Automatic Structuring of Topics for Natural Language Generation in Community Question Answering in Programming Domain

Abstract

The present article describes the methodology for the automatic generation of responses on Stack Overflow using GPT-Neo. Specifically, the formation of a dataset and the selection of appropriate samples for experimentation are expounded upon. Comparisons of the quality of generation for various topics, obtained using thematic modeling of the titles of questions and tags, were carried out. In the absence of consideration of the structures and themes of texts, it can be difficult to train models, so the question is being investigated whether thematic modeling of questions can help in solving the problem. Fine-tuning of GPT-neo for each topic is undertaken as a part of experimental process.