Twice-Trained Agglomerative clustering approach using topic modeling over Generic Semantic Core Knowledge Graph
Amani Mechergui, Wahiba Ben Abdessalem Karâa, Sami Zghal · 2023
Topic Modeling (TM) can act as a bridge linking unstructured text data to a structured knowledge representation in a Knowledge Graph (KG). The Latent Dirichlet Allocation (LDA) is a commonly used distributional term clustering technique in this regard. However, existing distributional term clustering approaches have not utilized LDA as a bottom-up training strategy with prior knowledge to cluster semantically related terms as concepts across different domains for building and enhancing a GSCKG. We propose to employ a Twice-Trained Agglomerative hierarchical framework using LDA over Generic Semantic Core KG (T2AggLDA-GSCKG), outlined in five steps. We aim to align term topics with the predefined Core Concepts (CCs) of SCKGs, thereby designating modules of a GSCKG that facilitate both its building and enrichment. Our goal is to adapt the LDA term clustering process by utilizing a topic seed-based LDA model to consider these CCs. During the 1stLDA training, we endeavor to identify hypernym and related relations among noun phrase patterns to construct SCKG, and during the 2ndtraining, we aim to enhance it. To achieve this, we will boost the 2ndtraining LDA input data by benefiting from CCs’ two prior knowledge techniques for topic-seed terms incorporation namely seed key knowledge injection and entity masking. The evaluation results show that our proposal has an overwhelming term clustering performance over CCs. It outperforms other unsupervised and semi-supervised distributional baselines on two datasets related to fish hunting and ontology domains, with nearly 13 times higher precision compared to that of normal LDA training.