Bitte wählen Sie ihr Lieferland und ihre Kundengruppe
The aim of the DARPA AIDA program was to detect disparate hypotheses about a common claim or question within text and image data, using three domains/scenarios: 1) the war in Ukraine 2013-14, 2) unrest in Venezuela, and 3) claims regarding the COVID virus. The program was separated into three main tasks, where Technical Area 1 (TA1) extracted knowledge graphs from text and images, TA2 fused knowledge graphs using coreference reasoning, and TA3 extracted hypotheses. The U Texas AIDA project was a TA3 project: It detected distinct hypotheses in knowledge graphs extracted and assembled by TA1 and TA2 teams at other sites. Research and development in the U Texas AIDA project resulted in two distinct approaches, one for the Ukraine and Venezuela scenarios, and a separate approach for the COVID scenario. For the first approach, the central idea was to use narrative coherence as a criterion for assembling hypotheses. We trained neural networks on the task, using for training synthetic data at scale by automatically generating Story Salads, mixtures of multiple narratives. In the evaluations, especially the 2020 hackathon, this architecture proved to be simple but effective, and able to quickly adapt to changes in the textual data associated with knowledge graph nodes. For the COVID scenario the task was to determine compatibility relations between pairs of claims. Here our approach used two types of classifiers working on English raw text data: a relatedness classifier to roughly group claims, and a natural language inference system to determine supporting and refuting relations.