A Comparative Analysis of Text Normalization Techniques for Enhanced Sentiment Analysis Performance

Sentiment analysis on Indonesian social media, particularly on platforms like Twitter, faces challenges due to the prevalence of informal and non-standard language. This variation complicates natural language processing (NLP) tasks, necessitating robust language normalization techniques. This study evaluates five normalization methods Jaro, Jaro Winkler, Levenshtein, Damerau-Levenshtein, and Smith-Waterman by comparing their effectiveness in enhancing sentiment classification accuracy on Indonesian Twitter data. The methods are tested against a standard word set (KBBI) and sentiment-labeled review data from the DANA application. Results indicate that Damerau-Levenshtein achieves the highest accuracy of 89.15% in classifying sentiments into positive and negative categories. Meanwhile, the Jaro method demonstrates strong consistency in handling errors like duplication and character insertion, albeit with slightly lower accuracy. These findings highlight the critical role of language normalization in improving NLP performance and provide actionable insights for developers working on sentiment analysis systems in the context of Indonesian social media.