Hate Speech / Toxic Comment Detection - Data Mining (CSE-362) Project

Hate Speech / Toxic Comment
Detection
Harshit Agrawal (18074019)
Ashish Kumar (18075068)
Sachin Srivastava (18075070)
Data Mining
(CSE-362) Project
Odd Semester
2020-21
Under the guidance of
Dr. Bhaskar Biswas
Department of
Computer Science
and Engineering
IIT (BHU) Varanasi
27.09.2020

Overview
Detection and Flagging of comments
containing text which may incite
violence, spread hate and induce
negativity.

Introduction
The project Hate Speech / Toxic Comment Detection aims at:
● Removing any textual material containing hate or toxicity.
● Online bullying has become a serious issue in recent years.
● The threat of abuse and harassment online means that many people stop
expressing themselves.
● Toxic comments spread negativity, racial harassment and depression
● Removing hate speech makes easy for everyone to express their opinions freely
over the internet.

Motivation & Applications
● Social networking: Message or tweet can be reported to the support team.
● Online meetings / webinars: Message can be hidden from attendees and shown
only to the moderators.
● Chatbot training: The model can give large negative feedback for any insulting
reply towards the user.
● Threatening messages: Directly reported to concerned authorities for
immediate actions.

Problem Visualisation Preprocessing Training
Description of the
Problem Statement
and Challenges
Preprocessing of
data into machine
readable format
Cleaning and
Visualisation of the
dataset
Training the model
using various methods
Conclusion
Conclusion and
Scope of
Improvement

Problem Description
Statement
Given any text or paragraph containing a few lines in natural language, the objective
is to classify it as belonging to one or more of the following categories:
clean, obscene, threatening, insulting, toxic, severely toxic and identity hate.
Model
Severely Toxic
Toxic
Insulting
Threatening
Obscene
Clean
Identity Hate
Text / Paragraph
(natural language)
INPUT
OUTPUT
ONE OR
MORE OF
THESE
CLASSES

Problem Description
● Model will output the probability of text to belong to each category.
● Based on certain threshold (which can be tuned as hyperparameter), the text
can be classiﬁed to be belonging to a category / set of categories.
Multi-class classification problem Multi-label classification problem
Seven different classes:
Clean, obscene, threatening,
insulting, toxic, severely toxic,
identity hate.
Any text or paragraph can be
abusive in multiple ways: it
can be both threatening &
insulting to someone

Dataset
● Dataset is made publicly available by Conversation AI.

Labels
● The dataset has been labeled on a crowdsourced annotation platform.
● Different users may have different opinions about a comment (judgement bias).
● Some offensive ‘ASCII arts’ which may or may not be labelled as toxic by users
depending on how it was rendered.
● Some extremely long comments got labelled to toxic without any signiﬁcant
reason.
● “Give us your address and we'll come and kill you” : Not marked as toxic!

Challenges
Noisy Data: Real world dataset from Internet users. Full of slangs, contractions,
non-english or meaningless words and spams.
Unbalanced Dataset: Large number of comments are clean. Unclean samples are
relatively rare and with six classes of toxicity problem is further aggravated.
Computational Capability: Natural Language Processing models like RNN requires
huge processing powers and powerful GPUs for efﬁcient training.
Memory Issues: More than 1 lakh training samples - difﬁcult to store and process in
internal memory at a time.

Solution Approach
Training the model
Data Post-processing
Data Pre-processing
Dataset
XGBoost
Transfer Learning
Extremely Random Trees
Get qualitative high-level
insights into the data
Conversational AI
Remove punctuations, stop
words, stemming, etc.
Reason out the possible
source of errors
Data Cleaning & Visualization
Support Vector Machines
Recurrent Neural Networks
Binary Relevance
Classifier Chains

● Checking for missing or null values.
● Adding extra column for ‘clean’ comments.

Class Imbalance: The toxicity is not spread evenly across
all the classes. Out of all the hate tags, ‘toxic’ tags are
43.58%, whereas ‘threat’ tags are 1.36%.
Clean comments are ~140k out
of the ~160k total comments.

● Comments obtained are from random users on internet
● Dataset is noisy. No ﬁxed standard of writing across all the comments
● Dataset must be cleaned before further steps to obtain meaningful results
● Remove non alphanumeric characters (ip addresses, time, date) and
punctuations.
● Convert everything to lower case letters.
● Detect and remove extremely long words (may cause overﬁtting), non english
and meaningless words.

