UNIT-5

Nature of Social Media Text

Social media text has a unique nature, setting it apart from traditional forms of writing.
Social media text is generally short, informal, and full of non-standard expressions.
Platforms like Twitter, Facebook, and Instagram host user-generated content that
includes emojis, hashtags, hyperlinks, abbreviations, and slang. Additionally, users
often mix languages in the same post (e.g., “This movie toh mast tha”), a phenomenon
called code-mixing. These characteristics make the text noisy and structurally
inconsistent, which poses challenges for computational analysis. Unlike formal written
text, social media messages lack proper grammar and punctuation, and are often filled
with sarcasm, humor, and emotional expressions.

Here are some key characteristics:

1. Informality and Conversational Tone: Social media text often mirrors spoken
language. You'll find:
●​ Contractions: Words like "gonna," "wanna," and "ain't" are common.
●​ Slang and Colloquialisms: Everyday, informal language and regional
expressions pop up frequently.
●​ Shorter Sentences: Brevity is key for быстрое consumption.
●​ Direct Address: Users often speak directly to their audience or other users.

2. Brevity and Immediacy: Platforms often impose character limits, encouraging

concise communication. This leads to:
●​ Abbreviations and Acronyms: Think "lol," "btw," "omg," and "tbh."
●​ Hashtags: Used to categorize content and join conversations.
●​ Real-time Updates: Information is often shared as it happens.

3. Multimodality: Text is often combined with other media:

●​ Emojis and Emoticons: Used to convey tone and emotion.
●​ Images and Videos: Visuals are a core part of many social media posts.
●​ Links: Sharing external content is common.

4. Interactivity and Engagement: Social media is designed for two-way

communication:
 ●​ Questions: Often used to encourage comments and discussion.
●​ Mentions and Tagging: Directly referencing other users.
●​ Calls to Action: Encouraging likes, shares, and comments.

5. Personalization and Identity: Users often craft their online persona through their
text:
●​ Personal Opinions and Experiences: Sharing individual thoughts and stories is
common.
●​ Brand Voice: Organizations develop a specific tone and style for their
communications.

6. Dynamic and Evolving Language: Social media language is constantly changing:

●​ New Slang: Terms and phrases emerge and fade quickly.
●​ Platform-Specific Norms: Different platforms have their own unique styles and
conventions.

7. Noisy and Unstructured Data: This informality and rapid creation can lead to:
●​ Grammatical Errors and Typos: Less emphasis on formal writing rules.
●​ Code-Mixing: Especially in multilingual societies, mixing of languages within a
single post is common.
●​ Ambiguity and Context Dependence: Meaning can be heavily reliant on the
surrounding conversation or shared understanding.

Issues and Challenges:

1. Misinformation and Disinformation: The speed and reach of social media make it
easy for false or misleading information to spread rapidly.This can have serious
consequences, impacting public opinion, health, and even safety.

2. Lack of Nuance and Context: The brevity of social media text can strip away
important context and nonverbal cues. This can lead to misunderstandings,
misinterpretations, and even conflict.Sarcasm, for example, is often lost in translation.

3. Cyberbullying and Harassment: The anonymity and public nature of many social
media platforms can create an environment where cyberbullying and harassment thrive.
This can have severe emotional and psychological impacts on individuals.
4. Privacy Concerns: Sharing personal information on social media, even in seemingly
innocuous posts, can lead to privacy breaches and data vulnerabilities. Information can
be used in ways users didn't intend or anticipate.

5. Emotional Contagion and Negative Comparisons: Exposure to the curated and

often idealized lives of others on social media can lead to feelings of inadequacy, envy,
and low self-esteem. Negative emotions can also spread quickly through online
interactions.

6. Addiction and Time Management: The design of social media platforms often
encourages compulsive use, leading to addiction and significant time wasted. This can
detract from productivity, real-life relationships, and overall well-being.

7. Filter Bubbles and Echo Chambers: Algorithms on social media often prioritize
content that aligns with a user's existing beliefs and interests. This can create filter
bubbles or echo chambers, limiting exposure to diverse perspectives and potentially
reinforcing biases.

