BSF: 2016257: Building Models for Reading Comprehension in Specialized Domains from Scratch

点击翻译按钮获取中文摘要

Machine learning algorithms are increasingly allowing people to search, structure, and access the textual information created daily in every possible domain. In areas with abundant annotated data, where supervised learning algorithms can be applied, algorithms for text understanding have had success in structuring text and providing natural language interfaces. When building a system in a new domain for which there is little to no data, however, data collection and annotation data can be prohibitively expensive. This project explores a protocol for developing text understanding systems that read text and provide a natural language interface in a particular domain (such as biology or history) -- this can allow specialized communities to have digital access to data that is otherwise locked in text. The project also trains students as part of an international collaboration -- this award supports travel of the US-based researchers to collaborate in a project funded by the US-Israel Binational Science Foundation. The project encompasses both data collection and model training, and considers the interaction between the two. To replace expert annotations it uses crowdsourcing workers in an iterative procedure that starts training a structured predictor from almost no data. It creates an interactive framework in which users ask questions and verify candidate answers that are later used to retrain the system. It aims to jointly train over multiple domains, and use domain adaptation methods to transfer knowledge from one domain to another. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.