A Report on Experimental Results of Lower Approximation of Word Embeddings Based on Vector-Based Rough Sets

Abstract

In recent years, embedding models such as Word2Vec and BERT have become the mainstream approach for representing concepts as vectors. However, a key challenge in embedding-based machine learning is the lack of interpretability resulting from its black-box nature. One possible approach to addressing this problem is to integrate rough set theory, which provides a framework for explainable concept representation, with embedding-based machine learning. Following this idea, the vector-based rough set (VRS) model has been proposed, wherein concepts are represented as vectors. This model inherits decision class approximation properties of probabilistic rough set theory. This paper presents the results of an empirical investigation into the relationship between the lower approximation in the VRS model and actual word embedding vectors.