8. Language Barriers and Code-Mixing: While connecting people globally, language

differences can still be a barrier. Additionally, the informal nature of social media often
leads to code-mixing (using multiple languages in one post), which can be challenging
for automated analysis and sometimes for understanding.

9. Data Overload and Noise: The sheer volume of social media text can be
overwhelming, making it difficult to find relevant and reliable information. Spam,
irrelevant content, and off-topic chatter contribute to this "noise."

10. Challenges for Natural Language Processing (NLP): The informal language,
abbreviations, slang, typos, and code-mixing common in social media text pose
significant challenges for NLP techniques used for analysis like sentiment analysis or
topic modeling.

Pre-Processing Of Social Media Text

1. Data Collection and Cleaning:
●​ Gathering: This is the first step, obviously! You're pulling text data from various
social media platforms (Twitter, Facebook, Instagram, etc.) using APIs or web
scraping (with ethical considerations in mind, of course!).
●​ Handling Duplicates: You might find the same post shared multiple times.
Removing duplicates ensures you're not overrepresenting certain data points.
 ●​ Removing Irrelevant Data: This could include bot posts, promotional content
(depending on your analysis goal), or posts in languages you're not focusing on.

2. Text Cleaning: This is where you tackle the nitty-gritty of making the text more
uniform and analyzable.
●​ Lowercasing: Converting all text to lowercase (e.g., "Hello" becomes "hello")
helps ensure that words are treated the same regardless of capitalization.
●​ Removing Punctuation: Punctuation marks (like commas, periods, question
marks) usually don't add much to the meaning of the words themselves in most
analyses.
●​ Handling Special Characters: Removing or replacing special characters (like
emojis, symbols) depending on your needs. Emojis, for instance, can carry
sentiment, so you might want to handle them differently (e.g., converting them to
text descriptions).
●​ Dealing with URLs and Mentions: You might want to remove URLs or replace
them with a placeholder (like "[URL]"). Similarly, you could remove or mark user
mentions (@usernames).
●​ Handling Hashtags: Decide whether to keep hashtags, remove the "#" symbol
and treat them as regular words, or extract them for separate analysis. Hashtags
often carry valuable information about the topic or sentiment.
●​ Addressing Numbers: Depending on your analysis, you might remove numbers
or convert them to a standard format.

3. Normalization: This step aims to bring words to a more standard form.

●​ Tokenization: Breaking down the text into individual words or "tokens." This is a
fundamental step for most NLP tasks.
●​ Stop Word Removal: Removing common words that don't carry much meaning
(like "the," "a," "is," "are"). The list of stop words can be customized based on
your needs.
●​ Stemming: Reducing words to their root form by removing suffixes (e.g.,
"running" becomes "run," "jumps" becomes "jump"). Common stemming
algorithms include Porter and Snowball.
●​ Lemmatization: Similar to stemming, but it aims to bring words to their dictionary
form (lemma) using morphological analysis (e.g., "better" becomes "good,"
"running" becomes "run"). Lemmatization is generally more accurate than
stemming but computationally more intensive.
●​ Handling Abbreviations and Acronyms: You might want to expand common
abbreviations (like "lol" to "laughing out loud") or handle them consistently. This
can be tricky as context is often needed.
 ●​ Dealing with Slang and Emoticons: This is one of the biggest challenges with
social media text. You might use dictionaries or look-up tables to replace
common slang terms with their standard equivalents or analyze the sentiment
conveyed by emoticons.

4. Feature Engineering (Optional but Often Crucial): After the basic cleaning and
normalization, you might want to create new features from the text that can be useful for
your analysis.
●​ N-grams: Considering sequences of words (e.g., "very good," "not happy"). This
can capture more contextual meaning than individual words.
●​ Sentiment Scores: Using sentiment analysis tools to assign a sentiment score
(positive, negative, neutral) to each piece of text.
●​ Topic Modeling: Identifying the main topics discussed in the text.
●​ Word Embeddings (Word2Vec, GloVe, FastText): Representing words as
dense vectors in a high-dimensional space, capturing semantic relationships
between words. These are often used as input for more advanced machine
learning models.

