R-SE2: A Regression Model for Software Effort Estimation Using GPT-3 Embeddings
Eliane Maria De Bortoli Fávero, Gabriel Junges Baratto, Dalcimar Casanova, Jefferson Tales Oliva · 2025
Software effort estimation by analogy (ABSEE) from requirements texts continues to present significant challenges, either in obtaining consistent historical data or inadequately representing contexts to infer effort. Large Language Models (LLMs) can contextualize the textual representation, significantly improving the results obtained in research in this area. Objective: This paper proposes the R-SE2 approach, which evaluates the performance of a regression model based on GPT-3 embeddings for software effort estimation, GPT-3 using exclusively textual software requirements. Method: The GPT-3 LLM was applied to textual features of user stories without fine-tuning. The generated representation was used in a deep learning architecture with a linear output. Results: The results indicate that R-SE2 outperforms the baseline. We highlight the results obtained by applying the proposed model in a single repository containing different projects, where the MAE value is 3.67 and the standard deviation is only 0.12, representing a 14% improvement over the baseline. Conclusion: The results were positive, confirming that LLMs without fine-tuning can be used to estimate ABSEE based on requirements texts. Among the method’s main advantages are its reliability, generalizability, speed, and low computational cost, allowing the inference of effort estimates for new and existing requirements.