Comparative Study of Sentiment Analysis Techniques: Traditional Machine Learning vs. Deep Learning Approaches
Grace Diri
, Philip Obiorah, Hongbo Du
, Philip Obiorah, Hongbo Du
InSITE 2025
• 2025
• pp. 23
Aim/Purpose
This paper presents a comparative analysis of sentiment classification models, focusing on the performance differences between lexicon-based sentiment analysis, traditional machine learning techniques, and transformer-based deep learning approaches across multiple benchmark datasets.
Background
While numerous methods exist for sentiment analysis, a systematic comparison across modeling paradigms and datasets remains limited. This study addresses this gap by evaluating representative models on IMDB, Yelp Polarity, and Amazon Polarity datasets using unified metrics and preprocessing pipelines.
Methodology
The evaluation includes (a) VADER as a lexicon-based baseline, (b) Logistic Regression, Decision Tree and Naive Bayes models built on three types of features: Bag of Words, TF-IDF, and Word2Vec as the traditional machine learning alternatives, and (c) two transformer models, Distil-BERT-SST-2 and RoBERTa-Sentiment, without fine-tuning as the deep learning methods. Performance is assessed using Accuracy, Precision, Re-call, and F1 Score.
Contribution
This study offers a comprehensive, side-by-side comparison of conventional and modern sentiment classification techniques, identifying model-feature synergies and highlighting limitations in generalizability across the domains.
Findings
Test results show that VADER, although efficient, underperforms on complex texts. Traditional models particularly Logistic Regression paired with TF-IDF yield strong performance.
Recommendations for Practitioners
DistilBERT demonstrates superior performance across all datasets, while RoBERTa’s performance suffers due to domain mismatch.
Recommendations for Researchers When computational resources are constrained, TF-IDF with Logistic Regression offers a competitive alternative to deep learning. For best performance, transformer models like DistilBERT should be prioritized.
Impact on Society
The study can improve applications in customer feedback systems, social media monitoring, and public opinion analysis, contributing to more responsive and adaptive services.
Future Research
Future work should explore domain-specific fine-tuning strategies and evaluate multilingual sentiment performance, especially for transformer models.
This paper presents a comparative analysis of sentiment classification models, focusing on the performance differences between lexicon-based sentiment analysis, traditional machine learning techniques, and transformer-based deep learning approaches across multiple benchmark datasets.
Background
While numerous methods exist for sentiment analysis, a systematic comparison across modeling paradigms and datasets remains limited. This study addresses this gap by evaluating representative models on IMDB, Yelp Polarity, and Amazon Polarity datasets using unified metrics and preprocessing pipelines.
Methodology
The evaluation includes (a) VADER as a lexicon-based baseline, (b) Logistic Regression, Decision Tree and Naive Bayes models built on three types of features: Bag of Words, TF-IDF, and Word2Vec as the traditional machine learning alternatives, and (c) two transformer models, Distil-BERT-SST-2 and RoBERTa-Sentiment, without fine-tuning as the deep learning methods. Performance is assessed using Accuracy, Precision, Re-call, and F1 Score.
Contribution
This study offers a comprehensive, side-by-side comparison of conventional and modern sentiment classification techniques, identifying model-feature synergies and highlighting limitations in generalizability across the domains.
Findings
Test results show that VADER, although efficient, underperforms on complex texts. Traditional models particularly Logistic Regression paired with TF-IDF yield strong performance.
Recommendations for Practitioners
DistilBERT demonstrates superior performance across all datasets, while RoBERTa’s performance suffers due to domain mismatch.
Recommendations for Researchers When computational resources are constrained, TF-IDF with Logistic Regression offers a competitive alternative to deep learning. For best performance, transformer models like DistilBERT should be prioritized.
Impact on Society
The study can improve applications in customer feedback systems, social media monitoring, and public opinion analysis, contributing to more responsive and adaptive services.
Future Research
Future work should explore domain-specific fine-tuning strategies and evaluate multilingual sentiment performance, especially for transformer models.
nature language processing, sentiment analysis, text classification, traditional machine learning, deep learning
54 total downloads
Back