Why is Pre-processing So Important for Social Media Text?

●​ Improved Accuracy: By cleaning and normalizing the text, you reduce noise
and inconsistencies, leading to more accurate results in your analysis.
●​ Better Model Performance: Machine learning models trained on clean and
well-processed data tend to perform significantly better.
●​ More Meaningful Insights: Pre-processing helps to focus on the actual content
and meaning of the text, allowing you to extract more valuable insights.

Aggressive and abusive content detection

Detecting aggressive and abusive content in social media text is a multifaceted
challenge requiring a combination of linguistic understanding, advanced machine
learning techniques, careful data handling, and ongoing ethical considerations. While
significant progress has been made, it remains an active area of research and
development.

There are primarily two main categories of approaches, often used in combination:

1. Rule-Based Approaches:
 ●​ Keyword Lists: This involves creating lists of offensive words, phrases, and
slurs. When these keywords are detected in a text, it can be flagged as
potentially aggressive or abusive.
●​ Regular Expressions: More sophisticated pattern matching can be used to
identify variations of offensive terms (e.g., using wildcards or character
substitutions).
●​ Limitations: Rule-based systems are often brittle, meaning they can be easily
bypassed by slight variations in language or the use of novel offensive terms.
They also struggle with context and may flag non-offensive language if it
happens to contain a keyword.

2. Machine Learning (ML) and Deep Learning (DL) Approaches:

These are the dominant approaches due to their ability to learn complex patterns and
handle nuanced language.

●​ Traditional Machine Learning:​

○​ Feature Engineering: This involves manually extracting relevant features

from the text, such as:
■​ N-grams: Sequences of words that capture local context.
■​ Character N-grams: Sequences of characters that can help
identify misspellings and variations of offensive terms.
■​ Sentiment Scores: Overall sentiment of the text.
■​ Presence of offensive keywords or patterns (as in rule-based
systems, but used as features).
■​ Stylistic features: Word count, sentence length, use of
capitalization, punctuation.
○​ Classifiers: These features are then fed into machine learning
models,like:
■​ Naive Bayes: A probabilistic classifier.
■​ Support Vector Machines (SVM): Effective in high-dimensional
spaces.
■​ Logistic Regression: A linear model for binary classification.
■​ Random Forest: An ensemble learning method.

●​ Deep Learning:​

○​ Word Embeddings: Techniques like Word2Vec, GloVe, and FastText

learn dense vector representations of words, capturing semantic
 relationships. These embeddings can automatically capture contextual
information.
○​ Recurrent Neural Networks (RNNs), including LSTMs and GRUs:
These are well-suited for processing sequential data like text and can
capture long-range dependencies.
○​ Convolutional Neural Networks (CNNs): While often used for images,
CNNs can also be effective for text by learning local patterns.
○​ Transformer Networks (e.g., BERT, RoBERTa): These state-of-the-art
models have shown remarkable performance in various NLP tasks,
including text classification. They excel at understanding context and
capturing complex relationships between words.
○​ Multi-task Learning: Training models to simultaneously detect different
types of abuse (e.g., hate speech, cyberbullying, threats) can improve
overall performance.

Challenges in Aggressive and Abusive Content Detection:

●​ Subjectivity and Context Dependence: What constitutes "aggressive" or
"abusive" can be subjective and heavily dependent on social and cultural context.
The same phrase might be acceptable in one community but offensive in
another.
●​ Evolving Language: New slang, offensive terms, and coded language
constantly emerge, requiring continuous updates to detection systems.
●​ Sarcasm and Irony: Detecting these nuances, where the literal meaning of
words contradicts the intended meaning, is very challenging for automated
systems.
●​ Implicit Bias in Data: Training data used for ML models may contain biases
reflecting societal prejudices, leading to models that unfairly target certain
groups.
●​ Code-Mixing and Multilingualism: Social media users often switch between
languages or mix them within a single post, making detection more complex.
●​ Variations in Spelling and Grammar: The informal nature of social media text
leads to frequent misspellings, abbreviations, and grammatical errors, which can
confuse detection systems.
●​ Evasion Techniques: Users may intentionally try to evade detection by using
character substitutions, leetspeak, or other obfuscation methods.
●​ Low-Resource Languages: Developing effective detection systems for
languages with limited available data is particularly challenging.
 ●​ Fine-grained Categorization: Distinguishing between different types of harmful
content (e.g., insults vs. threats vs. hate speech) requires more sophisticated
approaches.
●​ Scalability: Processing the massive volume of social media data in real-time
requires efficient and scalable detection systems.

