Files
Abstract
Comprehending knowledge is a crucial focus in Artificial Intelligence (AI). A knowledge graph is a structured repository for entities and their relationships, organized in a graph format. Graph nodes symbolize the entities and the edges embody the relations. Knowledge graphs have been employed in various applications, including social networks, question-answering, recommender systems, cybersecurity, healthcare, and finance.
Knowledge graphs mainly utilize text to store entity and relation information. Natural Language Processing (NLP) techniques have enabled machines to comprehend text efficiently, thanks to the recent advancements facilitated by deep-learning and language models. Given the NLP capabilities, supplementing the construction of knowledge graphs with triples became feasible. Extracting triples from text incorporates various tasks, such as Named Entity Recognition (NER) to extract entities from text, and Relation Classification (RC) to identify relations between entities. However, large-scale knowledge graphs are hard to complete. Therefore, the knowledge graph completion task targets establishing missing links between entities. Concurrently, encoding text to exploit entity representation is essential for the knowledge graph completion task.
This dissertation contributes to the existing body of research by examining the exploitation of NLP-based models for two knowledge graph-related tasks: knowledge graph construction and knowledge graph completion. Particularly, the dissertation describes our novel relations classification model, which contributes to the knowledge graph construction task. Furthermore, the dissertation describes our two separate knowledge graph completion models. The first model harnesses the graph's structural and textual information, and the second employs Large Language Models (LLMs). Finally, we show our survey conducted on recent few-shot Named Entity Recognition and Relations Classification models.
Our results suggest that the aforementioned models have achieved new scores on popular benchmarks. These benchmarks evaluate knowledge graph completion models and Relations Classification models. Beyond simply reporting these results, this dissertation contributes to the field by detailing the architecture of these models and making them publicly accessible. The models are featured to be fast, dynamic, easily customizable, and usable, facilitating their smooth adoption by the research community.
Knowledge graphs mainly utilize text to store entity and relation information. Natural Language Processing (NLP) techniques have enabled machines to comprehend text efficiently, thanks to the recent advancements facilitated by deep-learning and language models. Given the NLP capabilities, supplementing the construction of knowledge graphs with triples became feasible. Extracting triples from text incorporates various tasks, such as Named Entity Recognition (NER) to extract entities from text, and Relation Classification (RC) to identify relations between entities. However, large-scale knowledge graphs are hard to complete. Therefore, the knowledge graph completion task targets establishing missing links between entities. Concurrently, encoding text to exploit entity representation is essential for the knowledge graph completion task.
This dissertation contributes to the existing body of research by examining the exploitation of NLP-based models for two knowledge graph-related tasks: knowledge graph construction and knowledge graph completion. Particularly, the dissertation describes our novel relations classification model, which contributes to the knowledge graph construction task. Furthermore, the dissertation describes our two separate knowledge graph completion models. The first model harnesses the graph's structural and textual information, and the second employs Large Language Models (LLMs). Finally, we show our survey conducted on recent few-shot Named Entity Recognition and Relations Classification models.
Our results suggest that the aforementioned models have achieved new scores on popular benchmarks. These benchmarks evaluate knowledge graph completion models and Relations Classification models. Beyond simply reporting these results, this dissertation contributes to the field by detailing the architecture of these models and making them publicly accessible. The models are featured to be fast, dynamic, easily customizable, and usable, facilitating their smooth adoption by the research community.