Using Standardized Lexical Semantic Knowledge to Measure Similarity

The issue of sentence semantic similarity is important and essential to many applications of Natural Language Processing. This issue was treated in some frameworks dealing with the similarity between short texts especially with the similarity between sentence pairs. However, the semantic component was paradoxically weak in the proposed methods. In order to address this weakness, we propose in this paper a new method to estimate the semantic sentence similarity based on the LMF ISO-24613 standard. Indeed, LMF provides a fine structure and incorporates an abundance of lexical knowledge which is interconnected together, notably sense knowledge such as semantic predicates, semantic classes, thematic roles and various sense relations. Our method proved to be effective through the applications carried out on the Arabic language. The main reason behind this choice is that an Arabic dictionary which conforms to the LMF standard is at hand within our research team. Experiments on a set of selected sentence pairs demonstrate that the proposed method provides a similarity measure that coincides with human intuition.