Bidirectional encoder transformer based emotion analysis deep learning modeldevelopment

In this thesis, the BERT model, which is a new neural language model for Turkish sentiment analysis, was applied. Although natural language processing has recently progressed, representing compound meanings is a challenge. Traditional deep learning methods claim that sentences are an ordinary linear structure, that is, chains or sequences. In this thesis, an application has been developed by using a previously trained neural network in order to obtain more successful results for the Turkish sentiment analysis problem, taking into account the morphological structure of the words. BERT, a recent deep bidirectional demonstration of self-attention from unlabeled text, achieved remarkable results on many fine-tuned Natural Language Processing (NLP) tasks. This thesis, it is aimed to show the effectiveness of the BERT algorithm for Turkish, which is a morphologically rich language. Morphologically rich languages require intensive language preprocessing steps to model the data in a way that conforms to Machine Learning (ML) algorithms. In particular, fragmentation, and root finding tasks are needed to obtain an efficient data model to overcome data sparsity or high dimensional problems. In this context, the sentiment analysis problem, which is one of the common NLP research problems for Turkish, was chosen from the literature. The experimental performance of BERT was then compared with the basic ML algorithms. While eliminating heavy preprocessing tasks, improved results were obtained in the selected NLP problem compared to the basic ML algorithms.