Offensive and Hate text detection

Detecting offensive and hate text on social media is a closely related but distinct task
from detecting general aggressive or abusive content. While there's overlap, the focus
here is specifically on identifying language that is intended to insult, demean, or incite
hatred against individuals or groups based on protected characteristics.

Defining Offensive vs. Hate Speech:

●​ Offensive Language: This is a broader category encompassing insults,
profanity, and disrespectful language that may not target specific groups based
on their identity. It can be directed at individuals or general situations.
●​ Hate Speech: This is a more severe form of offensive language that targets
individuals or groups based on attributes like race, ethnicity, religion, gender,
sexual orientation, disability, etc. It often promotes violence, discrimination, or
prejudice.

Approaches to Detection:

1. Rule-Based Systems (with a focus on hate indicators):

●​ Hate Keyword Lists: These lists contain slurs, derogatory terms, and phrases
specifically targeting protected groups.
●​ Pattern Matching: Identifying patterns that indicate hateful intent, such as
combinations of negative sentiment words with group identifiers.
●​ Limitations: Suffers from the same issues as general rule-based systems
(brittleness, context insensitivity, evasion).

Specific Challenges in Offensive and Hate Text Detection:

●​ Fine-grained Categorization: Distinguishing between different levels and types
of offensive and hateful content (e.g., mild insults vs. severe threats, different
forms of hate speech targeting various groups).
●​ Implicit and Indirect Hate Speech: Hate can be expressed subtly through
coded language, dog whistles, and veiled references, making it difficult to detect.
 ●​ Target Group Identification: Identifying the specific target group of hate speech
can be challenging, especially when implicit references are used.
●​ Intersectionality: Recognizing hate speech that targets individuals based on
multiple intersecting identities (e.g., a Black woman) is a complex task.
●​ Cultural and Regional Variations: What is considered offensive or hateful can
vary significantly across cultures and regions.
●​ Euphemisms and Code Words: Hate groups often develop and use
euphemisms or code words to evade detection.
●​ Contextual Ambiguity: The same words or phrases can have different
meanings depending on the context. For example, a slur used within a
community as reclamation might be hateful when used by an outsider.
●​ The "Intent" Problem: Determining the speaker's intent is often impossible for
automated systems, but intent is a key factor in defining hate speech.

Categories of Hate speech

1. Targeted Groups: This is perhaps the most common way to categorize hate speech,
focusing on the specific groups being attacked. Examples include:

●​ Racism: Hate speech targeting individuals or groups based on their race or

ethnicity. This can include slurs, stereotypes, and incitement to discrimination or
violence.
●​ Sexism: Hate speech directed at individuals based on their gender, often
targeting women with misogynistic language, threats, or the promotion of harmful
stereotypes.
●​ Homophobia: Hate speech targeting individuals based on their sexual
orientation, often using derogatory terms, promoting discrimination, or inciting
violence against LGBTQ+ people.
●​ Transphobia: Hate speech directed at transgender individuals, often denying
their identity, using incorrect pronouns, or promoting discrimination and violence.
●​ Religious Discrimination (e.g., Anti-Semitism): Hate speech targeting
individuals or groups based on their religious beliefs or lack thereof. This can
involve stereotypes, conspiracy theories, and incitement to hatred or violence.
●​ Xenophobia: Hate speech targeting individuals based on their nationality or
origin, often portraying them as outsiders or threats.
●​ Ableism: Hate speech targeting individuals with disabilities, often using
derogatory language, stereotypes, or denying their rights and dignity.
●​ Ageism: Hate speech targeting individuals based on their age, often involving
stereotypes or discriminatory remarks.
2. Forms of Expression: Hate speech can be expressed in various forms:

