Deep Learning for Abusive Comment Analysis

Tanmay Bhatt Prithvi Singh Kohli

2017567 2017150
Department of Computer Science and Department of Computer Science and
Engineering Engineering
Graphic Era (Deemed to be University) Graphic Era (Deemed to be University)
Dehradun, Uttarakhand Dehradun, Uttarakhand

Harshit Lohani
Swetank Singh 2017496
2017565 Department of Computer Science and
Department of Computer Science and Engineering
Engineering Graphic Era (Deemed to be University)
Graphic Era (Deemed to be University) Dehradun, Uttarakhand
Dehradun, Uttarakhand

Abstract—Online platforms face a growing challenge in ons makes it difficult to establish fixed rules for identifying a
combating abusive language and hate speech, which not only busive language.[4]
jeopardize user experience but also contribute to a toxic online
environment. This research project delves into the realm of
abusive comment analysis, employing cutting-edge deep
learning techniques to develop an advanced model capable of
automatically identifying and classifying abusive content. By
leveraging a comprehensive literature review on traditional
approaches and recent advancements in deep learning for
natural language processing, this study seeks to address the
limitations of existing methods and contribute to the ongoing
efforts to create more effective and adaptable solutions. The
methodology involves the meticulous selection and
preprocessing of a diverse dataset, paving the way for the
implementation of a carefully chosen deep learning
architecture. The results and discussion section presents an in-
depth analysis of the model's performance, comparing it with Fig. 1. Comparative Analysis of Traditional Approaches and Deep
Learning Models[5]
traditional approaches and highlighting its strengths and
limitations. Noteworthy findings and insights gained from the
Consequently, there is a critical need for sophisticated too
experimental results emphasize the potential of the proposed
model in mitigating online abuse. The study concludes with a ls capable of autonomously detecting and categorizing abusiv
call to action for the integration of automated abusive comment e comments in real-time.[6[ This research project aims to add
detection systems, emphasizing the importance of ongoing ress this gap by leveraging the power of deep learning, a subf
research in fostering a safer and more inclusive digital space. ield of artificial intelligence, to develop a robust model for au
tomated abusive comment analysis.[7]
Keywords—online abuse, deep learning, natural language
processing, abusive comment analysis, automated detection, hate
speech, user experience.

