Vietnamese sentence similarity based on concepts

Abstract

We propose a novel method for measuring semantic similarity of two sentences. The originality of the method is the way that it explores the similarity of concepts referred to in the sentences using Wikipedia. The method also exploits Wiktionary to measure word-to-word similarity. The overall semantic similarity is a linear combination of word-to-word similarity, word-order similarity, and concept similarity. We build datasets consisting of 45 Vietnamese sentence pairs and then evaluate the method on these datasets. The results show that in the best cases, concept similarity help improving the performance of our method more than 15% point. The proposed method is language-independent and quite easy to employ. Therefore, one can readily adopt our method to measure semantic similarity for sentences written in other languages.