How long are comments?

Are longer comments more toxic?
Violin Plot Box Plot
No signiﬁcant correlation between length & toxicity is observed.

Violin Plot Box Plot
Slight correlation is observed. Lesser words increase the chances of toxicity.
Are longer comments more toxic?

A cursory look at comments reveal many comments which look like spam.
EPIC FAIL! EPIC FAIL EPIC FAIL! EPIC FAIL EPIC FAIL! ...
EPIC FAIL! EPIC FAIL EPIC FAIL! EPIC FAIL EPIC FAIL! ...
How to classify spammers?
Though spam detection is a wide area of research in its own we can use a very simple
approach which sufﬁce for our application.
Take comments with lot of repeated words (i.e. very less % of unique words).
Are spammers more toxic?

Lesser percentage of unique words increases the chances of toxicity.
Bar Plot

More percentage of unique words DOES NOT decrease the chances of toxicity.
KDE Plot

Multi-class classiﬁcation: A lot of comments
have only one tag, yet there do exist some
comment having 5 or 6 hate tags.
Heatmap of the correlation
between the tags.

[18:00 3/7/2010]: Hey man!!, I'm
really not trying to edit war. It's
just that this guy is constantly
removing relevant information and
talking to me through edits instead
of my talk page. He seems to care
more about the formatting than
the actual info. (@172.16.254.1)
hey man i m really not trying to
edit war it s just that this guy is
constantly removing relevant
information and talking to me
through edits instead of my talk
page he seems to care more about
the formatting than the actual info
DATA CLEANING
COMMENT CLEANED COMMENT

Data Preprocessing
Machine Readable Format
Cleaned Data
Data Preprocessing
out, what, most, any, off,
too, have, more, or, the,
ours, both, whom, and, of,
aren, her, does, from, if, not,
own, this, it, a, it's, hers,
why, who, now, been, me, ...
STEMMING
argue
argued
argues
argu
STOP WORDS
arguing

Data Preprocessing
Conversion of cleaned data to machine readable format.
Stop Words
● Commonly used words which don’t add any meaning to the text.
● E.g.: ‘is’, ‘are’, ‘this’, ‘at’, etc.
● Removed the stop words using NLTK stopwords corpus.

Data Preprocessing
Stemming
● Transforming a word into its base stem, i.e., a set of characters to construct a
word and its derivatives.
● Helps in reducing the vocabulary of words, which convey the same meaning.
● E.g. ‘argue’, ‘argued’, ‘argues’ and ‘arguing’ are reduced to the base stem ‘argu’.
● Used NLTK implementation of the Porter Stemming Algorithm - it contains a
set of rules to transform a word into its base stem.

Data Preprocessing
hey man i m really not trying to
edit war it s just that this guy is
constantly removing relevant
information and talking to me
through edits instead of my talk
page he seems to care more about
the formatting than the actual info
hey man realli tri edit war guy
constantli remov relev inform talk
edit instead talk page seem care
format actual info
DATA PREPROCESSING
CLEANED COMMENT COMMENT IN MACHINE READABLE FORM

● Constructs a Maximum Marginal Hyperplane
that best divides the dataset into classes.
● Hyperplane: decision plane that separates a
set of objects belonging to different class.
● Support Vectors: Data points which are
closest to the hyperplane.
● Margin: gap between the two lines on closest
class points.

● Generates optimal hyperplane iteratively to
minimize the error.
● Objective: Find hyperplane with maximum
possible margin between support vectors.
● Can be used for both linearly separable as
well as non-linearly separable data.

● For non-linear and inseparable planes, SVM
uses Kernel trick.
● Transforms input space to higher dimensional
space.
● Here, data points are plotted on x-axis and
z-axis, where z = x2
+ y2
.
● These data points can be segregated linearly.

Encoding Text Data
Count vectoriser and TF-IDF vectoriser:
● Used for text feature extraction
● Transform text → meaningful representation of numbers
● To ﬁt Machine Learning algorithm for prediction.
● Tokenizing, counting, normalizing are used to extract features from text.
● Tokenizing: Text → split into tokens & assigning integer value to each token.
● Counting: counting occurence of token in document
● Normalizing: Diminishing those tokens that occur in most of the documents.