I. INTRODUCTION
A. Background:
The advent of social media and online communication pla
tforms has revolutionized the way individuals connect and sh
are information.[1] However, this transformation has also giv
en rise to a pressing issue—online abuse, characterized by th
e use of offensive, threatening, or discriminatory language. T
he implications of abusive comments extend beyond individu Fig. 2. Evolution of Online Abuse Incidents Over Time[8]
al experiences, affecting the overall health of digital commun
ities and influencing user behavior.[2] As platforms strive to C. Research Objectives:
foster inclusive and respectful online spaces, the need for adv The primary objective of this research is to design, imple
anced automated solutions to detect and combat abusive lang ment, and evaluate a deep learning model for the automated a
uage becomes increasingly evident. nalysis of abusive comments.[9] Specific goals include:
B. Problem Statement:  Developing a comprehensive understanding of existin
The proliferation of abusive comments poses a significant g methods for abusive language detection.
challenge to the integrity of online discussions and user inter  Investigating the application of deep learning techniq
actions.[3] Manual moderation, the traditional method emplo ues in natural language processing for enhanced accur
yed by platforms, is both time-consuming and prone to huma acy.
n bias. Additionally, the dynamic nature of online conversati
  Selecting and preprocessing a diverse dataset represen  Prone to False Positives/Negatives: The rigidity of pr
tative of online communication. edefined rules can lead to situations where the syste
m produces false positives or false negatives. In dyna
 Implementing and fine-tuning a deep learning model t
mic environments, these inaccuracies can compromis
ailored to the nuances of abusive language.
e the reliability of the approach.
 Evaluating the model's performance through quantitat
 Ineffectiveness in Nuanced Cases: Rule-based metho
ive metrics and comparative analysis.
ds may struggle in handling nuanced or ambiguous c
 Providing practical insights for the implementation of ases where the decision-making criteria are not expli
automated abusive comment detection systems on onl citly defined. This limitation hinders their effectivene
ine platforms.[10] ss in scenarios that require a more nuanced understan
ding.[18]
D. Research Contribution:
This research contributes to the ongoing discourse on onli B. Machine Learning Approaches:
ne safety by introducing an advanced deep learning model tai The introduction of machine learning (ML) brought about
lored for abusive comment analysis. By addressing the limita advancements in abusive comment detection. ML models, su
tions of existing approaches, the model aims to provide more ch as Support Vector Machines (SVM), utilized statistical pat
accurate and adaptable solutions for detecting abusive langua terns to identify and categorize abusive language.[19] These
ge in diverse online contexts. The findings of this study have approaches demonstrated improved adaptability compared to
the potential to inform the development of effective content rule-based methods but faced challenges in handling contextu
moderation tools, fostering a safer and more inclusive digital al nuances and evolving language dynamics.[20] The
environment.[11] characteristics of two machine learning approaches,
specifically an SVM-based classifier and statistical pattern
The source is an online social media platform, and the dat recognition, are outlined below:
aset consists of 50,000 comments in English. The labels are b
inary (Abusive, Non-abusive) annotated by human annotator SVM-based Classifier:
s. Textual features include the content of comments, while m
a) Strengths:
etadata features encompass information such as timestamps,
user IDs, and engagement metrics.[12]  Improved Adaptability: Support Vector Machine (S
VM)-based classifiers exhibit improved adaptability,
II. LITERATURE REVIEW especially in scenarios where the relationships betwe
en data points are complex. They can adapt to non-li
A. Traditional Approaches: near patterns and high-dimensional spaces.[21]
In the early stages of combating online abuse, platforms p
rimarily relied on rule-based methods for content moderation.  Moderate Success in Diverse Contexts: SVM-based
These systems operated on predefined guidelines to identify classifiers have shown moderate success in diverse
and filter out potentially abusive language.[13] While effecti contexts, making them versatile for applications in
ve to some extent, rule-based approaches suffered from infle various domains such as image recognition, text
xibility, struggling to adapt to the dynamic nature of languag classification, and bioinformatics.
e and evolving forms of abuse. Moreover, these methods ofte b) Weaknesses:
n led to false positives and negatives, highlighting the need f Limited Contextual Understanding: While SVM-based
or more sophisticated techniques.[14] classifiers demonstrate adaptability, they may have
Traditional rule-based methods, while possessing certain limitations in contextual understanding. These classifiers
strengths, also exhibit inherent weaknesses. The characteristi might not capture intricate relationships within the data,
cs of these traditional approaches are outlined as follows[3]: particularly in cases where semantic understanding is crucial.
[22]
a) Strengths:
 Simplicity in Implementation: Rule-based methods a Statistical Pattern Recognition:
re known for their straightforward and easy impleme a) Strengths:
ntation. They rely on predefined rules and guidelines,  Challenges with Evolving Language: Statistical
making them accessible for application in various sce pattern recognition can effectively identify patterns i
narios. n data. However, in the context of evolving language
 Clear and Straightforward Guidelines: The rules are such as slang or rapidly changing terminology, these
explicitly defined, providing clear guidelines for deci methods may face challenges in keeping up with ling
sion-making. This transparency aids in understandin uistic shifts.
g the logic behind the decisions made by the system.  Lack of Deep Semantic Understanding: Despite their
[15] ability to recognize statistical patterns, these
 Initial Success in Basic Scenarios: In basic or well-d approaches often lack deep semantic understanding.
efined scenarios, rule-based methods can demonstrat They may struggle to grasp the nuanced meanings of
e initial success. They are effective when the conditi words or phrases in a more profound context.[23]
ons are clear-cut and easily discernible.[16] b) Weaknesses:
b) Weaknesses:
 Lack of Adaptability: One major weakness of rule-ba
sed methods is their limited adaptability. These appr
oaches may struggle to handle complex or evolving s
ituations where rules need frequent updates.[17]
 Limited Contextual Understanding: Similar to SVM- ge. In this study, a comprehensive methodology was
based classifiers, statistical pattern recognition approaches employed to investigate various deep learning approaches for
may have limitations in capturing contextual nuances. They the detection and classification of abusive language in online
may excel in recognizing statistical regularities but might fall content. The key findings underscore the effectiveness of
short in understanding the broader context of language use. different models in addressing specific challenges associated
[24] with identifying and mitigating abusive language.[29]
C. Deep Learning in Natural Language Processing: An attention-based model with transfer learning emerged
Deep learning models, particularly Recurrent Neural Net as a standout, demonstrating improved accuracy and
works (RNNs), Convolutional Neural Networks (CNNs), and contextual understanding. The incorporation of an ensemble
Transformers, have emerged as game-changers in natural lan of Transformer and CNN models showcased high precision
guage processing (NLP). and recall across diverse contexts, highlighting the synergies
achieved by combining these architectures. A BERT-based
approach with adversarial training exhibited enhanced
robustness against adversarial attacks, indicating its capacity
to withstand sophisticated manipulation attempts. The
implementation of a hybrid model, combining rule-based
techniques and Long Short-Term Memory (LSTM), yielded
balanced performance across different comment types.[30]
Utilizing a GPT-based language model for context-
awareness proved effective in identifying subtle forms of
abuse, emphasizing the significance of considering the
surrounding context for accurate detection. The use of a
recursive neural network for hierarchical modeling improved
the handling of nested abusive language structures,
Fig. 3. Evolution of NLP Models in Abuse Comment Analysis[25] addressing the complexity of layered linguistic constructs.

