Representing Sets in Embeddings of Relational Information

• Applicant: Professor Dr. Benjamin Roth
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; General and Comparative Linguistics, Experimental Linguistics, Typology, Non-European Languages
• Term: from 2018 to 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 406696030

The lack of computer-processable information is one of the main obstacles to successful artificial intelligence. Knowledge graphs represent large amounts of facts as a graph where entities (Persons, Locations, Organizations, etc.) are the nodes, and relations are the edges.Constructing knowledge graphs automatically, e.g., by using automatic methods for extracting information from textual corpora (such as Wikipedia and newspaper corpora), is a crucial building block to solving this problem.Extracting relations from text allows computers to further process relational information, and it allows humans to better search and browse.Traditional methods for relation extraction predict single facts by classification.This leads to problems when we consider a relation as a set of tuples (the pairs of arguments for that relation), since traditional methods for relation extraction are not formulated with a unified representation for the sets of entities they have to represent. Instead, they are trained to minimize the loss on the single facts seen during training, limiting their usability to the corresponding prediction of isolated facts.Knowledge graph embedding methods aim at finding vector representations for nodes and edges from which the original knowledge graph can be reconstructed with low error. Traditional approaches to knowledge graph embedding learn vector representations for single entities or facts, rather than for sets of entities or entity tuples.The research program proposed here is motivated by the vision of a representation of relational information that can deal with sets of entities as an integral part of a probabilistic model. The goal is to represent these sets of entities as regions in vector space, and to learn appropriate relational transformations operating on the regions. Relational information will be learned from knowledge graphs and natural language text.Such an approach will provide elegant solutions to several problems that current methods for automated knowledge graph construction face, including (1) representing uncertainty, (2) representing entity sets, (3) modeling intension rather than extension of concepts, and (4) reasoning with incomplete knowledge graphs.

• DFG Programme: Research Grants
• International Connection: Austria, USA
• Cooperation Partners: Professor Dr. Philipp Cimiano; Professor Dr. Andrew McCallum; Professor Dr. Hinrich Schütze