Comparative Analysis of Embedding Strategies for Turkish Sentiment Analysis

Ayaulym Parmash, Myrzaliyeva Zhibek, Selcuk Cankurt

2026

Closed access Metadata (CSL-JSON)
DOI 10.1109/icecco67619.2026.11488770
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Sentiment analysis for the Turkish language is particularly challenging due to the language's agglutinative nature, its complex morphology, and the relative paucity of high-quality labeled datasets compared to English. This paper presents a comprehensive comparative study of different feature representation methods and classification algorithms for Turkish sentiment analysis. We benchmark a traditional machine learning baseline-TF-IDF combined with Random Forest-against several Long Short-Term Memory (LSTM) deep learning models. To address lexical sparsity, we examine four embedding strategies: randomly initialized embedding layers, custom Word2Vec embeddings trained on our corpus, and pre-trained Turkish FastText models in both trainable and non-trainable configurations. Our experiments, conducted on three publicly available datasets, demonstrate that LSTM models consistently outperform the statistical baseline, providing a significant accuracy boost over TF-IDF. Notably, the non-trainable pre-trained Turkish FastText embeddings achieved the best performance, reaching a peak accuracy of 95.33%. These results indicate that leveraging large-scale external resources is particularly effective for morphologically rich languages, compared to relying solely on task-specific data. Moreover, freezing the weights of a pre-trained model can help prevent overfitting, yielding more robust generalization. This study provides a clear roadmap for researchers and practitioners in selecting the most effective embedding models for Turkish natural language processing tasks.