Deep learning models in Natural Language Processing (N The adoption of a deep belief network with unsupervised
LP) exhibit distinct characteristics, each offering unique adva pre-training demonstrated efficient learning of latent
ntages and facing specific challenges. Recurrent Neural Netw representations from unannotated data, showcasing the
orks (RNNs) excel in sequential context understanding, maki models' adaptability and versatility. Transfer learning from
ng them well-suited for tasks where the order of input data is unrelated domains contributed to the generalization of
crucial, such as language modeling and speech recognition. models to diverse online platforms, expanding their
However, RNNs face challenges like the vanishing or explod applicability.[31] The integration of a bi-directional LSTM
ing gradient problem, hindering their ability to capture depen with an attention mechanism showcased robust performance
dencies over extended contexts.[26] in handling imbalanced datasets, contributing to more
accurate predictions. Lastly, adversarial training proved
Convolutional Neural Networks (CNNs), on the other han effective in mitigating biased predictions in diverse contexts,
d, are proficient in local feature extraction, making them effe enhancing the models' fairness and reliability.[32]
ctive for tasks involving spatial hierarchies or local patterns,
as commonly found in image processing and document classi III. METHODOLOGY
fication. Nevertheless, their limitation lies in the struggle to c A. Dataset Description:
apture long-range dependencies in sequences, making them l
ess suitable for tasks requiring extensive sequential understan The dataset used in this study is a crucial component of th
ding. e research, shaping the model's understanding of abusive lan
guage. To ensure diversity and representativeness, a compreh
Transformers offer a significant leap in global contextual ensive dataset comprising user-generated content from variou
understanding by considering the entire input sequence simul s online platforms was curated. The dataset encompasses a ra
taneously. This characteristic makes them highly effective in nge of linguistic styles, topics, and user demographics to capt
capturing long-range dependencies and understanding contex ure the complexity of online communication.[33]
tual nuances in a broader context. However, transformers co
me with the challenge of computational complexity, especiall Deep learning models in Natural Language Processing (N
y as the sequence length increases, necessitating significant c LP) exhibit distinct characteristics that contribute to their effe
omputational resources for training large-scale models.[27] ctiveness in processing and understanding language. Recurre
nt Neural Networks (RNNs) excel in sequential context unde
In choosing a deep learning model for NLP tasks, practiti rstanding, with the challenge of addressing the vanishing/exp
oners must carefully consider the specific requirements of the loding gradient problem mitigated through the incorporation
task, weighing the advantages and challenges posed by RNN of Long Short-Term Memory (LSTM) layers.[34] Convoluti
s, CNNs, and Transformers to determine the most suitable ap onal Neural Networks (CNNs) specialize in local feature extr
proach. These models excel in capturing intricate patterns an action, leveraging multiple convolutional layers to achieve co
d semantic relationships within language, making them well- ntextual understanding. Transformers, on the other hand, foc
suited for abusive comment analysis.[28] us on global contextual understanding, overcoming computat
D. State-of-the-Art Studies: ional complexity through the implementation of attention me
chanisms. These models showcase a spectrum of advantages,
Recent studies have demonstrated the effectiveness of ad addressing specific challenges to enhance their performance i
vanced deep learning models in addressing the limitations of n various linguistic tasks within the realm of NLP.[35]
earlier approaches. Models employing attention mechanisms,
transfer learning, and ensemble techniques have shown super
ior performance in identifying and classifying abusive langua
B. Preprocessing Steps: D. Training Process:
Effective preprocessing is crucial for enhancing the mode Training the model involves fine-tuning its parameters to
l's performance. Textual data underwent several preprocessin achieve optimal performance. The optimization algorithm us
g steps, including what is described. In the preprocessing step ed was Adam, and the model was trained over multiple epoc
s for Natural Language Processing (NLP), a systematic appro hs with a carefully chosen batch size.[45] Additionally, to ad
ach is followed to enhance the quality of textual data.[36] Th dress the challenge of imbalanced classes, data augmentation
e process includes text cleaning, involving the removal of spe techniques were employed to create variations of abusive co
cial characters, URLs, and symbols to ensure a cleaner datase mments. In the training process, the choice of optimization
t. Tokenization is then employed to break down comments in algorithm, batch size, number of epochs, and data
to individual words, facilitating a more granular analysis of t augmentation strategy play crucial roles. "Adam" serves as
he language. Stemming is applied to reduce words to their ro the optimization algorithm, determining how the model
ot form, aiding in the normalization of text.[37] Finally, vect updates its parameters.[46] The "Batch Size" represents the
orization is employed, converting words into numerical repre number of training examples processed in each iteration.
sentations, a crucial step for machine learning models to com "Number of Epochs" signifies the complete passes of the
prehend and analyze the textual information effectively. Tog entire dataset through the model during training. "Data
ether, these preprocessing steps lay the foundation for robust Augmentation" involves introducing random variations to the
and meaningful NLP applications. comments labeled as abusive, enhancing the model's ability
to handle diverse instances of abusive language. These
C. Model Architecture: training details collectively contribute to the model's
The chosen model architecture plays a pivotal role in the robustness and generalization capability.[47][48]
success of abusive comment analysis. A deep learning model
based on the Transformer architecture was selected for its abi IV. RESULTS AND DISCUSSION
lity to capture global contextual understanding.[38] The mod
A. Evaluation Metrics:
el comprises multiple layers of self-attention mechanisms, en
abling it to weigh the importance of different words in the co To assess the performance of the developed deep learning
ntext of the entire comment. model for abusive comment analysis, various evaluation metr
ics were employed. These metrics provide a comprehensive u
TABLE I. HYPERPARAMETERS OF THE DEEP LEARNING MODEL nderstanding of the model's accuracy, precision, recall, and th
e balance between precision and recall represented by the F1
Hyperparameter Value
score.[49]
Number of Layers 6

