SE3M: A model for software effort estimation using pre-trained embedding models

doi:10.1016/j.infsof.2022.106886

2022-07-01

PRE

Eliane Maria De Bortoli Fávero *

Dalcimar Casanova

Andrey Ricardo Pimentel

DOI：10.1016/j.infsof.2022.106886

原文链接

原文求助

摘要

En 中文

Context: Software effort estimation from requirements texts, presents many challenges, mainly in getting viable features to infer effort. The most recent Natural Language Processing (NLP) initiatives for this purpose apply context-less embedding models, which are often not sufficient to adequately discriminate each analyzed sentence. Contextualized pre-trained embedding models have emerged quite recently and have been shown to be far more effective than context-less models in representing textual features. Objective: This paper proposes evaluating the effectiveness of pre-trained embedding models, to explore a more effective technique for representing textual requirements, which are used to infer effort estimates by analogy. Method: Generic pre-trained models went through a fine-tuning process for both approaches - context-less and contextualized. The generated models were used as input in the applied deep learning architecture, with linear output. The results were very promising, realizing that contextualized pre-trained embedding models can be used to estimate software effort based only on requirements texts. Results: We highlight the results obtained to apply the contextualized pre-trained model BERT with fine-tuning, applied in a single repository containing different projects, whose Mean Absolute Error (MAE) value is 4.25 and the standard deviation is only 0.17. This represents a result very positive when compared to similar works. Conclusion: The main advantages of the proposed estimation method are reliability, the possibility of generalization, speed, and low computational cost. Such advantages are provided by the fine-tuning process, enabling to infer effort estimation for new or existing requirements.

Keyword:

Software effort estimation

Pre-trained model

Context-less embedding

Contextualized embedding

Domain-specific model

BERT