Deep Knowledge Graph Representation Learning for Completion, Alignment, and Question Answering

Soumen Chakrabarti
2022 Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval  
A knowledge graph (KG) has nodes and edges representing entities and relations. Nodes have canonical entity IDs and edges have canonical relation IDs. E.g., in Wikidata, Barack Obama and Honolulu have canonical IDs Q76 and Q18094, and the relation "place of birth" has canonical ID P19. A fact triple involving these as subject, relation and object is written as (Q76, P19, Q18094), or, more colloquially, (Barack Obama, place of birth, Honolulu). Curating and structuring knowledge has been a human
more » ... pursuit since time immemorial, but large, collaboratively maintained KGs such as WordNet [23], Wikipedia, DBPedia, YAGO [37], Freebase [3], and Wikidata bloomed with the advent of the Internet. Wikidata has over 97 M entities, almost 10 k relations, and close to 1.5 B facts. KGs are of central interest to search, information retrieval (IR) and question answering (QA), as evidenced by Google's purchase of Freebase in 2008, followed by Bing's development of the Satori KG, and Amazon's product KG. The use of KGs in pre-neural search and QA did see some representation [14, 30] in search and IR conferences. However, with the widespread adoption of neural networks and deep learning, the last several years have witnessed a huge surge in KG-related papers in AI, machine learning (ML) and NLP conferences, but comparatively less action in IR conferences. This stands in contrast to dense passage retrieval (DPR), where the IR community embraced deep learning early [20, 25] . My motivation behind proposing this tutorial is to bridge the gap described above, give IR researchers a thorough overview of the best practices of
doi:10.1145/3477495.3532679 fatcat:nmw3yihxkzgf3jgdklv4whhtey