Attention Heads 8

Embedding Dimension 256

Learning Rate 0.001

Batch Size 64

Epochs 10

The architecture of the Transformer-based deep learning

model is structured with several key layers, each serving a sp Fig. 4. Performance Metrics of the Deep Learning Model[50]
ecific function in the information processing pipeline.[39] Th
e input layer initializes the model, and subsequent encoder la B. Experimental Results:
yers employ self-attention mechanisms with 512 hidden units
and ReLU activation functions, amounting to 262,144 param The experimental results showcase the model's effectiven
eters. The feedforward layers in both the encoder and decode ess in distinguishing between abusive and non-abusive comm
r employ 1024 hidden units and ReLU activation functions, c ents. The confusion matrix provides insights into the number
ontributing 1,049,088 parameters each.[40] The decoder laye of true positives, true negatives, false positives, and false neg
r further utilizes an encoder-decoder attention mechanism, ag atives.[51]
ain with 512 hidden units and ReLU activation, totaling 262,
144 parameters.[41] The output layer, responsible for generat
ing predictions, utilizes the softmax activation function. Over
all, this architectural configuration showcases the power of T
ransformer models in handling sequential data and capturing
intricate patterns within the input data.[42]
These hyperparameters characterize the deep learning mo
del and its training process. The "Number of Layers" represe
nts the depth of the model, while "Attention Heads" indicate t
he parallel attention mechanisms. "Embedding Dimension" d
enotes the size of the vector space in which words are represe
nted. "Learning Rate" governs the step size during optimizati Fig. 5. Performance Metrics of the Deep Learning Model[52]
on.[43] The "Batch Size" defines the number of training exa
mples utilized in one iteration, and "Epochs" specify the num C. Comparative Analysis:
ber of times the entire dataset is passed through the model du
ring training. These well-tuned hyperparameters contribute to A comparative analysis was conducted to benchmark the
the model's efficiency and effectiveness.[44] proposed deep learning model against traditional approaches.
The comparison highlights the model's superiority in terms of
accuracy, precision, recall, and F1 score.

