Syntactic-Semantic Similarity Based on Dependency Tree Kernel

The representation of words in the vector space generated using the Word2Vec model does not capture the syntactic similarity between sentences, but only the semantic similarity by considering the context of words. To address this problem, we propose to exploit the dependency tree that gives the syntactic grammar relations among words in a sentence that is combined with the vector representation of the similarity computation of words in sentences. We also adapt and investigate the effect of a dependency tree kernel that incorporates the dependency relation type with the similarity of tree nodes to measure the similarity between two sentences. Using three pre-trained embedding models and FARASA dependency parser, we conduct experiments on Arabic paraphrasing benchmark and SemEval-MSRvid dataset to evaluate the adapted dependency tree kernels. The results show that these kernel functions provide satisfactory results in terms of precision and recall with AraBERT embedding model while Aravec and Fasttext achieve the best correlation values.