Encoding Text Data
Count vectoriser:
● Converts text to lowercase, removes punctuations & stopwords.
● Builds vocabulary of text, containing text from all documents
● Encodes the text into vectors with dimension |V| x |D|, where
|V| is size of vocabulary, and |D| is the number of documents.
● Each element of the vector represents the frequency of the token in the
document.
● Issue: Certain words which occur in most of the documents won’t be very
meaningful in the encoded vectors.

Encoding Text Data
TF-IDF vectoriser:
● Reduces impact of tokens that occur in most of the documents.
● Value of a term in vector increases proportionally to frequency of words in
document
● If the token appears in most of the documents, then that term is penalized.
● Term Frequency - TF(t, d): Number of times a term ‘t’ appears in document ‘d’.
● Inverse Document Frequency - IDF(t): Inverse of document frequency DF(t) -
number of documents that contain a term t.

Encoding Text Data
TF-IDF vectoriser:
● TF(t, d): High frequency of term in document increases weight of TF-IDF (local
parameter).
● IDF(t): Low document frequency of term in whole corpus increases weight of
TF-IDF (global parameter).
● TF-IDF: Considers both local and global impact of a word in vocabulary.
● Performs better than Count Vectorizer for feature extraction.

Encoding Text Data
Count Vectorizer
● Simple way to tokenize text and
build vocabulary of words.
● Directly gives frequency of token
w.r.t index of vocabulary.
TF-IDF Vectorizer
● Compares no. of times word
appears in document with no. of
documents containing that word.
● Assigns the value by considering
overall document of words,
penalizing the words having high
document frequency.

Problem transformation
● Transforms the problem into separate single class classification problem.
● Methods Used - Binary Relevance, Classifier Chains
● Binary Relevance: Each label is treated as separate single class classification
problem.
● Classifier Chains: First classifier is trained on the input, then subsequent
classifiers are trained on input + all previous classifiers in the chain.

Binary Relevance:
Drawback: Doesn’t consider correlation between the labels, treats every target
variable as independent.
Four different single-class
classiﬁcation problem

Classiﬁer Chains:
● Forms chains in order to preserve label correlation.
● Generally performs better than binary relevance when correlation between
labels is involved.
Four different single-label problem
Yellow: input space, White: target variable

Decision Trees
● A non-parametric supervised learning method used for classiﬁcation and
regression.
● The goal is to create a model that predicts the value of a target variable by
learning simple decision rules inferred from the data features.
● In decision analysis, a decision tree can be used to visually and explicitly
represent decisions and decision making
● Growing a tree involves deciding on which features to choose and what
conditions to use for splitting, along with knowing when to stop

Decision Trees
● Tree models where the target variable can take a discrete set of values are
called classiﬁcation trees
● In these tree structures, leaves represent class labels and branches represent
conjunctions of features that lead to those class labels
● Advantages:
○ Able to handle multi-output problems.
○ The cost of using the tree is logarithmic in the number of training data points.
○ Uses a white box model
○ Requires little data preparation

Decision Trees
Shortcomings:
● Decision-tree learners can create over-complex trees: overﬁtting
● The problem of learning an optimal decision tree is known to be NP-complete.
● There are concepts that are hard to learn because decision trees do not express
them easily, such as XOR, parity or multiplexer problems.
● Decision tree learners create biased trees if some classes dominate.
● Decision trees can be unstable because small variations in the data might result
in a completely different tree being generated.
Some of these problems can be mitigated by using decision trees with an ensemble.

Decision Trees Ensembling
● Ensemble methods, combines several decision trees to produce better
predictive performance than utilizing a single decision tree.
● A group of weak learners come together to form a strong learner!
● Ensembled Trees are composed of a large number of decision trees, where the
final decision is obtained taking into account the prediction of every tree.
● Specifically, by majority vote in classification problems, and by the arithmetic
mean in regression problems.
● All of these ensemble methods take a decision tree and then apply either
bagging (bootstrap aggregating) or boosting as a way to reduce variance and
bias.

Bagging
● Bagging (Bootstrap Aggregation) is is a general technique for combining the
predictions of many models
● It is used when our goal is to reduce the variance of a decision tree.
● The idea is to create several subsets of data from training sample chosen
randomly with replacement.
● Now, each collection of subset data is used to train their decision trees.
● As a result, we end up with an ensemble of different models.
● Combination of all the predictions from different trees are used which is more
robust than a single decision tree.

