination of the Toxic Comment Analyzer project, achieving an

impressive accuracy of 92.41% with the integration of LSTM and linear
regression validates the efficacy of our chosen model architecture. This
amalgamation of advanced natural language processing and machine learning
techniques has yielded a powerful tool for discerning toxic language in digital
discourse.

❖ Sequencing Success with LSTM:

❖ The utilization of Long Short-Term Memory (LSTM) layers has played a

pivotal role in capturing intricate sequential dependencies within toxic
comments.
❖ LSTM's capability to retain context over extended sequences enhances the
model's sensitivity to nuanced linguistic patterns.

❖ Linear Regression’s Complementary Role:

❖ The inclusion of linear regression complements the deep learning
architecture, contributing to the model's interpretability and adaptability.
❖ The synergy between LSTM and linear regression has not only boosted
accuracy but has also provided insights into the linear relationships
between different features.

❖ Versatile Model Adaptability:

❖ The hybrid model's versatility is underscored by its adaptability to the

dynamic and varied nature of online language, ensuring robust
performance across diverse datasets.

❖ Technological and Ethical Significance:

❖ Beyond technical achievements, the Toxic Comment Analyzer holds ethical

significance by addressing the crucial need to mitigate toxic language
online.
❖ The accuracy achieved signifies a substantial step toward fostering a more
positive and respectful digital environment.

❖ Collaborative Milestone:

❖ The success of the project is a collaborative achievement, involving

meticulous data preprocessing, feature engineering, and iterative model
fine-tuning.
❖ Continuous collaboration and refinement will be key to the sustained
success and relevance of the Toxic Comment Analyzer.

In conclusion, the Toxic Comment Analyzer not only showcases technical

prowess but also symbolizes a commitment to responsible and inclusive
online communication. As we embark on further developments, this model
stands as a testament to the potential for technology to positively shape
online interactions, promoting an environment characterized by respect,
understanding, and constructive engagement.
FUTURE SCOPE OF WORK

❖ Enhanced Model Generalization:

Improve the model's generalization capabilities through training on
diverse datasets from various online platforms.
Ensure effectiveness across a wide range of contexts and user behaviors.

❖ Dynamic Real-time Monitoring:

Implement a real-time monitoring system for continuous updates based on
evolving linguistic trends and emerging patterns in toxic language.
Ensure the model stays relevant and adaptive over time.

❖ Multimodal Analysis:
Expand analysis to include multimodal content (images, videos) for a
comprehensive approach to toxicity detection.
Enhance the model's capability to handle diverse forms of media.

❖ Expandability and Interpretability:

Integrate techniques for model explain ability and interpretability to
enhance user trust.
Provide insights into why certain comments are classified as toxic,
fostering user understanding.

❖ User Feedback Loop:

Develop an interactive feedback loop allowing users to provide feedback
on the model's predictions.
❖ Enable continuous improvement through iterative retraining based on
user input.

❖ Cross-language Toxicity Detection:

Extend the model's capabilities to detect toxicity in multiple languages.
Address the global nature of online communication, ensuring inclusivity.

❖ Adaptive Learning and Transfer Learning:

 Explore adaptive learning techniques for the model to adapt to changing
user behavior.
Investigate transfer learning approaches to leverage pre-trained models
for related tasks.

❖ Customization for Different Platforms:

Tailor the Toxic Comment Analyzer for specific online platforms.
Account for platform-specific linguistic nuances and user interactions to
enhance accuracy and relevance.

❖ Collaboration with Online Platforms:

Collaborate with online platforms and social media companies for
integration into content moderation systems.
Contribute to creating a safer and more positive online environment
through proactive moderation.

❖ Ethical Considerations and Bias Mitigation:

Conduct in-depth analyses to identify and mitigate potential biases in the
model's predictions.
Ensure ethical deployment, minimizing unintended consequences on
diverse user groups.

❖ Scalability and Cloud Integration:

Optimize the model for scalability to handle large volumes of comments
efficiently.
Explore cloud-based solutions to facilitate easy deployment and
accessibility.

❖ Research on Emerging Toxicity Patterns:

Stay abreast of emerging trends in online toxicity.
Continuously update the model to effectively detect novel forms of harmful
language.

❖ Integration with Educational Initiatives:

Collaborate with educational institutions and online platforms to integrate
the Toxic Comment Analyzer into digital literacy programs.
Foster awareness about responsible online communication and provide
 tools for positive engagement.

The future scope of work for the Toxic Comment Analyzer encompasses
technical advancements, ethical considerations, and collaborative efforts to
create a safer and more inclusive online environment, presented through a
combination of bullet points and explanatory paragraphs.

25

REFERENCES

[1] - Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I.
(2019). Language Models are Few-Shot Learners. arXiv preprint
arXiv:2005.14165.
[2] - Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., ... &
Brew, J. (2020). Transformers: State-of-the-Art Natural Language
Processing. In Proceedings of the 2020 Conference on Empirical Methods
in Natural Language Processing: System Demonstrations (pp. 38-45).

[3] - Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal,
P., ... & Agarwal, S. (2020). Language Models are Few-Shot Learners.
arXiv preprint arXiv:2005.14165.

[4] - Kasneci, Enkelejda & Seßler, Kathrin & Kuechemann, Stefan &
Bannert, Maria & Dementieva, Daryna & Fischer, Frank & Gasser, Urs &
Groh, Georg & Gü nnemann, Stephan & Hü llermeier, Eyke & Krusche,
Stephan & Kutyniok, Gitta & Michaeli, Tilman & Nerdel, Claudia & Pfeffer,
Juergen & Poquet, Oleksandra & Sailer, Michael & Schmidt, Albrecht &
Seidel, Tina & Kasneci, Gjergji. (2023). ChatGPT for Good? On
Opportunities and Challenges of Large Language Models for Education.
103. 102274. 10.1016/j.lindif.2023.102274.

[5] - Dettmers, T., Pagnoni, A., Holtzman, A., & Zettlemoyer, L. (2023).
QLoRA: Efficient Finetuning of Quantized LLMs. ArXiv, abs/2305.14314.

[6] - Vaswani, A., Sukhbaatar, S., Child, R., Shazeer, N., Parmar, N.,
Uszkoreit J. & Polosukhin, I. (2018). On the Pitfalls of Measuring Accuracy
in Natural Language Processing Systems. arXiv preprint
arXiv:1802.09941.

[7] - Hugging Face. (2023). Hugging Face Transformers Library.

Retrieved from https://huggingface.co/transformers/

29

[8] - Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A.,
Cistac, P., Rault, T., Louf, R., Funtowicz, M. and Davison, J., 2019.
Huggingface's transformers: State-of-the-art natural language
processing. arXiv preprint arXiv:1910.03771.
[9] - Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y.,
Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S. and Bikel, D., 2023.
Llama 2: Open foundation and fine-tuned chat models. arXiv preprint
arXiv:2307.09288.

[10] - Sun, S., Zhang, Y., Yan, J., Gao, Y., Ong, D., Chen, B. and Su, J., 2023.
Battle of the Large Language Models: Dolly vs LLaMA vs Vicuna vs
Guanaco vs Bard vs ChatGPT--A Text-to-SQL Parsing Comparison. arXiv
preprint arXiv:2310.10190.

30