●​ Slurs and Epithets: Derogatory terms used to insult or demean individuals or

groups.
●​ Stereotypes: Harmful generalizations about entire groups of people.
●​ Dehumanization: Portraying targeted groups as less than human, often
comparing them to animals, insects, or diseases.
●​ Demonization: Presenting targeted groups as evil, malicious, or a threat to
society.
●​ Incitement to Violence: Direct calls for violence or harm against specific groups.
●​ Denial and Minimization of Atrocities: Downplaying or denying historical or
ongoing violence and discrimination against targeted groups (e.g., Holocaust
denial).
●​ Hateful Imagery and Symbols: Use of symbols, memes, and images to convey
hateful messages.
●​ Conspiracy Theories: Blaming specific groups for societal problems based on
unfounded and malicious theories.
●​ Justification of Discrimination: Arguments attempting to legitimize unequal
treatment or prejudice against certain groups.

3. Intensity and Severity: Hate speech can range in its intensity:

●​ Mild Insults: While offensive, they may not necessarily incite hatred or violence.
●​ Severe Vilification: Language that strongly denigrates and dehumanizes.
●​ Incitement: Explicitly or implicitly encouraging violence or discrimination.

4. Public vs. Private: While the focus is often on public expressions, hate speech can
occur in private settings as well. However, legal and social responses often differ.
Publicly expressed hate speech is generally considered more serious due to its
potential to influence a wider audience.

5. Online vs. Offline: The internet and social media have provided new avenues for the
rapid dissemination of hate speech, often with a sense of anonymity.

Hate text detection with deep learning

Deep learning has significantly advanced the field of hate speech detection due to its
ability to automatically learn complex patterns and contextual nuances in text data,
outperforming traditional machine learning methods that rely on manual feature
engineering.
Deep Learning Models for Hate Speech Detection:
●​ Recurrent Neural Networks (RNNs): Models like LSTMs and GRUs excel at
processing sequential data, making them suitable for understanding the temporal
structure of text and capturing long-range dependencies crucial for identifying
hate speech.
●​ Convolutional Neural Networks (CNNs): While initially designed for image
processing, CNNs can effectively capture local and global patterns in text by
applying filters over word embeddings, identifying key phrases and stylistic
elements indicative of hate speech.
●​ Transformer Networks: Architectures like BERT, RoBERTa, and other
pre-trained language models have achieved state-of-the-art results. Their ability
to understand context bidirectionally and capture intricate relationships between
words makes them highly effective in discerning subtle forms of hate speech.
Fine-tuning these models on specific hate speech datasets is a common
practice.
●​ Multi-task Learning: Training models to simultaneously recognize different
categories of harmful content, including hate speech, can improve the overall
performance and robustness of the detection system.
●​ Graph Neural Networks (GNNs): These models can capture the relational
structure of social networks and how hate speech propagates through user
interactions, providing a broader understanding of the phenomenon.
●​ Multimodal Learning: Combining text analysis with the processing of images,
videos, and audio can enhance detection accuracy, as hate speech often
manifests across multiple media types.

Challenges of Using Deep Learning for Hate Speech Detection:

●​ Subtlety and Context Dependence: Hate speech can be expressed indirectly,
using sarcasm, coded language, or dog whistles, which are difficult for deep
learning models to interpret without a strong understanding of context.
●​ Evolving Language: The constant emergence of new slang, offensive terms,
and evasion tactics requires continuous adaptation and retraining of deep
learning models.
●​ Bias in Data: Training datasets may inadvertently contain societal biases,
leading to models that unfairly target or misclassify speech from certain
demographic groups.
●​ Data Scarcity: High-quality, labeled datasets that cover the diverse forms of hate
speech are often limited, especially for low-resource languages.
●​ Explainability: Deep learning models are often "black boxes," making it
challenging to understand why a particular text was classified as hate speech,
which is crucial for debugging and ensuring fairness.
Key Considerations:
●​ Careful Data Annotation: High-quality, diverse, and expertly annotated datasets
are essential for training robust and unbiased deep learning models.
●​ Contextual Awareness: Models need to incorporate contextual information
beyond the text itself, such as user history and community guidelines.
●​ Bias Mitigation Techniques: Strategies to identify and reduce bias in both the
data and the models are crucial for ethical and fair detection.
●​ Human Oversight: Deep learning systems should ideally work in conjunction
with human moderators to handle ambiguous cases and provide crucial
contextual understanding.
●​ Continuous Adaptation: Models must be continuously monitored and updated
to address evolving language and new forms of hate speech.

