Comparison of deep learning methods for code generation: applications in the java language

Human beings have the desire to transfer their feelings and thoughts to other people. They try to use tools to provide these transfers. The most important of these tools is language. Language is the tool that enables people to understand each other. The desire of people to communicate has expanded from communicating with each other and increased the desire to communicate with the environment. People's efforts to communicate with the machines around them have also increased. Especially the desire of people to transfer their feelings and thoughts to machines has led to the emergence of programming languages. Although programming languages have different structures compared to natural languages, people have thought of them as a language. These ideas have developed the idea of translation applications between programming languages and natural languages. Today, with the development of technology and artificial intelligence, significant progress has been made on these studies. Especially hardware developments have led to a renewed interest in deep learning methods. In this thesis, the role of deep learning models in automatic code generation from natural languages to programming languages is extensively discussed. The main objective of the study is to analyse the effectiveness of these models by comparing the Long Short-Term Memory (LSTM) model for text-based code generation with the LSTM-based encoder-decoder architecture with an attention mechanism. Today, code generation processes are an important area that increases automation and efficiency in software engineering. In this study, code generation processes are simulated, and performance analyses are performed using the LSTM model. The LSTM model is known for its capacity to work with sequential data and its ability to learn long-term dependencies. However, it may be insufficient in generating code from long and complex texts. To overcome these deficiencies, an attention mechanism has been added to the LSTM-based encoder-decoder architecture. The attention mechanism contributes to more accurate and consistent code generation by enabling the model to focus on important inputs. In this study, the performance of both models is analysed within the framework of certain criteria. As a result of the comparisons, it was found that the LSTM-based encoder-decoder architecture with attention mechanism is more effective in code generation processes. This study demonstrates the potential of deep learning models in software development processes and draws attention to innovative applications such as natural language to programming language conversion. The findings obtained within the scope of the thesis are of nature to encourage the use of deep learning models in code generation and software development areas and aim to contribute to future research in this field. In this direction, the study offers new perspectives for researchers in artificial intelligence and software engineering disciplines.