Carnegie Mellon University

ARIEL: Analysis of Rare Incident-Event Languages

Information Extraction, Summarization, and Question Answering

Military and humanitarian crises arise unpredictably, requiring rapid response in parts of the world where only low-resource languages are spoken. ARIEL aims to produce an effective suite of algorithms and methods to cope with any low-resource incident language, offering basic but useful functionality in 24 hours, and building to increasingly useful and sophisticated capabilities in a week, month and beyond. Our research will enable key capabilities including topic identification in written and spoken language; identification of key entities (starting with locations and expanding to individuals and organizations); machine translation starting from rough gisting in one to seven days, with increasing capability over time; event detection; and, when possible, entity and event co-reference. These capabilities will be delivered through a flexible omnivorous architecture and interfaced to the analyst in collaboration with the TA2 performer, and will incorporate external TA1.1, TA1.2 and TA1.3 modules from other performers. The system will also improve with experience well beyond the 30-day target, as many crises arise quickly but continue over time.