SBIR Phase II: SunDIAL: Slot DIscovery And Linking

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase II project is to search and understand documents across domains. It is a powerful tool for consumers and businesses to route, categorize and ensure key points of a document are seen. This makes it easy for overwhelmed users to search, link topic information to knowledge bases and extract customized data, even if they can't think of the right keywords. Online search can infer the popularity or relevance of a page based on data from millions of users, including a user's location and what they've shown interest in; however, business users don't benefit from these features. Users view concepts in a summary linked to a knowledge base, allowing them to see summaries, triage documents on mobile devices, write rules to sort or filter them automatically, and see key data before ever opening the document. This Phase II project improves information extraction systems by generating answerboxes from complex documents. Consumers will be able to quickly get the gist of formal documents such as consumer credit contracts, insurance policies, industry solicitations, engineering and other difficult-to-read documents. This innovation advances knowledge discovery, information retrieval methods, information extraction, slot-filling, and knowledge-base population. This Small Business Innovation Research Phase II Project discovers unsupervised and unrestricted slots and fillers (attributes and values) to construct structured summaries based on keywords and underlying patterns in document collections. It extends the state of the art by not requiring a manually crafted catalogue of slots and complements supervised approaches by discovering new slots. While conventional Natural Language Processing (NLP) approaches are effective on well-formed sentences, the techniques described here are effective for semi-structured content such as section headers, lists and tables. Additionally, NLP based approaches for fact extraction and text summarization are primarily lexical, requiring further processing for disambiguating and linking to unique entities and concepts in a knowledge base. This approach further advances the state of the art by eliminating these steps as it identifies keywords and links to knowledge base concepts as a first step in the discovery process. This concept-linking enables terms to explicitly map to semantic concepts in other ontologies such that they are available to enable reasoning and understanding. 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.