Cyberbullying Detection

Cyberbullying detection systems use machine learning to identify and flag potentially
abusive content on social media or online platforms. These systems analyze text,
images and sometimes video data to detect patterns and language that suggest
cyberbullying behaviour.

Here’s a more detailed list of process:

1.​ Data Collection and preprocessing

Cyberbullying detection systems start by collecting data from various sources like
social media platforms, online forums, and other online communities.This data is
then preprocessed to clean it, remove irrelevant information, and prepare it for
analysis.
●​
2.​ Feature Extraction
●​ The preprocessed data is analyzed to extract relevant features that can be used
to identify cyberbullying.
●​ These features can include textual features like sentiment analysis, word
frequency, and the presence of specific keywords, as well as visual features for
image-based cyberbullying.
●​
●​ 3. Model Training
 ●​ Machine learning models are trained on labeled data to identify patterns
associated with cyberbullying.
●​ These models can be supervised (using labeled data) or unsupervised (learning
patterns from unlabeled data).
●​ Commonly used models include:
●​ Deep Neural Networks (DNNs): Convolutional Neural Networks (CNNs),
Long Short-Term Memory (LSTM) networks, and Bidirectional LSTM
(BLSTM) networks.
●​ Other Machine Learning Algorithms: Support Vector Machines (SVMs),
Random Forests, Logistic Regression, and Naive Bayes.
●​ Transfer Learning: Using pre-trained models like BERT, which have been
trained on vast amounts of text data, can significantly speed up the training
process and improve accuracy.
●​
●​ 4. Cyberbullying Detection
●​ The trained models are used to analyze new data and predict whether it contains
cyberbullying content.
●​ The system flags any suspicious content for review by human moderators or
administrators.
●​
●​ 5. Model Evaluation and Improvement
●​ The performance of the cyberbullying detection system is evaluated based on
metrics like accuracy, precision, recall, and F1-score.
●​ The system is continuously improved by retraining the models on new data and
fine-tuning the parameters to optimize its performance.
​
​ Specific Techniques and Approaches:
​ 1. Sentiment Analysis:​
Identifying the emotional tone of the text, which can be a clue to cyberbullying
behavior.
​ 2. Psycholinguistics:​
Analyzing the language patterns used in cyberbullying, such as the use of specific
words, phrases, and sentence structures.
​ 3. Toxicity Features:​
Using features that are specifically designed to detect harmful content, such as
profanity, insults, and threats.
​ 4. Deep Learning:​
Using deep neural networks to learn complex patterns in the data, such as the
relationships between words and sentences in a text.
​ 5. Ensemble Learning:​
Combining multiple machine learning models to improve accuracy and robustness.
​
​ Challenges and future directions:
​ 1. Context Matters:​
Cyberbullying can be subtle and context-dependent, making it difficult for algorithms
to detect.
​ 2. Evolving Language:​
Cyberbullying language and tactics constantly evolve, requiring ongoing research
and development to stay ahead.
​ 3. Multilingual Support:​
The majority of research has focused on English, but there's a need to develop
systems that can detect cyberbullying in multiple languages.
​ 4. Image and Video Analysis:​
Developing robust algorithms for detecting cyberbullying in images and videos is an
ongoing area of research.

