An Evolutionary Strategy For Concept-Based Multi-Domain Sentiment Analysis

Abstract

Inferencing the sentiment expressed within a document is still a challenging task, especially when it is necessary to consider the domain dimension. In order to improve inference algorithm effectiveness, one of the main challenges is to learn polarity values representing the concept-domain pair. In this paper, we present an approach which relies on evolutionary algorithms and exploiting semantic relationships for estimating domain-dependent polarities of opinion concepts. The SenticNet resource is used as a starting point for extracting both concepts and common-sense expression relevant to the sentiment analysis topic. Subsequently, the creation of semantic relations is performed by exploiting the alignments between SenticNet and WordNet. Finally, an evolutionary strategy has been implemented for learning the polarity values of conceptdomain pairs. Our approach has been validated by following the Dranziera protocol and obtained results demonstrated the suitability of the proposed solution.