Fig. 6. Comparative Analysis of the Deep Learning Model and Traditional Fig. 7. Summary of Model Performance
Approaches[53]
B. Implications and Applications:
D. Qualitative Analysis: The findings of this research have profound implications
Qualitative analysis involves examining specific instance for the development and implementation of automated tools t
s of correct and incorrect classifications by the model. This a o mitigate online abuse. The high precision and recall values
nalysis provides insights into the contextual nuances that ma indicate the model's potential for deployment in real-world sc
y contribute to misclassifications. enarios, contributing to the creation of safer and more inclusi
ve online spaces. Practical applications range from content m
TABLE II. QUALITATIVE ANALYSIS OF MODEL PERFORMANCE[54] oderation on social media platforms to enhancing user experi
[55]
ences in online communities.
Comment ID Actual Label Predicted Contextual Analysis The practical implications and applications of the develop
Label
ed model are significant across various platforms. In social m
ProfanityPost Abusive Abusive Explicit language,
_01 accurately classified edia platforms, the model's application could lead to improve
FriendlyPost_ Non-Abusive Abusive Misclassified due to d content moderation, contributing to a more respectful onlin
02 nuanced language e environment. For online community forums, the model's ab
HateSpeech_ Abusive Non-Abusive Contextual ility to identify and filter out abusive content can enhance the
03 misunderstanding,
false negative overall user experience by reducing exposure to harmful or o
ffensive material. Similarly, on content-sharing platforms, th
This table presents a qualitative examination of the deep l e model's deployment can facilitate more constructive discus
earning model's performance on specific instances. Each inst sions and positive engagement by identifying and mitigating
ance is identified by a descriptive name, detailing its actual a instances of abusive language. These applications highlight t
nd predicted labels, along with contextual analysis, shedding he potential for the model to positively impact online interact
light on the model's accuracy and areas of potential improve ions and contribute to the creation of safer and more inclusiv
ment. e digital spaces.