Revenge Posts detection

1. Rule-Based Systems:
These rely on explicitly defined rules created by human experts to identify potential
revenge posts.
●​ Keyword and Phrase Lists: Compiling lists of words and phrases commonly
associated with revenge, threats, insults, humiliation, and negative targeting
(e.g., "expose," "ruin," "get back at," specific derogatory terms aimed at an
individual).
●​ Regular Expression Matching: Using patterns to identify variations of these
keywords and phrases, as well as specific grammatical structures that might
 indicate vengeful intent (e.g., direct accusations combined with negative
sentiment).
●​ User Behavior Analysis Rules: Flagging users with a history of aggressive
communication, repeated targeting of specific individuals, or sudden spikes in
negative posts directed at someone.
●​ Contextual Rules: Defining rules that consider the surrounding context of
keywords. For example, "expose their secret" in a negative context might be
flagged, while "expose the truth" in a general discussion would not.
●​ Metadata Analysis Rules: Examining metadata associated with posts, such as
timestamps, frequency of posting about a specific individual, or the use of
specific hashtags known to be associated with online harassment.
●​ Image/Video Analysis Rules (Basic): For image/video content, basic rules
might involve flagging content reported by users as non-consensual intimate
imagery or matching against known hashes of such content (though this is more
reactive).

Limitations of Rule-Based Systems:

●​ Brittleness: Easily bypassed by variations in language, slang, or misspellings.
●​ Context Insensitivity: Can struggle to differentiate between genuine revenge
posts and other forms of negative expression.
●​ Scalability and Maintenance: Requires constant updating as language and
online behaviors evolve.
●​ Difficulty with Nuance: May miss subtle or indirect forms of revenge.

2. Traditional Machine Learning Approaches:

These involve extracting features from the text or metadata and training machine
learning classifiers.
●​ Feature Engineering:
○​ Sentiment Analysis Scores: Calculating the overall sentiment of the post
and the sentiment directed towards specific entities mentioned. Highly
negative sentiment towards a named individual could be a feature.
○​ N-grams: Analyzing sequences of words to capture local context and
identify patterns of vengeful language.
○​ Lexicon-Based Features: Using dictionaries of negative, aggressive, and
threatening words, as well as words related to reputation damage or
exposure. Counting the occurrences of these words.
○​ User-Centric Features: Features based on the poster's history (e.g.,
frequency of negative posts, targets of past negativity) and the recipient's
history (e.g., being frequently targeted).
 ○​ Network-Based Features: Analyzing the social network around the
poster and the target, looking for patterns of coordinated negative
behavior.
○​ Stylistic Features: Analyzing writing style, such as the use of aggressive
language markers (e.g., excessive capitalization, punctuation, rhetorical
questions with negative intent).
●​ Machine Learning Classifiers:
○​ Naive Bayes: Effective for text classification tasks.
○​ Support Vector Machines (SVM): Can handle high-dimensional feature
spaces.
○​ Logistic Regression: Provides probabilities of a post belonging to the
"revenge" class.
○​ Random Forests: Ensemble method that can handle complex
relationships between features.

Advantages of Traditional ML:

●​ Better Generalization: Can often generalize better to unseen data compared to
rigid rule-based systems.
●​ Can Learn Complex Patterns: Machine learning models can learn more
intricate relationships between features than simple rules.

Limitations of Traditional ML:

●​ Reliance on Feature Engineering: Performance is heavily dependent on the
quality and relevance of the manually engineered features.
●​ Difficulty with Semantic Understanding: May struggle with understanding the
underlying meaning and intent behind the language as well as deep learning
models.
●​ Handling Context Remains Challenging: While N-grams help, capturing
long-range dependencies and nuanced context is still difficult.

Hybrid Approaches:
Combining rule-based methods with traditional machine learning can be effective.
For example:
●​ Use rule-based systems to flag potentially problematic content based on
keywords or patterns.
●​ Then, use a machine learning classifier with more sophisticated features to
further analyze the flagged content and make a final determination.
Overall, detecting revenge posts without deep learning relies on carefully crafted rules
and the extraction of relevant features from the text and metadata to train traditional
machine learning classifiers. While these methods have limitations in handling nuanced
language and context compared to deep learning, they can still be effective when
well-designed and continuously updated.

Case Study: HateCircle and Unsupervised Hate Speech Detection incorporating

Emotion and Contextual Semantic

Study this on following parameters along with your assignment:

1. The Problem

2. The Proposed Unsupervised Framework

3. General Offensive Content

4. Examples and Results

5. Significance (for Non-Deep Learning Unsupervised Detection)

6. Conclusion