Random Forest
● Fitting decision trees on bootstrapped data serves to decorrelate them slightly.
● Since each node chooses what feature to split on in a greedy manner, trees can
still end up very correlated with each other.
● Random forests decorrelates the individual decision trees.
● When the CART algorithm is choosing the optimal split to make at each node,
the random forest will choose a random subset of features and only consider
those for the split.
● From that random subset of features, the algorithm will still select the optimal
feature and optimum split to make at each node.

Extremely Randomized Trees
● Much more computationally efﬁcient than random forests.
● Performance is almost always comparable. In some cases, they may even
perform better!
● Adds one more step of randomization to the random forest algorithm.
● Random Forest chooses the optimum split to make for each feature while Extra
Trees chooses it randomly - so it is faster.
● Both of them will subsequently choose the best feature to split on by comparing
those chosen splits.
● From this reason comes the name of Extra Trees (Extremely Randomized
Trees)

Extremely Randomized Trees
● Random forest uses bootstrap replicas it subsamples the input data with
replacement.
● Extra Trees use the whole original sample.
● In the Extra Trees sklearn implementation there is an optional parameter that
allows users to bootstrap replicas, but by default, it uses the entire input
sample.
● These differences motivate the reduction of both bias and variance.
● Using the whole original sample instead of a bootstrap replica will reduce bias.
● Choosing randomly the split point of each node will reduce variance.

Boosting
● Boosting is another ensemble technique to create a collection of predictors.
● In this technique, learners are learned sequentially.
● Early learners are fitted with simple models to the data. Then data is analysed
for errors.
● In other words, we fit consecutive trees (random sample) and at every step, the
goal is to solve for net error from the prior tree.
● When an input is misclassified by a hypothesis, its weight is increased so that
next hypothesis is more likely to classify it correctly.
● By combining the whole set at the end converts weak learners into better
performing model.

Gradient Boosting
● Gradient Boosting is an extension over boosting method.
● Gradient Boosting = Boosting + Gradient Descent
● It uses gradient descent algorithm for optimizing the given loss function.
● An ensemble of trees are built one by one and individual trees are summed
sequentially.
● Next tree tries to recover the loss (difference between actual and predicted
values).
● It can supports any differential loss function.

Extreme Gradient Boosting
● XGBoost is like gradient boosting on ‘steroids’!
● XGBoost improves upon the base GBM framework through systems
optimization and algorithmic enhancements.

System Optimization:
● Parallelization: XGBoost approaches the process of sequential tree building in
parallelized approach using all of the CPU cores during training.
● ‘Out-of-core’ computing: optimize available disk space while handling very
large datasets that do not fit into memory.
● Cache Optimization of data structures and algorithm to make best use of
hardware.
● Tree Pruning: XGBoost uses ‘max_depth’ parameter instead of negative loss
criterion, and starts pruning trees backward. This ‘depth-first’ approach
improves computational performance significantly.

Algorithmic Enhancements:
● Regularization: It penalizes more complex models through both LASSO (L1) and
Ridge (L2) regularization to prevent overfitting.
● Sparsity Awareness: XGBoost handles different types of sparsity patterns in
the data as well as missing values more efficiently.
● Weighted Quantile Sketch: XGBoost employs the distributed weighted
Quantile Sketch algorithm to effectively find the optimal split points among
weighted datasets.
● Cross-validation: The algorithm comes with built-in cross-validation method at
each iteration, taking away the need to explicitly program this search .

Recurrent Neural Networks
● They are the Neural Networks where the output from previous step are
fed as input to the current step.
● Have a hidden layer which remembers some information about a
sequence.

LSTM
● Vanilla RNNs face vanishing gradients. To eliminate vanishing gradient problem
and capture long term dependencies LSTM(Long Short Term Memory) models
were developed.
● LSTMs use three gates :
○ Input Gate
○ Output Gate
○ Forget Gate
● With the help of these gates and Cell state LSTMs are able to preserve past
information of sequential data.

Bi-LSTM
● Bi-LSTM couples two LSTM such that one LSTM is trained on sequence in
direction of increasing index and the other is trained in the direction of
decreasing index.
● Using the hidden states of both LSTM, Bi-LSTM uses both past and future
context to give the output.