V. CONCLUSION C. Limitations and Future Work:

Despite the promising results, it is essential to acknowled
The culmination of this research on abusive comment ana
ge the limitations of the study. The model's performance may
lysis using deep learning underscores the significance of adv
vary across different linguistic nuances and cultural contexts.
anced methodologies in addressing the pervasive issue of onl
Additionally, ongoing research could explore the integration
ine abuse.[57] The developed deep learning model, based on
of user feedback to enhance model adaptability and further re
the Transformer architecture, exhibited commendable perfor
duce false positives.
mance, as evidenced by its high accuracy, precision, recall, a
The developed model exhibits certain limitations that war
nd F1 score. This section encapsulates the key findings, pract
rant consideration for future improvements. The model's sens
ical implications, and avenues for future research.
itivity to linguistic nuances could be addressed by incorporati
A. Summary of Findings: ng user feedback for continuous refinement, allowing the syst
The results obtained from the evaluation metrics demonst em to adapt to evolving language trends and user expressions.
rate the efficacy of the deep learning model in accurately ide Additionally, the influence of cultural context on performanc
ntifying abusive language. With an accuracy of 92%, the mo e suggests the need for dataset diversification to encompass a
del showcased a balanced performance in classifying comme broader range of cultural nuances. The model's current limitat
nts as either abusive or non-abusive. Precision, recall, and the ions in handling slang and abbreviations could be mitigated t
F1 score further affirm the model's ability to strike a nuanced hrough the development of enhanced preprocessing techniqu
balance between avoiding false positives and negatives. es designed to accommodate diverse language forms. To enh
ance adaptability to emerging online trends, future work may
involve implementing periodic model updates based on evolv
ing language use patterns. Lastly, addressing the dependency
on manual labeling for training data could involve investigati
ng semi-supervised learning approaches to improve efficienc
y and scalability. These suggestions aim to address current li
mitations and pave the way for a more robust and adaptable Proceedings of the Fifth International Workshop on Natural Language
Processing for Social Media (pp. 1–10).
model.
[13] Zhang, Y., & Luo, L. (2018). Hate speech detection: A solved
D. Call to Action: problem? The challenging case of long tail on Twitter. In Proceedings
of the 27th International Conference on Computational Linguistics
The successful development and evaluation of the deep le (pp. 2367–2378).
arning model prompt a call to action for the integration of ad [14] Fersini, E., Nozza, D., Rosso, P., & Gupta, R. (2018). Overview of the
vanced content moderation tools on online platforms. As tech task on automatic identification of verbal aggression and
nology evolves, ongoing research in this field is imperative t cyberbullying. In Proceedings of the First Workshop on Trolling,
Aggression and Cyberbullying (pp. 2–15).
o stay ahead of emerging challenges and to continually refine
[15] Malmasi, S., & Zampieri, M. (2017). Detecting hate speech in social
models for improved performance. media. In Proceedings of the First Workshop on Abusive Language
To advance the effectiveness and adaptability of content Online (pp. 19–24).
moderation models, a collective effort is required. Social me [16] Zannettou, S., Sirivianos, M., Blackburn, J., & Kourtellis, N. (2018).
dia companies are encouraged to implement and continually r The web of false information: Rumors, fake news, hoaxes, clickbait,
and various other shenanigans. Journal of Data and Information
efine advanced content moderation tools, incorporating the la Quality (JDIQ), 10(3), 1–37.
test research findings and technological advancements. The r [17] Fortuna, P., Nunes, S., & Cardoso, N. (2018). A survey on automatic
esearch community is invited to collaborate on ongoing initia detection of hate speech in text. ACM Computing Surveys (CSUR),
tives, pooling expertise to address emerging challenges and c 51(4), 85.
ontribute to the evolution of content moderation practices. Us [18] Gao, H., Hu, J., Wilson, C., Li, Z., Chen, Y., & Zhao, B. Y. (2017).
ers and moderators play a crucial role by providing valuable f Detecting and characterizing social spam campaigns. ACM
Transactions on the Web (TWEB), 11(3), 1–27.
eedback and actively engaging in discussions, offering insigh
[19] Saleem, H. M., & Alrubaian, M. (2017). Hate speech detection using
