Analysing Objectives of Auxiliary Inputs in Semantic Web Machine Learning Systems

Abstract

With the increasing popularity of neuro-symbolic systems, the number of systems incorporating both symbolic knowledge and statistical machine learning models is on the rise, leading to a wide variety of system architectures. Authors have different rationales behind their architectural decisions to include certain components into their processing flows, however, these objectives have not been thoroughly investigated. In this chapter, we examine the objectives targeting quality attributes of systems that combine symbolic knowledge represented by Semantic Web Technologies with machine learning approaches, known as SWeML systems. Building on top of the previously introduced SWeML classification system, we conduct a comprehensive analysis of the objectives for adding auxiliary inputs to these systems, i.e., inputs that are not needed to fulfil the system’s primary purpose, but to enhance its capabilities. Specifically, we manually analyze 293 research papers that integrate multiple inputs, exploring the various objectives behind the inclusion of additional symbolic knowledge or sub-symbolic inputs, such as improving performance, reducing response times, and enhancing interpretability. Additionally, we relate the objectives to system characteristics, such as system architecture, targeted task, or application domain, uncovering trends and correlations that can provide deeper insights into the design choices and priorities of SWeML systems.