How dissimilar synonyms affect the results of experiments based on fine-grained knowledge co-occurrence networks

Abstract

Despite efforts to reduce the effects of dissimilar synonyms on the construction of knowledge networks, few researchers have examined the extent to which it affects the results of experiments. In this work, we developed a multi-tiered comparative analysis framework to investigate how the dissimilar synonym issue influences the topology structure and functional dynamics of knowledge networks. Specifically, Pearson correlation analysis was performed to quantify the relationship between topology structure variables in the ontology knowledge network and their counterparts in the raw term-based network. Subsequently, we applied the Levenshtein distance algorithm to assess sequence dissimilarity in the ordinal sequences between variable pairs. Finally, we calculated the difference between the topology variables of the same knowledge node in the two networks. To evaluate the effect of the dissimilar synonym issue, we applied our framework to the scenario of knowledge impact prediction and ranking. The experimental results show that (1) the similarity values of the ordinal sequences of eigenvector centrality, PageRank coefficient, betweenness centrality, and closeness centrality variables are respectively 0.410, 0.404, 0.342, and 0.407, which means the dissimilar synonym issue has a considerable effect on the topology structure calculations of knowledge networks; and (2) higher-ranked knowledge nodes show lower overlap rates between the raw term-based knowledge network and the ontology knowledge network, suggesting that the dissimilar synonym issue influences the reliability of detecting high-impact knowledge. (3) in the scenario of knowledge impact prediction, the dissimilar synonym issue has minimal effect on the task performance.