5.4. Knowledge Graph Generation

The taxonomy only extracts hierarchically related terms, through taxonomic relations. In the knowledge graph extraction step, additional non-taxonomic relations are added to the graph: synonymy and meronymy. It is composed of the aggregation of taxonomic relations (hypernymy/hyponymy), meronomic relations (whole/part_of) and synonymy relations. Note that, due to the fact that hypernym and hyponym are reverse relations (as are whole_of and part_of), only one relationship is represented in the final graph. In the same manner, all synonyms in a given synonymy cluster are related to only one central term.

The knowledge graph extension follows the same overall architecture of the taxonomy construction algorithm. This pipeline is composed of two components: pairwise relationship scoring and knowledge graph construction. The pairwise relationship scoring detects the likelihood of each type of relationship existing between any two terms. The knowledge graph construction component uses these likelihood scores to build the knowledge graph.

The key difference between the taxonomy construction and the knowledge graph construction are the algorithms used to implement each component. Whereas in the taxonomy use case the pairwise scoring used a binary classifier, in the knowledge graph pipeline a multi-class classifier is used instead. Also, a single search algorithm is used for taxonomy construction, while the knowledge graph construction is divided into synonymy cluster and further search algorithms. The details of each component are explained in the following sections.