Background: Sentiment classification of short-form social media text remains challenging due to informal grammar, sarcasm, and limited contextual length, with classical machine learning approaches and even standalone deep learning architectures often struggling to jointly capture local lexical patterns and longer-range sequential dependencies ^{[1, 2]}.

Objective: This study proposed and evaluated a hybrid convolutional neural network–long short-term memory (CNN-LSTM) architecture for three-class (negative, neutral, positive) sentiment classification, benchmarked against four baseline models: logistic regression, random forest, XGBoost, and a standalone bidirectional LSTM.

Method: A labeled corpus of short-form social media posts was used to train and evaluate all five models under 10-fold cross-validation. This study uses a simulated dataset created for academic training purposes; all performance metrics, training curves, and confusion matrix values were generated to reflect plausible patterns consistent with the published text-classification literature and do not represent results from an actual trained model or real annotated corpus. Model performance was evaluated using accuracy, precision, recall, and F1-score, with the hybrid architecture combining convolutional feature extraction with sequential LSTM encoding.

Key Results: The proposed hybrid CNN-LSTM model achieved a mean F1-score of 0.92 across 10-fold cross-validation, outperforming the bidirectional LSTM (0.87), XGBoost (0.85), random forest (0.81), and logistic regression (0.74) baselines. Training and validation loss curves showed stable convergence within approximately 20 epochs without evidence of severe overfitting. The confusion matrix on the held-out test set indicated the highest classification accuracy for the positive and negative classes, with comparatively more confusion involving the neutral class.