SEMANTIC KNOWLEDGE GRAPH REPRESENTATION AND ACQUISITION

Abstract

This thesis develops a systematic framework for effectively constructing and representing semantic knowledge graphs, aiding knowledge management in information systems.

Firstly, we introduce a biquaternion-based knowledge graph embedding method that integrates multiple geometric transformations (i.e., scaling, translation, Euclidean rotation, and hyperbolic rotation) to enhance representation and reasoning efficacy for various relation patterns.

Next, we present a novel method for document-level relation extraction, focusing on better integration of discriminability and robustness. Our loss function enhances discriminability in both probabilistic outputs and internal representations, along with a negative label sampling strategy to mitigate annotation noise.

Lastly, we explore the automatic identification of semantic nodes and relations from argumentative corpora. We propose a challenging argument quadruplet extraction task. To support this task, we build a large-scale dataset and propose a novel quad-tagging augmented generative approach.

Extensive experiments validate the superiority of our proposed methods over several strong baselines.