Enhanced sentiment classification of diverse Turkish texts using hierarchical gated attention and recurrent neural architectures

Sentiment classification is a critical task in natural language processing, with growing importance in areas such as social media analysis, customer feedback, and public opinion monitoring. With recent advancements in natural language processing and deep learning, techniques like neural networks and transformers have become increasingly significant in sentiment classification. Sentiment classification on Turkish texts presents distinctive linguistic challenges due to the language’s rich morphology and agglutinative structure. This study specifically focuses on these challenges and highlights embedding methods, attention networks and mechanisms, and Recurrent Neural Networks (RNNs). The study explores how character and word embeddings enhance model performance and text understanding, how attention mechanisms highlight and focus on important parts of the input, and how Long Short-Term Memory (LSTM) networks effectively capture long-term dependencies within the text. A hybrid model is developed by integrating LSTM and Hierarchical Gated Attention Network to analyze sentiments in short, unstructured and long, structured Turkish texts. In order to facilitate comparison of the performance of the proposed model, a new Turkish dataset has been created and labeled using Twitter data. Hierarchical Gated Attention Network performs a dual role by providing emphasis on significant information at the word and sentence levels, therefore better capturing the hierarchical nature of the text. Coupling this with the strong capacity of LSTM, the resulting framework is capable enough to conduct sentiment classification. The experimental results indicate that combining these techniques enhances sentiment classification accuracy while enabling a more nuanced and effective contextual representation of Turkish text data. The proposed hybrid model attained an accuracy of 81% and an F1-score of 72.04% on short, unstructured texts, whereas it reached 93.97% accuracy and an F1-score of 88.01% on long, structured documents.