Exploring Sentence Embedding Structures for Semantic Relation Extraction
Alexander Kalinowski, Yuan An · 2021
Sentence embeddings encode natural language sentences as low-dimensional, dense vectors and have improved NLP tasks, including relation extraction, which aims at identifying structured relations defined in a knowledge base from unstructured text. A promising and more efficient approach would be to embed both the text and structured knowledge in low-dimensional spaces and discover alignments between them. We develop such an alignment procedure and evaluate the extent to which sentences carrying similar senses are embedded in close proximity sub-spaces, using that structure to align them to a knowledge graph. Our experimental results show that embedding spaces generated from simple models outperform those from more complicated approaches for the alignment and relation extraction task. We also show that clusterability can serve as a proxy for alignment accuracy, leading us to conclude that better structured spaces drive better semantic applications.