Vietnamese Context-Sensitive Malicious Spelling Error Correction

Spelling errors targeting specific keywords that users intentionally generate has seriously degraded the performance of social media control systems. In this paper, we show the severe effect of those misspellings and propose using a spelling correction approach for those targeted words based on context called word embedding. The data that we use within the limits of our work are Vietnamese spam email and hate speech. Also, we introduce a new and effective way to extract real misspellings to create reasonably synthetic data provided for our experiments. Our correction system results in a favorable performance on both synthetic and real data compared to Google.