Transfer learning for semantic similarity measures based on symbolic regression

Article type: Research Article

Authors: Martinez-Gil, Jorge^{a; *} | Chaves-Gonzalez, Jose Manuel^b

Affiliations: [a] Software Competence Center Hagenberg Softwarepark, Hagenberg, Austria | [b] University of Extremadura - Department of Computer Systems Engineering Centro Univ. Mérida, Mérida, Spain

Correspondence: [*] Corresponding author. Jorge Martinez-Gil, Software Competence Center Hagenberg Softwarepark 32a, 4232 Hagenberg, Austria. E-mail: jorge.martinez-gil@scch.at.

Abstract: Recently, transfer learning strategies have become ideal for reusing acquired knowledge through a training phase. The key idea is that reusing such knowledge brings advantages such as increased accuracy and considerable resource savings. In this work, we design a novel strategy for effective and efficient transfer learning in semantic similarity. Our approach is based on generating and transferring optimal models obtained through a symbolic regression process being able to stack evaluation scores from several fundamental techniques. After an exhaustive empirical study, the results lead to high accuracy in addition to significant savings in terms of training time consumed in most of the scenarios considered.

Keywords: Knowledge engineering, Transfer learning, Semantic textual similarity

DOI: 10.3233/JIFS-230141

Journal: Journal of Intelligent & Fuzzy Systems, vol. 45, no. 1, pp. 37-49, 2023