Semantic Knowledge Representation and Acquisition

Abstract:

This thesis aims to develop a series of methods that automatically extract entities and relationships from unstructured corpora, and effectively represent the structured semantic data, supporting the process of knowledge discovery, knowledge management, and knowledge utilization in intelligent information systems.

In the first study of this thesis, we proposed a novel biquaternion-based knowledge graph embedding method, which integrated multiple geometric transformations, viz., scaling, translation, Euclidean rotation, and hyperbolic rotation, enhancing the representation and reasoning efficacy for various relation patterns.

To facilitate the discovery of relational knowledge, we developed a novel framework for document-level relation extraction, which aimed to achieve better integration of both discriminability and robustness. Specifically, our loss function endowed high discriminability to both probabilistic outputs and internal representations. We also innovatively customized entropy minimization and supervised contrastive learning to address multi-label and long-tailed learning problems.

Lastly, we investigated how to automatically identify semantic nodes and relations from the argumentative corpus. We for the first time proposed a challenging argument quadruplet extraction task, which can provide an all-in-one extraction of four argumentative components, i.e., claims, evidence, evidence types, and stances. To support this task, we constructed a large-scale dataset and proposed a novel quad-tagging augmented generative approach. Extensive experiments demonstrated the empirical superiority of our proposed approaches over several strong baselines.