Sporathe knowledge network
ES
Universidad Nacional de Colombia

Bogotá - Ingeniería - Maestría en Ingeniería - Ingeniería de Sistemas y Computación · 2022

Understanding and Implementing Deep Neural Networks for Unconditional Source Code Generation

Code Generation is a relevant problem in computer science, supporting the automation of tasks such as code completion, program synthesis, and program translation. In recent years, Deep Learning approaches have gained popularity in the code generation problem, and some of these approaches leverage Language Models. However, the existing studies mainly focus on evaluation using machine learning metrics. Additionally, the generation process can be classified into conditional or unconditional (i.e., open-ended) approaches depending on the input context provided to the models. This research proposes CodeGenXplainer, a suite of interpretability methods for Unconditional Language Models of source code. CodeGenXplainer comprises four methods leveraging multiple source code features such as embedding representations, code metrics, compilation errors, and token distributions. Additionally, this research presents an empirical study to validate CodeGenXplainer using publicly available data and extensive sampling of code snippets. Furthermore, CodeGenXplainer provides a base conceptual framework that allows studying multiple complementary perspectives based on machine-generated code. Results show that the studied models can generate code exhibiting similar properties to human code, particularly in terms of code metrics, compilation errors, and token-level information; nonetheless, machine-generated code presents issues with the semantic elements of the code. (Texto tomado de la fuente)

Full text 126 pages with text

Read the full thesis Record in the repository

Contents

  • A.8 Divergence: Compilation errors distributions - RNNs (1)100
  • A.9 Divergence: Compilation errors distributions - RNNs (2)101
  • A.1 Divergence: Compilation errors distributions - RNNs (3)102

Its library's thesaurus