Authors:
Maísa Kely de Melo
1
;
2
;
Allan Victor Almeida Faria
1
;
3
;
Li Weigang
1
;
4
;
Arthur Gomes Nery
1
;
5
;
Flávio Augusto R. de Oliveira
6
;
1
;
Ian Teixeira Barreiro
1
;
7
and
Victor Rafael Rezende Celestino
8
;
1
Affiliations:
1
LAMFO - Lab. of ML in Finance and Organizations, University of Brasilia, Campus Darcy Ribeiro, Brasilia, Brazil
;
2
Department of Mathematics, Instituto Federal de Minas Gerais Campus Formiga, Formiga, Brazil
;
3
Department of Statistics, University of Brasília, Federal District, Brazil
;
4
Department of Computer Science, University of Brasilia, Campus Darcy Ribeiro, Brasilia, Brazil
;
5
Department of Economics, University of Brasilia, Federal District, Brazil
;
6
Ministry of Science, Technology and Innovation of Brazil, Federal District, Brazil
;
7
Department of Economics, University of São Paulo, Ribeirão Preto, Brazil
;
8
Department of Business Administration, University of Brasilia, Federal District, Brazil
Keyword(s):
Automation of Systematic Literature Review, Few-shot Learning, Meta-Learning, Transformers.
Abstract:
Systematic Literature Review (SLR) studies aim to leverage relevant insights from scientific publications to achieve a comprehensive overview of the academic progress of a specific field. In recent years, a major effort has been expended in automating the SLR process by extracting, processing, and presenting the synthesized findings. However, implementations capable of few-shot classification for fields of study with a smaller amount of material available seem to be lacking. This study aims to present a system capable of conducting automated systematic literature reviews on classification constraint by a few-shot learning. We propose an open-source, domain-agnostic meta-learning SLR framework for few-shot classification, which has been validated using 64 SLR datasets. We also define an Adjusted Work Saved over Sampling (AWSS) metric to take into account the class imbalance during validation. The initial results show that AWSS@95% scored as high as 0.9 when validating our learner with
data from 32 domains (just 16 examples were used for training in each domain), and only four of them resulted in scores lower than 0.1. These findings indicate significant savings in screening time for literature reviewers.
(More)