Multiple Semantic-based Similarity And Relatedness Measurements In WordNet

With the development of artificial intelligence technology,measuring semantic similarity and relatedness have more and more applications in many scientific fields such as:Natural Language Processing,human-computer interaction,information retrieval,semantic disambiguation,biomedical domain etc.How to better judge the degree of semantic similarity and relatedness is the key problem to be studiedIn recent years,along with the study of the Semantic Web,ontology has emerged.Ontology gives out formal description of a particular domain through concepts,relations and instance.Structured domain ontology is often used for semantic similarity and relatedness measurement,in which WordNet is used as one.Most previous work,including edge-based and information content-based methods,merely relies on complete semantic relationship in terms of“is-a”relation(hypemymy and hyponymy)provided by WordNet,which are subjected to semantic vacancy and the reliability of the linguistic sources.After all,WordNet,though organized by experts from many research domains,is still a manual dictionary and semantic misunderstanding may inevitably exist in WordNet.Therefore,only adopting complete semantic relationship would constrain the performance of single relationship-based algorithms.We propose a multiple-semantic complementary model based on“is-a”relation.We argue that the "15-a" relation,though complete,has semantic understanding.Therefore we incorporate multiple incomplete semantic relationships provided by WordNet,in which the final semantic similarity and relatedness is obtained by the maximum contribution via all useful semantic relationships.The merits of our proposed model lie in three aspects:(1)We find out multiple semantic relationships that can be used in semantic similarity and relatedness computation in WordNet.And then we proposed a multiple-semantic complementary computation model that using multiple semantic relationships to conduct the semantic similarity and relatedness measurement.Instead of using rough summation of fine-tuning weighting parameters to scale the eontribution of each semantic relationship,we adopt the maximum of them.(2)We use the least common descendant between concepts to give a specific supplement for their commonalities according to the genetics theory to reduce the semantic distance between two concepts and to strengthen the semantic interaction.(3)We apply our proposed model in adjectives and adverbs sematic similarity and relatedness computation by building up a semantic distance for adjectives and adverbs through incomplete semantic relationships.Since adjectives and adverbs are not organized in hypernym and hyponym in WordNet,there are no standard methods to evaluate their sematic similarity and relatedness.To verify the effectiveness of our proposed model,we conduct the comparison of human rating on widely recognized datesets MC30,RG65,AG203,SimLex999 and TOEFL.We show that competitive results can be achieved more than just use of single semantic relationship even in large datesets.In particular,our proposed method can help to alleviate the computing burden of the semantic computing of adjectives and adverbs.In standard evaluations,our method break the bottleneck of semantic similarity and relatedness measurement and make a great contribution in accuracy optimization.