Inproceedings3571
Towards Holistic Concept Representations: Embedding Relational Knowledge, Visual Attributes, and Distributional Word Semantics
Towards Holistic Concept Representations: Embedding Relational Knowledge, Visual Attributes, and Distributional Word Semantics
Published: 2017
Oktober
Buchtitel: The Semantic Web – ISWC 2017
Verlag: Springer
Referierte Veröffentlichung
BibTeX
Kurzfassung
KnowledgeGraphs(KGs)effectivelycaptureexplicitrelationalknowl- edge about individual entities. However, visual attributes of those entities, like their shape and color and pragmatic aspects concerning their usage in natural lan- guage are not covered. Recent approaches encode such knowledge by learning la- tent representations (‘embeddings’) separately: In computer vision, visual object features are learned from large image collections and in computational linguistics, word embeddings are extracted from huge text corpora which capture their distri- butional semantics. We investigate the potential of complementing the relational knowledge captured in KG embeddings with knowledge from text documents and images by learning a shared latent representation that integrates information across those modalities. Our empirical results show that a joined concept rep- resentation provides measurable benefits for i) semantic similarity benchmarks, since it shows a higher correlation with the human notion of similarity than uni- or bi-modal representations, and ii) entity-type prediction tasks, since it clearly outperforms plain KG embeddings. These findings encourage further research towards capturing types of knowledge that go beyond today’s KGs.
Download: Media:Towards Holistic Concept Representations.pdf