Deep learning methods for text analysis

Sumit Dubey

Diplomová práce

Abstract:

This study explores how deep learning techniques enhance sentiment analysis, focusing on metrics like accuracy, precision, recall, and F1-score. It evaluates how advanced feature extraction, hyperparameter tuning, and LSTM architectures improve model performance. The research compares an improved model with multiple LSTM layers and regularization techniques to a baseline model with a single LSTM layer. Data preprocessing techniques like tokenization and vectorization split customer review data into training and testing sets. Both models are trained for 150 epochs using the Adam optimizer. Evaluation is done through learning curves, F1-score, accuracy, and confusion matrix analysis. The improved model shows better accuracy, generalization, and reduced overfitting, making it more suitable for real-world sentiment analysis applications. The study contributes to advancing deep learning-based text analysis.

Keywords

Sentiment Analysis, LSTM Networks, Transformers & BERT, Text Classification, Named Entity Recognition (NER)

Jazyk práce: angličtina

Datum vytvoření / odevzdání či podání práce: 30. 3. 2025

Obhajoba závěrečné práce

• Vedoucí: doc. Ing. Arnošt Veselý, CSc.
• Oponent: Richard Vágner, externi

Česká zemědělská univerzita v Praze

Provozně ekonomická fakulta

Magisterský studijní program:
Informatics