De-Obfuscating Hate: A Specialized Spell Corrector for Social Media Cyberbullying Detection

TL;DR

Social media is a "Wild West" of linguistics where users intentionally mangle words to bypass filters. This paper presents a specialized spell checker that uses ten targeted heuristic approaches—ranging from removing repeated "heeeeey" characters to substituting visual mimics like "@" for "a". By achieving ~90% correction accuracy, it significantly boosts the performance of downstream bullying detection models.

Problem & Motivation: The Intentional Misspelling Challenge

In digital safety, the "Mute Word" feature is a common defense. However, bad actors are clever; they use Non-word errors (like "siht" for "shit") and Visual substitutions (like "bi+(h" for "bitch") to evade censorship.

Traditional spell checkers fail here because:

1. They assume errors are accidental (fat-finger syndrome), not adversarial.
2. They cannot handle the massive word lengths caused by "elongation" (e.g., "weeeekeeend").
3. The search space in a full dictionary is too computationally expensive for real-time filtering.

Methodology: A Ten-Heuristic Arsenal

The authors don't rely on a single algorithm. Instead, they deploy a pipeline that mimics how humans perceive misspelled text.

1. The Preprocessing Shield

Before algorithmic correction, the system normalizes "chat speak" (e.g., "ppl" → "people") and converts emoticons to sentiment-bearing words (e.g., "<3" → "love").

2. The Core Correction Approaches

The methodology is split into five functional categories:

• Deletion: Handles character elongation ("hellooooo") and plural errors.
• Insertion (The Star Performer): The Adding Character Approach (ACA) proved most effective, especially for words where vowels were removed (e.g., "goin" → "going").
• Transformation: Replaces similar shapes (Group: p, q, g, y) or similar sounds (Group: k, c, ck, ke).
• Separation: Splits merged phrases like "thatawkwardmoment" into "that awkward moment".
• Reordering: Corrects anagrams and swapped neighbors ("spiol" → "spoil").

Fig 1: The Four-Step Correction Pipeline involving indexed lookup and heuristic selection.

3. Smart Candidate Selection

If multiple heuristics suggest different corrections, the system calculates the Levenshtein Distance between the original noise and the dictionary candidates, picking the one with the smallest edit cost.

Experiments & Results: Real-World Impact

The system was tested on the Formspring.me and Twitter Sentiment Analysis datasets.

Interestingly, the impact on classification varied by model. For Twitter, the CNN Classifier saw a jump from 95.14% to 96.10% accuracy after correction. However, for Formspring, Naive Bayes performed better. This suggests that the "contextual" nature of CNNs (looking at 3-word windows) benefits more from the "un-merging" of words than simple Bag-of-Words models.

Table: Success rates per category. Deletion and Insertion dominate social media corrections.

Critical Analysis & Conclusion

Takeaways

The paper proves that "Cleaning is as important as Modeling." By specifically targeting the intentional ways people bypass filters, the authors created a tool far more useful for online safety than a standard Microsoft Word-style spell checker. Notably, the Adding Character Approach (ACA) is the unsung hero of social media NLP, as users often drop characters for speed.

Limitations

• The First Letter Bias: Most heuristics assume the first letter is correct. In highly adversarial "leetspeak," this might not hold.
• Compound Errors: The system struggles when a word has both sound and shape errors simultaneously due to computational complexity.
• Pseudonyms: Random handles without the "@" symbol are still often misidentified as errors.

Future Work

The authors aim to incorporate synonym detection and better prediction of when a dataset actually needs correction versus when the original "noise" might carry its own unique dialectal information.