Identifying the driving factors of word co-occurrence: a perspective of semantic relations

This study aims to investigate and identify the driving factors of word co-occurrence from the perspective of semantic relations between frequently co-occurring words. Natural sentences in a corpus of news articles were used as co-occurrence windows to extract co-occurring word pairs, and the distance of those two words was not limited. ConceptNet (a semantic knowledge base) was used to annotate the semantic relation between co-occurring words. To solve the problem that some co-occurring word pairs fail to match direct semantic relations in ConceptNet, we proposed a relation annotation method by connecting them with an intermediate word. Results showed that six semantic relations in ConceptNet, (i.e., RelatedTo, IsA, Synonym, HasContext, Antonym, and MannerOf) were important factors directly inducing word co-occurrence. The combination of some of those semantic relations was an important factor indirectly driving word co-occurrence. Also, syntactic analysis and lexical semantic theories were combined to analyze the direct and indirect semantic relations. In this analysis, we found that the factors driving word co-occurrence in sentences could be classified into three relation categories: collocation and modification, hyponymy, and synonym and antonym. These findings can help explain the phenomenon of word co-occurrence and improve the method and application of co-word analysis.