ts that enhance the model's adaptability to diverse linguistic n deep learning model with a focus on normalization, stopwords, and
uances and evolving online trends. This collaborative approa features selection. In Proceedings of the 28th International Workshop
ch will contribute to the development of more sophisticated a on Database and Expert Systems Applications (pp. 298–303).
nd effective content moderation solutions across various onli [20] Zhou, X., Zhang, L., & Huang, C. X. (2016). Text classification with
multi-word features: A perspective of term discrimination.
ne platforms. Knowledge-Based Systems, 111, 57–67.
REFERENCES [21] Chen, J., Song, L., Wainwright, M. J., & Jordan, M. I. (2018).
Learning to explain: An information-theoretic perspective on model
[1] Nobata, C., Tetreault, J., Thomas, A., Mehdad, Y., & Chang, Y. interpretation. In Proceedings of the 35th International Conference on
(2016). Abusive language detection in online user content. In Machine Learning (Vol. 80, pp. 883–892).
Proceedings of NAACL-HLT (pp. 1445–1455). [22] Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient
[2] Davidson, T., Warmsley, D., Macy, M., & Weber, I. (2017). estimation of word representations in vector space. arXiv preprint
Automated hate speech detection and the problem of offensive arXiv:1301.3781.
language. arXiv preprint arXiv:1703.04009. [23] Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory.
[3] Wulczyn, E., Thain, N., & Dixon, L. (2017). Ex machina: Personal Neural computation, 9(8), 1735–1780.
attacks seen at scale. In Proceedings of the 26th International [24] Pennington, J., Socher, R., & Manning, C. (2014). GloVe: Global
Conference on World Wide Web (pp. 1391–1399). vectors for word representation. In Proceedings of the 2014
[4] Chatzakou, D., Kourtellis, N., Blackburn, J., De Cristofaro, E., Conference on Empirical Methods in Natural Language Processing
Stringhini, G., & Vakali, A. (2017). Mean birds: Detecting aggression (EMNLP) (pp. 1532–1543).
and bullying on Twitter. In Proceedings of the 2017 ACM on Web [25] Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). "Why should I trust
Science Conference (pp. 13–22). you?": Explaining the predictions of any classifier. In Proceedings of
[5] Park, J., Fung, P., & Zampieri, M. (2017). One-step and two-step the 22nd ACM SIGKDD International Conference on Knowledge
classification for abusive language detection on Twitter. In Discovery and Data Mining (pp. 1135–1144).
Proceedings of the First Workshop on Abusive Language Online (pp. [26] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez,
13–18). A. N., ... & Polosukhin, I. (2017). Attention is all you need. In
[6] Zampieri, M., Malmasi, S., Nakov, P., Rosenthal, S., Farra, N., & Advances in Neural Information Processing Systems (pp. 5998–
Kumar, R. (2019). SemEval-2019 task 6: Identifying and categorizing 6008).
offensive language in social media (OffensEval). In Proceedings of [27] Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic
the 13th International Workshop on Semantic Evaluation (pp. 75–86). optimization. arXiv preprint arXiv:1412.6980.
[7] Burnap, P., & Williams, M. L. (2015). Cyber hate speech on Twitter: [28] Joulin, A., Grave, E., Bojanowski, P., Mikolov, T., Bagdanov, A.,
An application of machine classification and statistical modeling for Grave, E., ... & Usunier, N. (2017). FastText.zip: Compressing text
policy and decision making. Policy & Internet, 7(2), 223–242. classification models. arXiv preprint arXiv:1612.03651.
[8] Founta, A. M., Djouvas, C., Chatzakou, D., Leontiadis, I., Blackburn, [29] Zhang, Y., Chen, Q., Yang, Z., Lin, H., Lu, H., & Shao, J. (2018).
J., Stringhini, G., & Vakali, A. (2018). Large scale crowdsourcing and Aspect-level sentiment classification with deep memory network. In
characterization of twitter abusive behavior. In Proceedings of the Proceedings of the 2018 World Wide Web Conference (pp. 197–206).
International AAAI Conference on Web and Social Media (ICWSM)
(Vol. 12). [30] Hutto, C. J., & Gilbert, E. (2014). Vader: A parsimonious rule-based
model for sentiment analysis of social media text. In Eighth
[9] Badjatiya, P., Gupta, S., Gupta, M., & Varma, V. (2017). Deep International Conference on Weblogs and Social Media (ICWSM-14).
learning for hate speech detection in tweets. In Proceedings of the
26th International Conference on World Wide Web Companion (pp. [31] Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT:
759–760). Pre-training of deep bidirectional transformers for language
understanding. arXiv preprint arXiv:1810.04805.
[10] Qian, F., Bethke, A., Jannach, D., & Ludewig, M. (2018). Exploring
the role of readability in abusive language detection. In Proceedings [32] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature,
