Comparison of deep learning models in natural language processing

Artificial intelligence has shaped the functioning of life today by adding the concepts of efficiency and productivity. However, machine learning, a sub-branch of artificial intelligence, has come to the forefront especially in the last two decades and has become the most demanded technological field. Artificial intelligence and especially machine learning has also manifested itself in the competition and success of companies and organizations working in this field. Machine learning is the use of models that computers have obtained through training over data to perform certain tasks. One of the fields that use these tasks is natural language processing. The field of natural language processing covers a variety of topics involving the computational processing and understanding of human languages. In the last few years, the field of natural language processing has been further enhanced by the use of deep learning models. This thesis presents a performance comparison of deep learning architectures and methods in the text generation subdomain of natural language processing. In this context, the text generation functions of 7 transformative models downloaded from the huggingface website were used to calculate similarity ratios with the measurement metrics used in this field. The aim was to measure the similarity and semantic integrity between the reference texts and the texts produced by the models. The experiments showed that model size has a significant effect on performance. Among the 5 measurement metrics, BertScore, which is transducer-based, was found to be more successful than n-gram-based metrics.