LSTM beats Random Forest and Naive Bayes at tweet sentiment analysis with 0.92 ROC-AUC

A new study published on arXiv in July 2026 tested how far an LSTM neural network outperforms five classical machine learning algorithms at sentiment analysis of Twitter posts. The recurrent network won on every key metric, scoring 80% accuracy on the test set and a ROC-AUC of 0.92.

What was compared, and how

The authors took a Kaggle dataset of tweets and applied standard NLP preprocessing: tokenization (splitting text into words), lemmatization (reducing words to their base form) and stop-word removal. Six algorithms were then trained to classify tweets into three categories: positive, negative and neutral.

The models tested:

• Logistic regression
• Random Forest
• Naive Bayes classifier
• Gradient Boosting
• LSTM (Long Short-Term Memory), a recurrent neural network that accounts for word order

How well did the LSTM perform?

The LSTM was the best of all the approaches tested:

• Training accuracy: 90.98%
• Test accuracy: 80.00%
• Micro-average ROC-AUC: 0.92

A ROC-AUC of 0.92 means the model correctly separates positive from negative content in roughly 92% of cases. That is a strong result for noisy social media data full of slang and irony. The gap between 90.98% (training) and 80% (test) points to moderate overfitting, typical of neural networks without additional regularization.

"The results show that LSTM outperforms conventional machine learning

methods at capturing the contextual and sequential aspects of text," the authors note.

Classical algorithms train faster and are easier to interpret, but they treat text as a set of independent words, ignoring order and context. An LSTM, by contrast, "remembers" the previous words as it processes each new one. That is why it understands that "not bad" is not "bad", while "didn't like it at all" carries strong negative sentiment.

Why this matters for social media analysis

Twitter (now X) generates hundreds of millions of posts every day. Automated sentiment analysis lets brands track audience reaction to a product in real time, and lets researchers study shifts in public opinion without manually reading millions of messages.

Modern commercial systems rely on transformer models such as BERT and RoBERTa, which significantly outperform LSTM but demand substantial computing resources. LSTM remains a practical choice for teams on a limited budget: it is lighter, faster at inference and works well enough without a powerful GPU cluster.

What this means

The work confirms a consistent pattern in NLP: for text where word order and context matter, deep learning reliably beats classical ML. The experiment is easy to reproduce. An open Kaggle dataset and standard libraries (TensorFlow, PyTorch) make it possible to run a similar pipeline in a few hours.

Frequently asked questions

How does LSTM differ from BERT in sentiment analysis?

LSTM processes text sequentially, word by word, keeping the earlier context in memory. BERT is a transformer model that analyzes the whole text at once in both directions and is pretrained on huge corpora. On most modern benchmarks BERT beats LSTM by 5-15 percentage points, but it requires considerably more computing resources.

What is ROC-AUC, and why is 0.92 a good result?

ROC-AUC shows how well a model separates classes: 0.5 is random guessing, 1.0 is perfect classification. A value of 0.92 is considered high for sentiment analysis on social media, where data is noisy with slang, typos and sarcasm.