Skip to content

Hate Speech Detection

Detect hate speech across 6 target categories with severity levels and leetspeak evasion handling.

Why hate speech detection?

Content moderation for social media platforms, forum management, brand safety, and research on online toxicity. Malaysian hate speech is particularly challenging because it uses code-switching, local slurs, religious references, and leetspeak to evade basic filters.

Load module

import malaysian_manglish_nlp as mnlp

result = mnlp.detect_hate_speech("Text to check here")
print(result)
# {'is_hate': False, 'severity': 'none', 'categories': [], 'score': 0.12}

Basic usage

Non-hate speech

mnlp.detect_hate_speech("Sedap gila nasi lemak kat sini!")
# {'is_hate': False, 'severity': 'none', 'categories': [], 'score': 0.05}

mnlp.detect_hate_speech("Aku tak setuju dengan polisi kerajaan")
# {'is_hate': False, 'severity': 'none', 'categories': [], 'score': 0.15}

Criticism is not hate speech

The module distinguishes between legitimate criticism/opinion and actual hate speech. Political disagreement, product complaints, and negative reviews are not flagged.

Detected hate speech

result = mnlp.detect_hate_speech("Offensive text here")
# {'is_hate': True, 'severity': 'high', 'categories': ['race'], 'score': 0.92}

6 target categories

Category Description
race Racist content targeting ethnicity
religion Content attacking religious groups
gender Sexist or misogynistic content
nationality Xenophobic content targeting nationality
disability Content mocking disabilities
sexual_orientation Homophobic or transphobic content

Check severity

mnlp.get_severity("text here")
# 'none' | 'low' | 'medium' | 'high'

Check target groups

mnlp.get_target_groups("text here")
# [] or ['race', 'religion', ...]

Severity levels

Level Description Action
none Clean text No action
low Mildly offensive, borderline Flag for review
medium Clearly offensive Moderate
high Severe hate speech Remove immediately 
# Quick binary check
mnlp.is_hate("some text")
# True or False

Leetspeak and evasion detection

Malaysians sometimes use leetspeak or spelling tricks to evade filters. malaysian-manglish-nlp detects common patterns:

# Number substitution
# "b0d0h" instead of "bodoh"
# Racial slurs with @, $, 0 substitutions

# Deliberate misspelling
# Split words, extra characters

# Code-switched slurs
# Mixed BM/EN offensive phrases

The module normalizes text internally before detection:

# These evasion attempts are caught:
# - Character substitution (0→o, @→a, $→s, 1→i)
# - Deliberate spacing ("b o d o h")
# - Repeated characters ("booodohh")
# - Abbreviated slurs

Batch processing

texts = [
    "Sedap gila makanan ni!",
    "Normal sentence here",
    "Another clean text",
    "Offensive content here",
]

results = mnlp.hate_detect_batch(texts)
for text, result in zip(texts, results):
    status = "🚫 HATE" if result['is_hate'] else "✅ Clean"
    print(f"{status} | {text[:40]}")

Integration with content moderation

def moderate_comment(text):
    """Example moderation pipeline."""
    # Check hate speech
    hate = mnlp.detect_hate_speech(text)

    if hate['severity'] == 'high':
        return {'action': 'block', 'reason': 'hate_speech'}

    if hate['severity'] == 'medium':
        return {'action': 'flag', 'reason': 'review_needed'}

    # Check profanity too
    profanity = mnlp.detect_profanity(text)
    if profanity['severity'] == 'high':
        return {'action': 'censor', 'censored': mnlp.censor(text)}

    return {'action': 'approve'}

# Usage
moderate_comment("Normal friendly comment")
# {'action': 'approve'}

CLI usage

# Full analysis (includes hate speech check)
$ mnlp analyze "text to check"

# With JSON
$ mnlp analyze "text" --json

How it works

  1. Normalization - leetspeak decoding, repeated char reduction
  2. Lexicon matching - curated hate speech dictionaries for Malaysian context
  3. Pattern detection - known hate speech phrases and structures
  4. Context analysis - distinguish insults from hate speech
  5. Severity classification - score determines none/low/medium/high
  6. Category assignment - which target group(s) are attacked

Performance

Metric Score
Binary F1 86.3%
Severity accuracy 79.8%
Category F1 (macro) 74.5%
Leetspeak detection 82.1%
False positive rate 3.2%
Throughput 18,000 texts/sec

False positives

3.2% false positive rate means some legitimate text gets flagged. Always provide an appeal mechanism in production moderation systems.

See also