A comment on social media may appear harmless at first glance. However, sarcasm, coded expressions, altered spellings, or ordinary words used in certain situations can carry messages of hatred that are not immediately obvious. Identifying this type of content is difficult not only for human moderators but also for computer systems.

Hate speech refers to words or content that insult, attack, or spread hatred against a person or group. It is linked to prejudice, intolerance and conflict between communities and can spread quickly through social media, deepen social divisions and affect the psychological well-being of those targeted.

Why Is Hate Speech Difficult to Monitor?

Billions of users post content on social media every day. The amount of text, images, videos and comments uploaded within a short period makes it almost impossible for human moderators to review everything manually.

Social media platforms always depend on users to report harmful content. However, the reporting and review process may take time. By the time a harmful post is removed, it may already have been viewed, shared or copied by many users.

Another major challenge is that there is no universally accepted definition of hate speech. An expression viewed as offensive in one country or culture may be regarded as freedom of expression in another.

Hate speech is also sometimes grouped with cyberbullying, offensive language and profanity, although these terms do not carry the same meaning. For example, a rude comment may be offensive without targeting a person based on identity or group membership.

How Machine Learning Helps

Machine learning allows large amounts of online content to be processed automatically. Instead of checking every post individually, computer models are trained to recognise language patterns associated with harmful or hateful content.

Researchers have applied different machine learning approaches to hate speech detection, such as classical machine learning, ensemble learning and deep learning.

Early approaches relied on classical machine learning methods that examine words, term frequencies and language patterns associated with hateful content. These methods are easy to implement, although they are less effective when hate speech is expressed indirectly.

Ensemble learning combines the decisions of multiple models to produce a final prediction, which may provide more consistent results than using one model alone.

More recent studies have used deep learning. These models examine the relationships between words and the overall meaning of a sentence. This allows them to identify less obvious forms of hate speech expressed through stereotypes, sarcasm, figurative language or indirect references.

However, even an advanced model does not truly understand language in the same way as a human being. It makes predictions based on patterns learned from the data used during training.

The Challenge of Subtle and Changing Language

Online hate speech does not always use direct insults. Users may change spellings, replace letters with symbols or use coded terms to avoid detection. As language changes, systems trained using older data may fail to recognise newer slang and emerging forms of online abuse (Yin & Zubiaga, 2021).

Sarcasm is another major challenge. A sentence may appear positive when read literally, but communicate the opposite meaning in context. Cultural references, humour and local expressions can also be difficult for a computer model to interpret correctly.

This challenge is relevant in multilingual environments such as Malaysia, where users may combine Bahasa Melayu, English, dialects, abbreviations and informal expressions in the same sentence.

Can a Model Work Everywhere?

The performance of a hate speech detection model depends on the data used during training. A model may perform well when tested using data similar to its training examples but become less accurate when applied to content from another platform, topic, country or community. This problem is known as poor generalisation (Yin & Zubiaga, 2021).

For example, a model trained on political discussions or content from one country may perform poorly when applied to different topics or cultural settings. A high accuracy score, therefore, does not guarantee effective performance in every real-world situation.

Technology Alone Is Not Enough

Automated detection systems can help platforms identify potentially harmful content more quickly. However, they should not be treated as a complete solution. A system that is not carefully developed may remove legitimate content, misunderstand humour or unfairly affect certain communities.

Machine learning can help human moderators identify suspicious content, but difficult cases still require human judgement. Effective moderation depends on clear community guidelines and cooperation between computer scientists, linguists, social scientists and affected communities.

As online communication continues to grow, hate speech detection systems must become more accurate, fair and sensitive to cultural differences. Artificial intelligence can help identify harmful content, but it should support human judgment rather than replace it. Responsible detection requires reliable technology, clear policies and careful consideration of language, culture and freedom of expression.

Prepared by: Dr. Wan Noor Hamiza Wan Ali, Senior Lecturer, Faculty of Artificial Intelligence, Universiti Teknologi Malaysia