Deep Learning for NLP - Part 9

Name: Deep Learning for NLP - Part 9
Rating: 5.0 (2 reviews)

Hate Speech Detection

Created byManish Gupta

Last updated 9/2021

English

English [Auto],

What you'll learn

Deep Learning for Natural Language Processing
Hate Speech Detection
DL for Hate Speech Detection
Multimodal Hate Speech Detection
Analysis of hate speech detection results
DL for NLP

Course content

1 section • 12 lectures • 2h 15m total length

Introduction2:25
Survey hate speech detection, covering datasets, feature-based NLP methods, and deep learning approaches, including multimodal text-image analysis. Discuss interpretability, challenges, and limitations in current mechanisms.
Why is hate speech detection important?11:19
Explore why hate speech detection matters, drawing on an ADL survey of online harassment across protected classes, social media impact, and legal commitments to curb abuse and prevent real-world harm.
Hate speech datasets11:47
Explore hate speech datasets across social networks, news portals, and Wikipedia, distinguishing toxic language from offensive language, and including multimodal and antisemitism-focused corpora for robust detection.
Feature based approaches7:41
Explore feature based traditional machine learning for hate speech prediction, using dictionaries, bag of words, POS, dependency parsing, embeddings, topic models, and various classifiers including logistic regression, SVM, and MLP.
Deep learning methods - 114:30
Explore early deep learning methods for hate speech detection, from CNNs with random and glove embeddings to multi-modal two-tower models using text and metadata features, with interleaving training achieving results.
Deep learning methods - 214:35
Explore deep learning methods for hate speech detection, focusing on data augmentation techniques—oversampling, word-level and paraphrase-based methods, and conditional generation—and strategies to mitigate bias through explanation-driven regularization.
Deep learning methods - 321:35
Explore deep learning methods for hate speech detection, including character-level adversarial robustness using word and character embeddings, multilevel classification with diverse embeddings and specialized losses, and hierarchical conditional variational models.
Multimodal hate speech detection - 117:42
Examine multimodal hate speech detection across Instagram cyberbullying datasets by fusing text and image features through supervised and unsupervised deep learning, graph-based representations, and dynamic multimodal fusion.
Multimodal hate speech detection - 215:11
Explore multimodal hate speech detection using Amaechi 150k and hateful memes, comparing text-only vs image-only vs multimodal fusion with ocr and inception v3 features, noting text-only models often perform best.
Analysis of hate speech detection results9:16
Analyze hate speech detection results, showing a long tail with low uniqueness in hate tweets, and explore interpretation methods using embeddings, attention, and saliency analyses.
Challenges and limitations5:38
Explore challenges and limitations of hate speech detection, including annotation disagreement, culture-sensitive definitions, evolving language, and interpretability, plus multimodal labeling and adversarial attacks.
Summary3:57
Explore traditional and deep learning methods for hate speech detection, including CNNs, multitask losses, multilevel embeddings, and multimodal fusion of text, metadata, and images with image captioning and sentiment analysis.

Requirements

Basics of machine learning
Basic understanding of deep learning models

Description

Since the proliferation of social media usage, hate speech has become a major crisis. On the one hand, hateful content creates an unsafe environment for certain members of our society. On the other hand, in-person moderation of hate speech causes distress to content moderators. Additionally, it is not just the presence of hate speech in isolation but its ability to dissipate quickly, where early detection and intervention can be most effective. Through this course, we will provide a holistic view of hate speech detection mechanisms explored so far.

In this course, I will start by talking about why studying hate speech detection is very important. I will then talk about a collection of many hate speech datasets. We will discuss the different forms of hate labels that such datasets incorporate, their sizes and sources. Next, we will talk about feature based and traditional machine learning methods for hate speech detection. More recently since 2017, deep learning methods have been proposed for hate speech detection. Hence, we will talk about traditional deep learning methods. Next, we will talk about deep learning methods focusing on specific aspects of hate speech detection like multi-label aspect, training data bias, using metadata, data augmentation, and handling adversarial attacks. After this, we will talk about multimodal hate speech detection mechanisms to handle image, text and network based inputs. We will discuss various ways of mode fusion. Next, we will talk about possible ways of building interpretations over predictions from a deep learning based hate speech detection model. Finally, we will talk about challenges and limitations of current hate speech detection models. We will conclude the course with a brief summary.

Who this course is for:

Beginners in deep learning
Social science students with an inclination towards data science
Humanities students
Python developers interested in data science concepts
Masters or PhD students who wish to learn deep learning concepts quickly
Deep learning engineers and developers
Employees of Social media companies