Small-scale sentiment classification suffers from data scarcity, which limits model generalization. This study systematically compares three text augmentation strategies under a controlled, reproducible framework.

• EDA (Easy Data Augmentation) based on token-level synonym replacement, insertion, deletion, and swapping.
• Back-translation using English↔French MarianMT models to create high-fidelity paraphrases.
• Contextual token substitution (NLPaug-style) with pre-trained language models for semantics-preserving edits.

Methods: Experiments use a 5,000-sample IMDb subset with 100% augmentation, 10-fold cross-validation, and fixed seeds, comparing traditional classifiers (Logistic Regression, Random Forest) and a fine-tuned BERT base model on accuracy, F1, AUC, and effect sizes.

• All augmentation strategies yield significant and statistically robust performance gains over non-augmented baselines.
• Contextual augmentation delivers the most consistent improvements for BERT, reaching about 97% test accuracy on the augmented setting.
• EDA and back-translation provide larger relative gains for traditional models, especially Random Forest, while exhibiting different diversity–cost trade-offs.

Lip reading plays a crucial role in applications such as speech recognition, assistive technologies for the hearing-impaired, and biometric authentication. However, performance often degrades when the speaker varies, the environment changes, or the speech sounds are visually similar. To overcome these challenges, we propose LipLingo:

• Based on the LipNet architecture
• Enhanced with standardized preprocessing for consistent mouth-region representation
• Uses rotational validation for better generalization
• Combines 3D convolutional layers (spatio-temporal features) and bidirectional recurrent layers (sequence modeling)
• Optimized with Adam optimizer and CTC loss

Results:

• 95% character-level accuracy
• 87% sentence-level accuracy
• Outperforms baseline LipNet