📄 The DARPA Wikidata Overlay: Wikidata as an ontology for natural language processing

Summary

Wikidata is the largest crowdsourced knowledge base about the world (over 102 million items). However, it is entity-centric. Events there are mostly named by deverbal nouns, and their typical participants (who acts, toward whom the action is directed) are not linked to these events. This is insufficient for information extraction, in particular for determining the roles of entities in events. The authors worked in the DARPA AIDA, KAIROS and MAA programs and created the DARPA Wikidata Overlay. It is a manually curated and validated subset of Wikidata in which event concepts are mapped to PropBank rolesets. First, over 200 types from the LDC annotation ontology were manually linked to Qnodes. Each mapping was checked at least twice. Then the upper-middle level of the hierarchy was extracted and participant roles were added. In addition, temporal Qnodes and Pnodes were mapped to the relations of Allen's interval logic. The resource is distributed in JSON format. Its value lies in combining established PropBank-based event extraction methods with the broad coverage of Wikidata and its inference capabilities. It also speeds up the adaptation of the ontology to new domains.

Key points

• Version 5.4.5 contains 276 entities, 5,167 events (5,164 of which have PropBank roles), 216 relations (144 of which are linked to PropBank) and 8 temporal relations. The data are available in JSON on GitHub (e-spaulding/xpo).
• PropBank roles are given in two variants: as generic ones (ARG0, ARG1) and as event-specific ones (for example, attacker and victim).
• The main mapping difficulty was that many nominalizations in Wikidata denote the result of an action rather than the process itself (sign.01 → signature). Therefore, a trade-off between coverage and semantic precision had to be found. The annotators also identified hundreds of gaps in the Wikidata event hierarchy.
• The Overlay was initially built only on DWD. This is a Wikidata dump with information up to 2010, restricted in this way because of the "Time Machine" requirement of the AIDA program. Later, the full Wikidata became the source.
• The authors plan to transfer the roles directly into Wikidata. To this end, they propose dedicated Qnodes for event roles and new Pnodes (role index, role function, selectional preference, etc.). From approximately 11,400 rolesets, they expect to obtain 25–40 thousand new Qnodes.