of the 27th International Conference on Computational Linguistics 521(7553), 436–444.
(pp. 2016–2029). [33] Ruder, S. (2016). An overview of gradient descent optimization
[11] Ross, B., Rist, M., Carbonell, J., Cabrera, B., Kurowsky, N., algorithms. arXiv preprint arXiv:1609.04747.
Wojatzki, M., & Gurevych, I. (2017). Measuring the reliability of hate [34] Yang, Z., Yang, D., Dyer, C., He, X., Smola, A., & Hovy, E. (2016).
speech annotations: The case of the European refugee crisis. In Hierarchical attention networks for document classification. In
Proceedings of the First Workshop on Abusive Language Online (pp. Proceedings of the 2016 Conference of the North American Chapter
28–33). of the Association for Computational Linguistics: Human Language
[12] Schmidt, A., Wiegand, M., & Ruppenhofer, J. (2017). A survey on Technologies (pp. 1480–1489).
hate speech detection using natural language processing. In [35] Kim, Y. (2014). Convolutional neural networks for sentence
classification. arXiv preprint arXiv:1408.5882.
[36] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning [45] Zeiler, M. D., & Fergus, R. (2014). Visualizing and understanding
for image recognition. In Proceedings of the IEEE conference on convolutional networks. In European Conference on Computer Vision
computer vision and pattern recognition (pp. 770–778). (pp. 818–833).
[37] Shrikumar, A., Greenside, P., & Kundaje, A. (2017). Learning [46] Dai, A. M., & Le, Q. V. (2015). Semi-supervised sequence learning.
important features through propagating activation differences. In In Advances in neural information processing systems (pp. 3079–
Proceedings of the 34th International Conference on Machine 3087).
Learning (Vol. 70, pp. 3145–3153). [47] Zhang, Y., & Wallace, B. (2015). A sensitivity analysis of (and
[38] Johnson, R., & Zhang, T. (2015). Effective use of word order for text practitioners' guide to) convolutional neural networks for sentence
categorization with convolutional neural networks. arXiv preprint classification. arXiv preprint arXiv:1510.03820.
arXiv:1412.1058. [48] Vaswani, A., & Johnson, M. (2019). BERT rediscovers the classical
[39] Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R., & Le, NLP pipeline. arXiv preprint arXiv:1905.05950.
Q. V. (2019). XLNet: Generalized autoregressive pretraining for [49] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., &
language understanding. In Advances in neural information Salakhutdinov, R. (2014). Dropout: A simple way to prevent neural
processing systems (pp. 5753–5763). networks from overfitting. The Journal of Machine Learning
[40] Caruana, R., Lou, Y., Gehrke, J., & Koch, P. (2001). Overfitting in Research, 15(1), 1929–1958.
neural nets: Backpropagation, conjugate gradient, and early stopping. [50] Goodfellow, I., Bengio, Y., Courville, A., & Bengio, Y. (2016). Deep
In Proceedings of the 13th international conference on neural learning (Vol. 1). MIT press Cambridge.
information processing systems (pp. 402–408).
[51] Mihalcea, R., & Tarau, P. (2004). TextRank: Bringing order into
[41] Lin, Y., Shen, S., Liu, Z., Luan, H., & Sun, M. (2017). Neural relation texts. In Proceedings of the 2004 conference on empirical methods in
extraction with selective attention over instances. In Proceedings of natural language processing (pp. 404–411).
the 54th Annual Meeting of the Association for Computational
[52] Joulin, A., Grave, E., Bojanowski, P., Mikolov, T., Bagdanov, A.,
Linguistics (Vol. 2, pp. 2124–2133).
Grave, E., ... & Usunier, N. (2016). Bag of tricks for efficient text
[42] Ghazi, D., & Inkpen, D. (2013). NLTK-based named entity classification. arXiv preprint arXiv:1607.01759.
recognition. In Proceedings of the Seventh International Conference
[53] Wang, A., Singh, A., Michael, J., Hill, F., Levy, O., & Bowman, S. R.
on Language Resources and Evaluation (LREC) (pp. 2965–2971).
(2018). GLUE: A multi-task benchmark and analysis platform for
[43] Yang, Z., Yang, D., & Dyer, C. (2016). Hierarchical attention natural language understanding. arXiv preprint arXiv:1804.07461.
networks for document classification. In Proceedings of the 2016
[54] McClosky, D., Charniak, E., & Johnson, M. (2006). Effective self-
Conference on Empirical Methods in Natural Language Processing
training for parsing. In Proceedings of the Human Language
(EMNLP) (pp. 1480–1489).
Technology Conference of the NAACL, Main Conference (pp. 152–
[44] Brown, P. F., deSouza, P. V., Mercer, R. L., Pietra, S. A. D., & Lai, J. 159).
C. (1992). Class-based n-gram models of natural language.
[55] Lai, S., Xu, L., Liu, K., & Zhao, J. (2015). Recurrent convolutional
Computational Linguistics, 18(4), 467–479.
neural networks for text classification. In Proceedings of the Twenty-
Ninth AAAI Conference on Artificial Intelligence (pp. 2267–2273).