Combining Knowledge Graphs with Semantic Similarity Metrics for Sentiment Analysis

Abstract

The paper proposes a new semantic similarity method with an asymmetry coefficient. The motivation behind this idea is that in some cases it is justified to break the symmetry while comparing certain entities. Such semantic similarity in some cases might be desirable from the psychological point of view. It allows us to enrich embedding methods with knowledge graphs by adding additional information about the specificity of a concept. For the evaluation of the proposed solution, the method has been used as a component for determining a sentiment of reviews. The values of the asymmetry coefficient for the selected set of chosen pairs of words were computed and compared. We present the results of series of experiments comparing the accuracy of different methods in the context of sentiment analysis in various configurations.