Word Embeddings
● Learned representation for text where words that have the same meaning have
a similar representation.
● Low-dimensional and dense. Number of features is much smaller than than size
of vocabulary.
● Vector values are learned in a way that resembles a neural network.

Pretrained Embeddings
● Unstructured textual data needs to be processed into some structured,
vectorized formats which can be understood and acted upon by machine
learning algorithms.
● Obtained by unsupervised training of a model on a large corpora of text.
● Captures the syntax and semantic meaning of a word by training on a large
corpora.
● Increases efﬁciency of the model.
● Several pre trained word-embeddings are present namely, Word2Vec, FastText,
Glove.

Word2Vec
● Statistical method for efﬁciently learning a standalone word embedding from a
text corpus.
● King - Man + Woman = Queen
● Two popular models are SkipGram and CBOW.

Word2Vec
CBOW Model
● Learns the embeddings by predicting the current word based on its context.
● Trains faster and works better for frequent words.
SkipGram Model
● Learns by predicting the surrounding words given a current word.
● Works well with small data and can represent rare words.

FastText
● Represents each word as n-gram of characters to allow the embeddings to
understand its preﬁxes and sufﬁxes. E.g., the word vector, apple, could be
broken down into separate word vectors units as ap, app, ple.
● Uses skip-gram model for learning embeddings.
● Works well on rare words as the word in broken down in n-grams to get its
embeddings.
● Even works well on the words not present in the vocabulary because the
n-gram character vectors are shared with other words.

Glove
● Stands for Global Vectors.
● An unsupervised learning algorithm for obtaining vector representations for
words.
● Learns by constructing a co-occurrence matrix that counts how frequently a
word appears in a context.
● Performs better on word analogy and named entity recognition.
● Does not rely only on local statistics (local context information of words), but
also incorporates global statistics (word co-occurrence) to obtain word
vectors.

Receiver Operating Characteristic
● An ROC curve is a graph showing the performance of a classification model at
all classification thresholds.
● This curve plots two parameters:
○ True Positive Rate (TPR) : Fraction of positive data points classified as positive.
○ False Positive Rate (FPR) : Fraction of negative data points classified as positive.

Receiver Operating Characteristic
● An ROC curve plots TPR vs. FPR at different classification thresholds.
● Lowering the classification threshold classifies more items as positive, thus
increasing both False Positives and True Positives.

Area Under the ROC Curve
● AUC measures the entire two-dimensional area underneath the entire ROC
curve from (0,0) to (1,1).
● More the area under curve better the classiﬁer!
● AUC provides an aggregate measure of performance across all possible
classiﬁcation thresholds.
● AUC can be interpreted as the probability that the model ranks a random
positive example more highly than a random negative example.
● AUC ranges in value from 0 to 1. A model whose predictions are 100% wrong
has an AUC of 0.0; one whose predictions are 100% correct has an AUC of 1.0.

Area Under the ROC Curve
● AUC is desirable for the following two reasons:
○ AUC is scale-invariant. It measures how well predictions are ranked, rather than their absolute
values.
○ AUC is classiﬁcation-threshold-invariant. It measures the quality of the model's predictions
irrespective of what classiﬁcation threshold is chosen.

Results and Comparison
Model Mean AUC_ROC Score
Support Vector Machines (Binary Relevance) 0.66
Support Vector Machines (Classiﬁer Chains) 0.67
Logistic Regression (Binary Relevance) 0.73
Logistic Regression (Classiﬁer Chains) 0.76
Extra Trees 0.93
XGBoost 0.96
LSTM without pretrained embeddings 0.97
LSTM with FastText embedding 0.96
LSTM with Glove embedding 0.88
LSTM with Word2Vec embedding 0.85

Conclusion and Further Works
● Classical Models like SVM and Logistic Regression fail to achieve high
AUC_ROC score.
● Classiﬁer Chain method has a slight edge over Binary Relevance.
● Tree based ensembling methods and Recurrent Neural networks can achieve a
very high AUC_ROC score.
● Transfer learning reduces the execution time but does not help with
performance.
● Further Improvements can be achieved by experimenting with complicated
neural networks and use of state of the art Transformers.

Hate Speech / Toxic Comment Detection - Data Mining (CSE-362) Project

More Related Content

Similar to Hate Speech / Toxic Comment Detection - Data Mining (CSE-362) Project (20)

Recently uploaded (20)

Hate Speech / Toxic Comment Detection - Data Mining (CSE-362) Project