Enhancing Multi-Domain Sentiment Classification Using Adaptive Lexicons and Machine Learning
Abstract
Multi-domain sentiment classification remains challenging because lexical polarity and expression patterns vary across application contexts. This study evaluates an enhanced iVADER framework that combines domain-specific lexicon adaptation with text preprocessing and machine-learning-supported feature representations. Five public English-language datasets were analyzed: Amazon Cells Labelled (1,000 records), Yelp Labelled (1,000), Tweets with Sarcasm and Irony (19,000), Apple Twitter Sentiment Texts (4,500), and ChatGPT Tweets (2,500), yielding 28,000 records. A stratified 80:20 split produced 22,400 training records and 5,600 test records. Preprocessing included noise and URL removal, case folding, tokenization, stopword removal, WordNet lemmatization, stemming, and negation handling. The framework expanded domain-specific lexical information and combined it with TF-IDF, n-grams, bag-of-words, and embedding-based representations. Enhanced iVADER achieved pooled accuracy of 92%, precision of 90%, recall of 89%, and an F1-score of 0.91, compared with 85%, 82%, 75%, and 0.80 for conventional VADER. Across the five datasets, accuracy ranged from 87% to 94%, with an unweighted mean of 91.0% and a sample standard deviation of 2.74 percentage points. The results indicate that domain-adaptive lexical information can improve VADER-based sentiment classification across the observed domains while retaining an interpretable scoring process.