Information extraction pipelines for knowledge graphs.

Area of Science:

• Artificial Intelligence
• Data Science
• Natural Language Processing

Background:

• Numerous knowledge graph (KG) completion methods exist but operate in isolation.
• The collective performance and limitations of these disjointed approaches are not well understood.
• Existing KG completion research lacks a unified framework to integrate and compare diverse methodologies.

Purpose of the Study:

• To extend the Plumber framework by integrating 40 reusable components for KG completion subtasks.
• To enable dynamic generation of 432 distinct knowledge extraction pipelines.
• To optimize pipeline selection based on input sentences using a transformer-based model.

Main Methods:

• Extended the Plumber framework with 40 components for KG completion subtasks (e.g., coreference resolution, entity linking, relation extraction).
• Developed a system to dynamically generate 432 distinct knowledge extraction pipelines.
• Trained a transformer-based classification model to select optimal pipelines based on input sentence embeddings.
• Evaluated Plumber on standard datasets across DBpedia, Wikidata, and Open Research Knowledge Graph.

Main Results:

• Plumber effectively generates KG completion pipelines tailored to input sentences.
• The framework demonstrated superior performance compared to all baseline methods, irrespective of the KG used.
• Analysis revealed insights into component synergies, failure cases, and limitations.

Conclusions:

• Plumber provides a unified and effective approach to knowledge graph completion.
• Dynamic pipeline generation optimizes KG completion performance across diverse datasets and KGs.
• The framework facilitates a deeper understanding of the strengths and weaknesses of various KG completion techniques.