A Quadruplication Multilingual and Multilevel Topic Seeding Approach Towards a Bottom-Up Graph Generation and Enhancement

Multilevel topic-seeding models serve as a connection that links unstructured data to a well-structured knowledge representation in a Knowledge Graph (KG). However, these approaches face challenges in representing KGs for multilingual term clustering, which involves refining the combination of predefined hierarchy, unstructured multilingual data, and structural knowledge. We propose “Quadruplicated Multilingual and Multilevel Latent Dirichlet Allocation” (4Multi2LDA) that integrates topic seed terms into multilevel modeling. Our objective is to generate distinct clusters by adjusting term clusters according to a predefined hierarchy. 4Multi2LDA involves monolingual/multilingual distributional term clustering over Core Concepts (CCs) across different domains, forming upper KGs. This process contributes to the semi-automatic learning and supplementation of a Bottom-Up Universal Upper Knowledge Graph (BU3KG) from unstructured data. In the initial training phase, we focus on establishing “the bottom level” of BU3KG. During the second training phase, we enhance the KG by constructing its “upper level” . In the third and fourth training phases, we aim to further enrich the constructed bottom and upper levels of the KG with new seed terms. The findings demonstrate that 4Multi2LDA exhibits outstanding performance in term clustering over CCs. It surpasses the performance of monolingual and multilingual baselines trained on four datasets.